[PDF]

ARTIFICIAL INTELLIGENCE: GETTING SMARTER PDF Free Download

1 / 311
0 views•311 pages

ARTIFICIAL INTELLIGENCE: GETTING SMARTER PDF Free Download

ARTIFICIAL INTELLIGENCE: GETTING SMARTER PDF free Download. Think more deeply and widely.

GLOBAL VIEW
MAY 2023
ARTIFICIAL INTELLIGENCE:
GETTING SMARTER
A global search for the real opportunities when investing in
Artificial Intelligence
See Disclosure Appendix of this report for important disclosures and analyst certifications
Published: 31 May 2023 20:02 UTC | Revised 31 May 2023 20:02 UTC
BERNSTEIN
Colin McGranahan,Global Director of Research colin.mcgranahan@bernstein.com +1 212 407 5824
Michael W. Parker,Head of APAC & EMEA Research michael.parker@bernstein.com +44 207 170 5056
Jay Huang,Greater China Director of Research jay.huang@bernstein.com +852 2918 5746
May 31, 2023
PORTFOLIO MANAGER'S SUMMARY
On a recent spring morning in Tokyo, one of our Directors of Research was interviewing a junior
associate candidate. The conversation turned — in a way only finance job interviews can — to
the subject of the economic life of cruise ships. To expedite the discussion, both the DoR and the
candidate got out their phones. The research manager — using a large and well-known search
engine — was instantly served articles about the average age of cruise ship passengers (48
and falling in our post-pandemic, experience economy), along with a ton of advertisements for
cruises. Meanwhile, the candidate — using ChatGPT — got the right answer immediately: the
average age of a cruise ship in service today is roughly 18 years. "Do you always use ChatGPT for
search?," the candidate was asked. "No, I use the legacy tool too. But only for the simple stuff."
The speed at which Artificial Intelligence (AI) in general and OpenAI's GPT suite in particular
has, in the last six months, taken over seemingly every conversation was the genesis of this
Blackbook. We asked our analysts to identify where, why, and how AI is going to create, destroy,
and transfer value in their sectors. We also asked the analysts to identify the companies in their
sectors that could win and lose over the next three to five years as a result of AI. Finally, we looked
for contrarians, asking: has everyone simply lost their minds (again)?
In this Blackbook, we have contributions from two dozen analysts, ranging from analysis sizing
Large Language Model training as a source of demand for computing power globally, a rank-
ordering of winners and losers in China's internet sector, a review of the coming mass adoption
cycle of AI in manufacturing, and an exploration of the use of AI in biologics drug discovery. And
— of course — we ask: will AI kill Google Search?
The speed of development in AI may be best gauged by the rate at which the technology
decouples from its framing devices. In 2023, discussing AI in terms of chess, Go, autonomous
driving, TikTok, the Turing Test, or even the Singularity feels arcane and slightly embarrassing.
Even the revelation, generally at this point in any AI essay or thought piece, that the above text was
written by ChatGPT, already feels very winter 2023. Once or twice in a decade, a new technology
will become ubiquitous, seemingly overnight: Google, iPhone, Zoom…and now GPT. There is no
Will-They-Won't-They when it comes to our coverage companies — roughly 600 of them globally
— and AI. The question is no longer the rate of adoption. What matters — and what we search for
in this Blackbook — are the unexpected consequences.
Many of the binaries we apply in analyzing new technologies — adoption, sophistication,
competition, substitution, and scale — have simply been overwhelmed by AI's universal embrace.
The AI landscape has become so rich and diverse that we believe the only way to analyze this
shift is at an industry-by-industry level globally. A single voice or framework can sometimes work
when analyzing new technologies. But, as they say, only for the simple stuff.
ARTIFICIAL INTELLIGENCE: GETTING SMARTER 1
BERNSTEIN
2 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
BERNSTEIN
TABLE OF CONTENTS
STRATEGY
ASIA QUANT STRATEGY: QUANT-AS-A-SERVICE COULD BE A GAME-
CHANGER
7
- Rupal Agarwal
CONSUMER
OTAS AND HOTELS: BEGINNING OF A RAPID ADOPTION CURVE? 21
- Cherry Leung, Richard J. Clarke, FCA, Pearl Xu, Kate Xiao, CFA, Niall Mitchelson
OPPORTUNITIES FOR AI IN US RESTAURANTS 31
- Danilo Gargiulo, Bill He
HAIER SMART HOME: OPPORTUNITIES FOR AI IN HOME APPLIANCES 45
- Melinda Hu, Charles Gou, Shirley Yang
US APPAREL & SPECIALTY RETAIL: TO BUILD OR TO BUY? OPPORTUNITIES
FOR AI IN RETAIL
57
- Aneesha Sherman, Shradha Mani, Jessica Tian
GLOBAL LUXURY GOODS: LVMH AND FARFETCH LEAD THE AI REVOLUTION 71
- Luca Solca, Renny Shao, Clementine Flinois
OCADO: AI IN ONLINE GROCERY AUTOMATION AND OPTIMIZATION 83
- William Woods, Eric Chen, CFA, Alexander Nielsen
INDUSTRIALS
AI IN AGRICULTURE: RISE OF THE GROWBOTS 97
- Chad Dillard, Nicholas J. Green, Miguel Marques, CFA, Ellen Lundstrom, ACA
EUROPEAN AIRLINES & AIR TECH: OPPORTUNITIES FOR AI IN FREIGHT
FORWARDING
109
- Alex Irving, CFA
INDIA INDUSTRIALS & INFRA: OPPORTUNITIES FOR AI IN THE POWER
SECTOR
121
- Nikhil Nigania, Anusha Madireddy
FINANCIALS
ASIA INSURANCE: CHATBOTS TO MAKE "LIFE" BETTER? 137
- Tianjiao Yu, Cally Yang
ARTIFICIAL INTELLIGENCE: GETTING SMARTER 3
BERNSTEIN
HEALTHCARE
GLOBAL CDMOS: USE CASES OF AI IN BIOLOGICS DRUG DISCOVERY 151
- Nithya Balasubramanian, Parth Shah
EU MEDTECH: AI IN MEDICAL IMAGING 161
- Lisa Bedell Clive, Jonathon Unwin
US HEALTHCARE SERVICES: WILL ROBOTS REPLACE DOCTORS? 173
- Lance Wilkes, Amir Farahani, CFA, William Robbins
TECHNOLOGY
US E-COMMERCE: OPPORTUNITIES AND RISKS FROM AI 189
- Nikhil Devnani, CFA, Eva Zhang
SOUTH & SOUTHEAST ASIA TECH: OPPORTUNITIES FOR AI IN CONSUMER
TECH
201
- Venugopal Garre, Ankit Agrawal, CFA
GLOBAL AUTOMATION: LATEST DEVELOPMENTS IN THE MANUFACTURING
SECTOR
213
- Jay Huang, Ph.D., Dien Wang, Ph.D., Weibin Liang, Ph.D.
INDIA TECHNOLOGY, MEDIA & INTERNET: OPPORTUNITIES FOR AI IN
TECHNOLOGY SERVICES
223
- Rahul Malhotra, Sanjit Shinde
GLOBAL SOFTWARE: AI IS CORE TO THE FUTURE — HIGHLIGHTS FROM
ADOBE SUMMIT CONFERENCE
235
- Mark L. Moerdler, Ph.D., Firoz Valliji, CFA, Sahr Singh
BOTTOM-UP APPROACH TO SIZING THE LLM SILICON OPPORTUNITY (PART 1
— INFERENCE)
243
- Stacy A. Rasgon, Ph.D., Akhilesh Kumawat, Alrick Shaw
BOTTOM-UP APPROACH TO SIZING THE LLM SILICON OPPORTUNITY (PART 2
— TRAINING)
253
- Stacy A. Rasgon, Ph.D., Akhilesh Kumawat, Alrick Shaw
WILL AI KILL GOOGLE SEARCH? 269
- Mark Shmulik
CHINA AI: GETTING SMARTER...SCORING THE GENERATIVE AI CONTENDERS 277
- Robin Zhu, Boris Van, Mark L. Moerdler, Ph.D., Ronald Ma, Ke Li, Xuan Ji
4 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
BERNSTEIN
Strategy
STRATEGY 5
BERNSTEIN
6 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
ASIA QUANT STRATEGY: QUANT-AS-A-
SERVICE COULD BE A GAME-CHANGER
HIGHLIGHTS •Adoption of AI in asset management is low, but we expect AI to take higher mindshare
and investment dollars going forward. AI has been around for decades and has been
central to quant investors. However, the broader investment industry is quite behind —
according to the 2019 CFA Institute survey, only 10% of portfolio managers (out of 230)
use AI/ML techniques, while 75% of the analyst community doesn't use any AI or big data
technique. The CFA Fintech survey in 2020 showed that AI adoption in risk management also
remains low — only 5% said they used AI widely. The difficulty in explaining AI/ML models,
data mining/overfitting, the changing nature of financial markets, and lack of talent have been
some key reasons for the limited acceptance. But AI growth will likely be strong. A survey
of 500 wealth/asset managers in 2021 by Deloitte highlighted AI as the number one area
to attract the largest amount of investments over the next two years — growing by 11% (a
52% jump over past levels). The outperformance of the AI-powered hedge fund index (9.8%
CAGR) versus the broader hedge fund index (5.8% CAGR) since 2010 and even versus the
S&P500 during 2010-19 makes the case stronger for AI. The increased development of AI-
driven products since ChatGPT highlights that this AI wave is going to be much stronger and
fruitful than the last.
•Use cases of AI/ML in asset management. AI techniques can be used to improve
portfolio management, generate trading signals, manage risks, enhance investment advice
within wealth management, and capture ESG-specific risks. For quants, we note that
exploring alternative datasets to identify new sources of alpha remains a key focus area. We
demonstrate a simple example of how AI can help in portfolio construction: We used simple
clustering analysis on our last published 30 stock Q+F portfolio, and were able to: (1) identify
stocks with similar return profiles across different sectors, and (2) improve portfolio returns
— the portfolio picked by a clustering model was able to generate 13% CAGR since 2018
(IR=0.5) versus 10% CAGR (IR=0.43) by a normal portfolio.
•Quant-as-a-Service (QaaS) has true potential. Discretionary investment firms find it hard
to embrace the latest AI developments due to the shortage of resources/talent. QaaS could
speed up the AI adoption, as asset managers could get a full stack of quant solutions from
a third party. Open-source software (Qlib, PyAlgo trade, etc.), community-based platforms
(Kaggle, Numerai, etc.) or trading APIs (Alpaca) have made life easier for people with relevant
quant skills. However, there aren't many companies providing end-to-end AI solutions for the
investment community — some notable firms are Evalueserve, Fractal Analytics, and Acuity
KP (all private), but the scope for growth within the space is huge.
BERNSTEIN
ASIA QUANT STRATEGY: QUANT-AS-A-SERVICE COULD BE A GAME-CHANGER 7
AI ADOPTION IN ASSET
MANAGEMENT REMAINS
LOW
While AI has been around since the 1950s, the past decade saw an increased interest in the
space from asset managers, primarily led by increased computing power and cloud storage, and
availability of bigger datasets. Since ChatGPT went viral, there has been a significant interest
in AI again as indicated by the sharp pickup in Google search trend on AI keyword (Exhibit 1).
ChatGPT has made AI discussions more mainstream and, even in the asset management industry,
there is renewed demand — we ourselves have had multiple inbounds asking how all this would
change quants. We would argue that AI is not new for quants. In fact, it has been central to quant
strategies for years. We all know about the big quant hedge funds such as DE Shaw, Two Sigma,
AQR, and Renaissance Technologies (all private) to name a few, which have been adopting a data-
driven and systematic way of investing, using AI techniques, advanced mathematical models,
and quantitative analysis to build investment frameworks. However, the use of AI remains low
currently in the wider asset management industry.
CFA Institute survey: In 2019, the CFA Institute conducted the AI/big data survey "AI
Pioneers in Investment Management" for investment management industry to identify high-
impact applications of AI and big data, as well as best practices. The survey was sent to a
randomized sample of CFA Institute charterholders, and there were a total of 513 respondents
(52% from the Americas,18% from Asia Pacific, and 30% from Europe, the Middle East, and
Africa). Respondent occupations spanned from both buy side and sell side equity as well as
credit analysts (N=159), portfolio managers (N=230), and private wealth managers (N=124).
The results show that AI adoption in the investment management industry was low — only
10% of portfolio managers who responded to the survey used AI/ML techniques to improve
the investment process, and most PMs (~50%) still use traditional linear regression and back
tests for investment decision-making. Within the analyst community, 75% of respondents didn't
use any AI and big data techniques for their analysis. Those who used, the two most popular
techniques cited were scraping third-party websites and using NLP for textual data processing
(Exhibit 2 and Exhibit 3). In a similar CFA Fintech survey in 2020, results from respondents
revealed that AI adoption in risk management also remains low — only 5% said they used AI
widely, while more than 50% were either not using AI at all or were still in the investigating stage
(Exhibit 4).
Bernstein survey: In 2019, our global quant team conducted a survey as part of our quant
conference in London, and the results were quite similar. AI and big data were regarded as
relatively unimportant among other topics in a group of ~200 asset managers (and bear in
mind that this is from a group of investors with an over-representation of quants). When we
asked specifically about incorporating AI or machine learning into their investment process,
most respondents (~70%) said they are "researching" it. However, we did see an increase over
previous years in the number of respondents saying that such approaches were already live,
suggesting there was some increase in adoption. In an earlier work, our global quant team in
2018 highlighted that the AUM for AI hedge funds was ~US$1Bn. While it might have grown
since then, the scale would still be too small. The first actively managed ETF to fully utilize AI as
a method for stock selection, AIEQ, has also seen a fair bit of volatility in its AUM — moving from
US$70Mn to US$200Mn from 2017-18, then seeing another surge post Covid-19, reaching US
$248Mn, however, moving back to US$120Mn AUM recently. In an industry where some asset
managers have trillions of dollars under management, US$120Mn is nothing.
BERNSTEIN
8 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
EXHIBIT 1: The search trend for AI has dramatically
increased after ChatGPT went viral
0
20
40
60
80
100
120
4/1/2018 4/1/2019 4/1/2020 4/1/2021 4/1/2022
Google Search Trend for Artificial
Intelligence
Source: Google.com, Bernstein analysis
EXHIBIT 2: AI adoption in asset management remains
low — of the 230 portfolio managers surveyed, only 10%
said they use AI to improve their investment process
50%
49%
33%
19%
10%
0% 20% 40% 60%
Run a backtest of a strategy
Regression analysis to find a
linear relationship
None
Run a backtest of an algorithm
Artificial intelligence/machine
learning to find a nonlinear
relationship or estimate
Investment Strategy and Process
Conducted by Portfolio Manager (N=230)
Note: Survey answers are in multiple choices. Data is based on CFA 2019 survey
"AI Pioneers in Investment Management."
Source: cfainstitute.org, Bernstein analysis
EXHIBIT 3: Of the 159 buy-side analysts surveyed, 75%
said they don't use AL/ML techniques; among those
who used AI, the two most popular techniques cited
were scraping third-party websites and using NLP
75%
14%
10%
9%
8%
6%
2%
0% 50% 100%
None
Scraping third-party websites
Using NLP to read large tracts
of text, transcripts etc
Using deep learning to gauge
sentiment in social media
Extracting alpha from
unstructured data
Using robotic process
automation
Using deep learning to count
cars in parking lots
Artificial Intelligence/Machine Learning
Use Cases for Industry and Company
Analysis by Analysts (N=159)
Note: Survey answers are in multiple choices. Data is based on CFA 2019 survey
"AI Pioneers in Investment Management."
Source: cfainstitute.org, Bernstein analysis
EXHIBIT 4: Even for risk management, AI application
remains low — of the 216 participants surveyed, only 5%
said they use AI widely, while 55% are either not using
AI at all or are still in the investigating stage
32%
23%
23%
9%
9%
5%
0% 10% 20% 30% 40%
Not at all
Investigating its use
Using in limited applications
In the process of
implementing these tools
Don't know/Unsure
Widespread use
Use of AI in Risk management (N = 216)
Note: Data is based on 2020 CFA Institute Fintech survey covering regions of the
Americas, Asia Pacific, and EMEA.
Source: cfainstitute.org, Bernstein analysis
BERNSTEIN
ASIA QUANT STRATEGY: QUANT-AS-A-SERVICE COULD BE A GAME-CHANGER 9
KEY CONCERNS ABOUT
AI IMPLEMENTATION IN
INVESTMENT MANAGEMENT
INDUSTRY
Exhibit 5 highlights some key challenges and concerns that have often been highlighted as
reasons for this slower adoption. One of the biggest issues has been with respect to explainability
— because of the models being too complex with numerous input parameters, it becomes
difficult to explain the rationale that drives investment decisions. An asset manager should
know why a strategy is outperforming or underperforming, but this becomes hard to know in
an AI-driven instance. The regulatory requirement for asset managers to be transparent and
accountable makes it harder to rely on an AI-driven strategy — if something goes wrong, one
cannot blame a computer model. Data mining has been another long-standing pushback that
quants have faced, and with more and more datasets available, the chances of spurious patterns
being picked up to predict future returns raises questions on the reliability of such models.
Given financial markets are driven by many factors, including expectations and interpretations
of market participants, which keep on changing with time, relying on models that primarily
use historical trends and patterns, and hence are unlikely to pick any meaningful changes in
external drivers, market structures, etc., becomes another bottleneck in the acceptance of such
techniques. Another more practical reason has been the lack of relevant talent and required
skillset. While there has been a rise in the number of people who would call themselves "data
scientists," we don't find the pool big enough for faster adoption of quant/AI techniques in the
wider asset management industry.
EXHIBIT 5: Key challenges/concerns for using AI techniques in investment process
Source: Bernstein analysis
BERNSTEIN
10 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
CASE FOR INCREASED
ADOPTION OF AI GOING
FORWARD
One of the biggest support for increased adoption of AI within the asset management industry
would be the evidences of strong, consistent, and less volatile performance of AI-powered
funds versus the broader market. Exhibit 6 shows the performance of the Eurekahedge Artificial
Intelligence Hedge Fund Index (an equal-weighted index of 15 constituent funds that utilize AI
and machine learning in their trading processes) versus the S&P500 and versus a broader hedge
fund index. Since 2010, this index has been able to consistently beat the broader hedge fund
index, generating 9.8% CAGR versus 5.8% CAGR by Eurekahedge Hedge Fund Index. While
S&P500 total returns have been slightly better at 12.6% per year, most of that outperformance
has come only post Covid-19. During 2010-18, the AI hedge fund (12.8% CAGR) was able to
outperform the S&P 500 (12.3% CAGR) while showing lower volatility.
While excess returns would draw more attention from asset managers toward AI, we find that
AI is where most incremental dollars would be going over the next two years based on a recent
survey conducted by Deloitte (collecting data from 500 international wealth and asset managers
representing over 33 trillion dollars of AUM in 2021). AI/machine learning/NLP was highlighted
as the number one area which would attract the largest amount of investments in the next two
years — growing by 11%, a 52% jump over past levels (Exhibit 7). AI stands out as one of the
technologies with the most positive impact, according to the survey. Hence, while AI adoption has
been increasing within the asset management industry over the last decade, it could see a much
faster adoption and investment going forward as there is an overall push toward digitization post
Covid-19, and now with the success of ChatGPT, there is a realization that AI could be a much
more powerful tool.
EXHIBIT 6: The Eurekahedge AI Hedge Fund Index has
been outperforming the normal hedge fund index
0
100
200
300
400
500
600
1/1/2010
10/1/2010
7/1/2011
4/1/2012
1/1/2013
10/1/2013
7/1/2014
4/1/2015
1/1/2016
10/1/2016
7/1/2017
4/1/2018
1/1/2019
10/1/2019
7/1/2020
4/1/2021
1/1/2022
10/1/2022
AI Fund Performance vs. Broader Market
Performance
The Eurekahedge AI Hedge Fund Index
SP500 total return
The Eurekahedge Hedge Fund Index
Note: The Eurekahedge AI Hedge Fund Index is an equally weighted index of
15 constituent funds. The index is designed to provide a broad measure of the
performance of underlying hedge fund managers who utilize AI and machine
learning theory in their trading processes. The S&P500 total return index
including reinvestment of dividends is used here.
Source: Bloomberg, Bernstein analysis
EXHIBIT 7: AI/machine learning/NLP was ranked #1 in
terms of seeing the biggest increase in investment in
the next two years
Source: Deloitte ThoughtLab Wealth and Asset Management 4.0, Bernstein
analysis
BERNSTEIN
ASIA QUANT STRATEGY: QUANT-AS-A-SERVICE COULD BE A GAME-CHANGER 11
USE CASES OF AI IN ASSET
MANAGEMENT
There are multiple use cases for AI in the asset management industry — AI can be implemented
for portfolio management, risk management, generation of trading strategies, and better
management of investor wealth. The surge of ESG investing has brought forward another
interesting use case of AI. Exhibit 8, Exhibit 9, and Exhibit 10 highlight these use cases with some
examples of available products.
EXHIBIT 8: Use cases of AI in asset management
Application
Key areas Improvements over traditional methods
Some examples of
available products
Portfolio
management
º Hierarchical analysis to replace the mean–variance framework of Markowitz with tree based structure
which requires fewer estimates, capture all information in covariance matrix while obtaining more stable and
robust portfolio weights
º Using neural networks to make optimal asset allocation decision under complex multi-objective constraints
(e.g., value at risks constraints) while also incorporate views of future asset performance under Black
Letterman framework
º Evolutionary algorithms solve optimization problems under complex constraints (e.g., number of assets and
holding thresholds) and incorporate additional objectives (e.g., model risk)
1. BlackRock Alladin
2. Axyon IRIS
Trading
º Collecting/ analyzing big data more efficiently and conduct more insightful financial analysis to produce
stronger signal
º Machine learning/Deep learning to combine various technical analysis and fundamental analysis features
º NLP to process new, unstructured sources of data such as textual data from news articles or online sources
º AI approaches complement traditional market impact models by capturing non-linearity relations of
variables and identifying important factors related to order books
º Cluster analysis approach to identify comparable assets to conduct market impacts analysis for assets with
insufficient data history
º Reinforcement learning techniques learn to map each combination of input variables for order books,
known as a “state,” to trading actions such that transaction costs are minimized
1. Sentient Technologies
2. Kavout
3. Kensho
4. Bloomberg’s liquidity
assessment tool/Sentiment
tool
Risk
management
º NLP to extract information from textual or image data sources including news articles, online posts,
financial contracts, central bank minutes and statements, and social media to generate better predictions of
market crashes and other major macroeconomic outcomes
º Unsupervised AI approaches can detect anomalies in risk model output by evaluating all projections
generated by the model and automatically identifying any irregularities
º NNs can also be used to devise better systematic risk factors than conventional linear model. NNs models
can capture nonlinearities and interactions of covariates, including firm characteristics and macroeconomic
variables risk premia
º NNs and SVMs also perform particularly well when estimating loss given default (defined as the economic
loss when default occurs)
1. IBM Watson's Financial
Services Risk and
Compliance solution
2. FIS Global
3. Kx Systems
Wealth
management
º NLP to incorporate textual data and provide chatbots
º Incorporate machine learning predictions and make specific recommendations about investment strategies,
analyze portfolios, change asset allocations, and offer other proactive support based on a customer’s
investment goals and risk profile data, which is collected and processed through NLP/Chatbots
º AI powered relationship management tools for effective interaction as well as data collection
º Integrate all AI applications of portfolio management, trading, and portfolio risk management into Robo
advisors to produce portfolios with better out-of-sample performance for investors but also rebalance
portfolios, automatically managing the portfolio’s risks and
minimizing transaction costs while reducing behavioral biases
1. Schwab Intelligent
2. Betterment
3. Wealthfront
ESG
º NLP to look through unstructured data and scan for ESG specific keywords in patent filings
º NLP to process information from statements/conversations/transcripts that are sustainability-related and
gauge levels of the company’s commitment to mitigating climate risks
º Clustering analysis to pick up ESG related stocks for ESG-aligned portfolio construction
º Machine learning algorithm to assign sustainability score to portfolios or individual stocks
1. Clarity AI
2. Acadian asset
management's NLP model
3. Auto-CIO platform
Artificial Intelligence (including Machine learning) applications
Source: cfainstitute.org, Mckinsey, Bernstein analysis
BERNSTEIN
12 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
EXHIBIT 9: Alternative data is a valuable source for investment management — key focus for many quant investors
Source: caserta.com, The Book of Alternative Data, alternativedata.org, Bernstein analysis
EXHIBIT 10: Some examples of AL/ML applications in asset mangement
Source: Company websites, Risk.net, cfainstitute.org, Bernstein analysis
BERNSTEIN
ASIA QUANT STRATEGY: QUANT-AS-A-SERVICE COULD BE A GAME-CHANGER 13
SIMPLE EXAMPLE OF USING
AI TO IMPROVE PORTFOLIO
CONSTRUCTION
In our last joint work with fundamental analysts (Asia Quant Strategy: 30 stocks for the quarter
- Opportunities for Growth and Value investors), we recommended 30 stocks which are both
preferred by quantitative models and fundamental analysis. We show below how a simple
clustering analysis machine learning algorithm in this case can help in portfolio construction. The
portfolio picked by the clustering model was able to outperform the portfolio of all 30 stocks.
Here we demonstrate the application of the machine learning clustering algorithm on grouping
the Quant + Fundamental portfolio into clusters. Clustering analysis is a category of unsupervised
learning that allows the algorithm to discover hidden structures in data without providing any
guidance. We used K-means technique for the demonstration, which is widely used in both
academia and industry. K-means starts with centroids as cluster centers and group samples
based on feature similarities (measured by squared Euclidean distance): (1) randomly pick K
centroids from the samples; (2) assign samples to the nearest centroids; (3) move the centroids
to the center of the samples surrounded; and (4) iterate the above processes until the clusters
remain unchanged or reach iteration limits.
We applied K-means to the 30-stocks portfolio using historical performances as the feature. The
algorithm grouped the 30 stocks into 15 clusters, within which stocks are expected to perform
similarly. While there is no surprise when stocks in the same sectors are grouped together, the
clustering of stocks from different sectors can be insightful (Exhibit 11).
EXHIBIT 11: Our recent Quant + Fundamental portfolio (30 stocks) can be grouped into 15 clusters based on
historical performances
LARSEN & TOUBRO
AMBUJA CEMENTS
AIR CHINA
TRIP COM
GEELY PING AN
ICBC
CMB
ZTO
TENCENT
HUA HONG
PDD
WUXI BIO
SAMSUNG
ELECTRONICS
SUN PHARM
HANS LASER
KAKAO
ESTUN
KINGDEE
YONYOU NETWORK
NOVATEK
MEDIATEK
GREAT WALL MOTOR
-1%
0%
1%
2%
3%
4%
5%
0 2 4 6 8 10 12 14 16
arithmetic mean return
cluster group
Quant + Fundamental portfolio clustering analysis
Note: Clustering algorithm is K-means. Data history starts from July 2018 till March 2023. Some stocks in Q+F portfolio are eliminated due to limited price history.
Source: Bloomberg, Bernstein analysis
Effectiveness of the strategy: Effective use of a cluster algorithm could provide higher
rewards (Exhibit 12). For the 30-stocks portfolio, we constructed a simple strategy that utilizes
cluster analysis with quarterly and semi-annual rebalance frequency. On each rebalance date,
we established clusters using K-means and only keep stocks with the highest past 12 months
risk-adjusted returns while eliminating redundant stocks in each cluster (if there is more than one
stock in the cluster). The portfolio is then constructed based on remaining stocks with all stocks
being equal-weighted. The intention is to improve risk/return profile as well as diversification
BERNSTEIN
14 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
benefits by eliminating highly correlated stocks. The backtesting result in Exhibit 13 shows
the strategy using a clustering algorithm generates higher return at a similar level of volatility
and improves the overall Information Ratio (IR) from 0.43 to 0.45 (quarterly rebalanced)/0.5
(semi-annual rebalanced). The absolute returns of the semi-annual rebalanced portfolio based
on clustering recorded a 13% CAGR since 2018 versus 12% per year by quarterly rebalanced
clustering portfolio versus 10% per year for a simple buy and hold portfolio.
EXHIBIT 12: Clustering algorithm can quickly identify
pairs that performed similarly including those across
sectors
0
50
100
150
200
250
8/1/2018
11/1/2018
2/1/2019
5/1/2019
8/1/2019
11/1/2019
2/1/2020
5/1/2020
8/1/2020
11/1/2020
2/1/2021
5/1/2021
8/1/2021
11/1/2021
2/1/2022
5/1/2022
8/1/2022
11/1/2022
2/1/2023
Cluster 1 Historical Performance
AMBUJA CEMENTS LARSEN & TOUBRO
0
20
40
60
80
100
120
140
160
8/1/2018
11/1/2018
2/1/2019
5/1/2019
8/1/2019
11/1/2019
2/1/2020
5/1/2020
8/1/2020
11/1/2020
2/1/2021
5/1/2021
8/1/2021
11/1/2021
2/1/2022
5/1/2022
8/1/2022
11/1/2022
2/1/2023
Cluster 2 Historical Performance
AIR CHINA TRIP COM
0
50
100
150
200
250
8/1/2018
11/1/2018
2/1/2019
5/1/2019
8/1/2019
11/1/2019
2/1/2020
5/1/2020
8/1/2020
11/1/2020
2/1/2021
5/1/2021
8/1/2021
11/1/2021
2/1/2022
5/1/2022
8/1/2022
11/1/2022
2/1/2023
Cluster 7 Historical Performance
SUN PHARM SAMSUNG ELECTRONICS
0
100
200
300
400
500
600
8/1/2018
11/1/2018
2/1/2019
5/1/2019
8/1/2019
11/1/2019
2/1/2020
5/1/2020
8/1/2020
11/1/2020
2/1/2021
5/1/2021
8/1/2021
11/1/2021
2/1/2022
5/1/2022
8/1/2022
11/1/2022
2/1/2023
Cluster 13 Historical Performance
MEDIATEK NOVATEK
Source: Bloomberg, Bernstein analysis
EXHIBIT 13: A strategy using clustering algorithm that
picks up stocks with highest risk-adjusted returns
while eliminating redundant stocks in each cluster
improves overall IR compared with a simple buy and
hold strategy
0
50
100
150
200
250
6/1/2018
9/1/2018
12/1/2018
3/1/2019
6/1/2019
9/1/2019
12/1/2019
3/1/2020
6/1/2020
9/1/2020
12/1/2020
3/1/2021
6/1/2021
9/1/2021
12/1/2021
3/1/2022
6/1/2022
9/1/2022
12/1/2022
Total Performances of Strategy Using
Cluster Algorithm vs. Buy and Hold
Cluster strategy quarterly rebalanced
Buy and Hold
Cluster strategy semi-annual rebalanced
IR-0.45
IR-0.43
IR-0.5
Note: Portfolios are equal-weighted. All returns are in USD.
Source: Bloomberg, Bernstein analysis
QAAS COULD BE THE
SOLUTION TO INCREASE AI
ADOPTION AT A FASTER PACE
As pointed out in Exhibit 5, apart from model implementation challenges, talent constraint is one
of the big obstacles for faster AI adoption in the investment management industry, especially for
discretionary investment firms. Unlike specialized data-driven, technology-based quant firms,
which already have the infrastructure to implement AI techniques, fundamental investment firms
usually find it difficult to embrace the latest AI developments either due to a shortage of tech
resources or shortage of AI talent, not to mention the high costs to deploy those resources.
We therefore find QaaS as a key solution for faster AI adoption, especially for traditional
investment management firms. The idea is for companies to offer a full stack of quant solutions
as offerings, in turn reducing the need for asset managers to build the capabilities internally.
While there are many open sources and community-based platforms that have made life much
easier for people/firms with the right technical skills to use these products, we don't see many
companies providing end-to-end quant solutions. There are firms such as Fractal Analytics
and Evalueserve that provide AI-based solutions for multiple industries, including for financial
services, and look well-positioned to be a leader in the space. Another key player would be
Acuity KP, which provides cost-competitive services by setting up dedicated data science and
quant services for hedge funds. QAS (Quantitative Analysis Service Inc.) (private), which provides
multiple customized quant solutions for screening, idea generation, risk management, etc., and
BERNSTEIN
ASIA QUANT STRATEGY: QUANT-AS-A-SERVICE COULD BE A GAME-CHANGER 15
Hudson & Thames (private) are relatively smaller companies, but are trying to develop ready-to-
implement solutions. The point is: there are some players, but the idea of providing QaaS is not too
popular yet and has the capability to make a real difference in moving the investment community
toward AI.
We have highlighted the different types of quant solution providers and the ecosystem that
exists today (Exhibit 14 to Exhibit 17). Service providers are the closest to the idea of QaaS —
they either provide complete data-driven AI solutions to investment strategies, trading, portfolio
construction (QAS, Acuity KP, and Hudson & Thames) or specialize in certain areas including
NLP, textual information extraction, trading executions (Quantitative Brokers, Kensho, and Alpha
Sense) (all private). Community-based platforms usually provide great resources for fast
implementation of AI models for users who have basic technical skills. They usually offer research
environment, backtesting function, portfolio construction tools, and trading API for live trading
(Quantconnect, Numerai, and Quantiacs). In addition, they can also function as great education
hubs for quant investment or data science learning (Kaggle, a subsidiary of Google). Among
those platforms, those with crowdsourcing business models are worth looking into — investing
by utilizing the "wisdom of crowd" that adopt robust models from individuals with limited costs
(Numerai and Quantiacs). Firms with necessary expertise can also use open-source libraries for
fast model prototyping — there are libraries covering the entire chain of quant finance (Qlib and
Quantlib), trading (PyAlgoTrade) and backtesting (Backtrader). Firms can also seek for affordable
trading/data alternatives via trading/data API (Alpaca and Finnhub), which is also suitable for
individuals and small start-ups.
BERNSTEIN
16 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
EXHIBIT 14: QaaS — service providers
Logo Name Description QaaS inspiration
Quantitative Broker
Quantitative Broker provides research-driven trading
solutions to reduce transaction costs, trade complex
structures, measure execution performance and
dramatically improve workflow efficiencies. Its client
base includes the world's biggest asset managers, hedge
funds, CTAs, global banks, public and private companies
including the buy-side and sell-side.
QB offers custom algorithms to each instrument on
execution, analytics suite on market analysis and
automation solution on workflow, which can come in
handy for clients who lack necessary resources.
KENSHO
Kensho leverages S&P’s data to build machine learning
applications, specialized in natural language data,
including complex documents and speech, and builds
machine learning models that add layers of structure to
unstructured and semi-structured data.
Kensho is not dedicated to finance but the services on
NLP can be valuable for extracting information from
natural language data for investment decisions.
AlphaSense
AlphaSense uses machine learning and natural language
process extract insights from an extensive universe of
public and private content—including company filings,
event transcripts, news, trade journals, and equity
research. Clients include JP Morgan, Wells Fargo, Melvin
Capital etc.
Accelerate the financial research process by efficient
extraction of key insights from natural language sources.
Hudson & Thames
Hudson&Thames promotes the scientific method within
investment management by codifying frameworks,
algorithms, and best practices to offer ready to use
implementations and intellectual property. It offers
consulting as well as ready to use quant finance library,
all based on academic journals.
Hudson&Thames offers ArbitrageLab, MLfinLab and
PortfolioLab python libraries that target at pairs trading,
machine learning applications and portfolio
management. The algorithms are implementations
based on academic research and suitable for quick
deploy of high quality academic research ideas.
Acuity Knowledge Partners
Acuity KP is a leading provider of research, analytics and
business intelligence to the financial services sector. It
supports over 400 financial institutions and consulting
companies through specialist workforce of over 4,000
analysts and delivery experts across its global delivery
network.
Acuity KP offers AI/data science services to asset
managers, hedge funds as well as investment banks and
provides AI based products for cost management,
extract natural languages data and model building.
Quantitative Analysis Service,
Inc.
QAS has 40 years of experience providing alpha
generation solutions to institutional investors. Its
solutions include investment instrument selection, risk
management, ETF model portfolio building and
technology projects.
QAS is experienced in offering complete portfolio
solutions and also active in publishing research.
Evalueserve
Evalueserve aims to empower enterprise clients with AI-
driven products and solutions that optimize decision-
making and drive actionable outcomes. For financial
institutions, they offer AI solutions to investment
research and risk management.
Evalueserve has a strong global footprint and has AI
optimized products for insights generation, data
extraction and risk modeling.
Fractal Analytics
Fractal Analytics is a multinational artificial intelligence
company that provides services in consumer packaged
goods, insurance, healthcare, life sciences, retail and
technology, and the financial sector.
Fractal Analytics current AI solutions to financial sector
are retail banking focused, however, with their expertise
and scale, they can easily expand their AI footprints to
asset management.
QUANT AS A SERVICE (QaaS) Service Providers
Note: This is not an exhaustive list.
Source: Company websites, Bernstein analysis
BERNSTEIN
ASIA QUANT STRATEGY: QUANT-AS-A-SERVICE COULD BE A GAME-CHANGER 17
EXHIBIT 15: QaaS — platforms
Logo Name Description QaaS inspiration
Kaggle
Kaggle, a subsidiary of Google LLC, is an online
community of data scientists and machine learning
practitioners. Kaggle allows users to find and publish
data sets, explore and build models in a web-based data-
science environment, and enter competitions to solve
data science challenges. Although it is not dedicated to
finance, many hedge funds including Two Sigma, Jane
Street held trading strategy competitions on Kaggle.
A high profile data science platform that offers
education, community on best practice on machine
learning model building. The competitions are great
platforms to showcase skills.
Numerai
Numerai is an AI-run, crowd-sourced hedge fund.
Participants make machine learning models to submit
signals, Numerai then combines them all into one big
ensemble model (the Meta Model) for trading decisions.
Participants who provide accurate signals are rewarded
with Crypto.
Innovative business model that utilizes the "Wisdom of
Crowd" when themselves lack the expertise/personals.
The flagship fund has generated 52.39% return since
2019.
Quantiacs
Quantiacs is a crowd-sourced quant platform hosting
algorithmic trading contests and a marketplace serving
investors and quants. It holds regular contests and
winning strategy will be allocated capital for live trading
and receive 10% of the profit.
The platform offers complete developing ecosystem
including software, data, servers and can invest in
multiple assets including stocks, futures and crypto. The
platform is able to run diversified portfolios of strategies,
since only winning strategies will be deployed after
rigorous live tests, risks are diminished.
Cloud Quant
Cloud Quant is a high-performance quantitative research
platform, it serves alternative data providers by
providing a data showcasing service while also offering
data services to data buyers.
Users can utilize the platform for quant research as well
as purchase all kinds of data including various alternative
data.
QuantConnect
QuantConnect is an open-source, cloud-based
algorithmic trading platform for equities, FX, futures,
options, derivatives and cryptocurrencies. QuantConnect
serves over 100,000 quants from 170+ countries, with
customers including hedge funds and brokerages, as well
as individuals such as engineers, mathematicians,
scientists, quants, students, traders, and programmers.
QuantConnect provides market data, backtesting engine
and a cluster computer for trading strategies across
multiple assets, including equities, futures, options,
cryptocurrencies, CFDs and FX. QuantConnect
established a marketplace that provides the community
the freedom to license their alpha-generating insights to
quantitative funds. Hedge funds can leverage the
marketplace to search for algorithms that fit their
specific criteria and license them for a monthly fee.
QUANT AS A SERVICE (QaaS) Platforms
Note: This is not an exhaustive list.
Source: Company websites, Bernstein analysis
EXHIBIT 16: QaaS — open-source software
Library Description
Qlib
An AI-oriented Quantitative Investment Platform by
Microsoft. Full ML pipeline of data processing, model training,
back-testing; and covers the entire chain of quantitative
investment: alpha seeking, risk modeling, portfolio
optimization, and order execution.
QuantLib
The QuantLib project is a free/open-source library for
modeling, trading, and risk management, aimed at providing a
comprehensive software framework for quantitative finance.
It is written in C++.
PyAlgoTrade
PyAlgoTrade is a Python Algorithmic Trading Library with
focus on backtesting and support for paper-trading and live-
trading.
Backtrader Python Backtesting library for trading strategies.
QUANT AS A SERVICE (QaaS) - Open-Source Libraries
Note: This is not an exhaustive list.
Source: awesome-quant, company websites, Bernstein analysis
EXHIBIT 17: QaaS — trading APIs
API Description
Alpaca
Alpaca offers trading API to developers to trade with
algorithms, connect with apps, and build services.
Finnhub
Finnhub aims to democratize financial data and provide API
of institutional-grade financial data to investors, fintech
startups and investment firms.
QUANT AS A SERVICE (QaaS) - Trading/Data API
Note: This is not an exhaustive list.
Source: awesome-quant, company websites, Bernstein analysis
Rupal Agarwal rupal.agarwal@bernstein.com +65 6230 2358
BERNSTEIN
18 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
BERNSTEIN
Consumer
CONSUMER 19
BERNSTEIN
20 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
OTAS AND HOTELS: BEGINNING OF A
RAPID ADOPTION CURVE?
HIGHLIGHTS •AI chat technology is poised to bring revolutionary changes to future travel, breaking
the decade long "location/date" default search. Travelers face multiple pain points today,
from finding travel ideas to making reservations. AI has the potential to revolutionize the
travel industry by transforming travel platforms into users' personal assistant, leveraging
their access to service providers and massive user data. A mature chat AI can generate
a comprehensive itinerary with recommended flights, hotels, tourist spots, and in-city
transportation, with actionable links all in one go. Real-time user data enables AI to tailor
recommendations without further action from customers, enhancing the overall travel
experience.
•Cherry Leung: OTAs can fortify their competitive edge over direct booking channels
of hotels and airlines with the potential for AI to seamlessly connect all elements of
a traveler's journey. Trip.com launched TripGen on its mobile app, which provides instant
travel tips and itinerary suggestions in text format. Although still in the experimental stage, its
commercial potential is prodigious. Once this technology is scaled and aptly deployed, OTAs
can significantly improve the accuracy and relevance of their recommendations to travels. Yet,
OTAs will need to effectively integrate their preferences into the chatbot technology in a way
to create a mutually beneficial product solution for customers and suppliers.
•Richard Clarke: For Western OTAs, AI could allow Big Tech and new entrants to gain
traction, and impact the OTAs' existing upsell model. Big Tech (Google in particular) have
become a disruptive force to OTAs, allowing greater price discovery and disintermediation
of OTAs toward direct bookings. AI may compound this disruptive effect. Equally, AI may
allow new tech-forward entrants; examples such as Hopper (private), eDreams (private), and
Trip.com are already gaining share. A key challenge for OTAs will be reconciling a useful
recommendation tool with their upsell revenue tool, i.e., providing a long list of hotels and
charging to be at the top. In theory, the value of being the recommended hotel on a chatbot
would be enormous, but this would mean the platform was not using any "intelligence" but
just putting forward the highest commission paying hotel.
INVESTMENT IMPLICATIONS Trip.com (TCOM.US) — Outperform. We shouldn't get too excited about TripGen by Trip.com
at this stage as it doesn't close the loop from suggesting to booking. In the long term, leading
OTAs can leverage AI to uplift its "backbone" role in the travel ecosystem with their deep-rooted
relationship with travel operators.
Booking (BKNG.US) — Underperform; Expedia (EXPE.US) — Market-Perform. Recent
years have shown that OTAs can be disruptive either by top of the funnel disintermediation or
by tech-forward new entrants, and we see the advent of AI as equally likely to continue that
disruption as reverse it. Our overall thesis on Booking is that consensus is too optimistic on the
revenue and margin path, given this disruption.
BERNSTEIN
OTAS AND HOTELS: BEGINNING OF A RAPID ADOPTION CURVE? 21
Airbnb (ABNB.US) — Outperform. Airbnb is different — it plays in a rapidly growing vertical
where it benefits from a unique product set and loyal customers. Being at the top of its funnel
and offering a less filterable product, Airbnb is a more natural beneficiary of AI.
HOW IS AI BEING USED IN
TRAVEL SO FAR?
Although Trip.com is the first to the market with an AI chatbot built into an OTA app, travel has
been using some form of AI or machine learning for a while, notably to assist with service and
marketing.
•Service agents: Expedia launched an AI-powered service agent in June 2021 (Exhibit 1). This
was primarily a reaction to Covid-19 and the confusion around customers' right to cancel/
change bookings, or the eligibility of credit usage. In a period when trying to change bookings
over the phone could result in hours of waiting, such a service was well-received. eDreams,
the leading flight OTA, had a similar module that it said was able to deal with 50% of queries,
cutting down phone wait times for those that needed it and, it says, being a key driver of the
rapid expansion of its subscription service. Some chatbots can also arrange bookings and
make some suggestions — if the questions are specific enough.
•Listing suggestions: AI has also been used to rank suggestions to consumers. eDreams
started using AI in 2021 to rank suggestions to users (again mainly for subscribers). Its travel
search engine is supported by a set of algorithms that make predictions on travel proposals
that aim to better meet each traveler's needs, considering the user’s previous selections
together with dozens of additional criteria and data points as well as aggregated data from 1.7
billion monthly searches from users. As a result, 72% of users with a previous search made
now order the first recommended travel option provided to them via search. This represents
an increase of +17% since the technology was introduced at the end of 2021.
•Flight and hotel price forecasting: Hopper notably uses machine learning to predict the
future direction of flight and hotel prices, and monetizes this by selling price freezing and
insurance products.
•Advertising: Google may not be employing AI yet in a guest's hotel search, but it is using
AI to better target marketing campaigns for hotels. The AI-powered tool, called Power Max,
enables hotels to create advertisements in multiple formats and reach more travelers through
a single campaign across YouTube, Display, Search, Discover, Gmail, and Maps. According to
Google, Bangkok-based Minor Hotels used Performance Max for travel goals to accelerate its
campaign creation process and grow its reach on Google. The hotel group received budget
optimization recommendations per property, enabling it to cut down on cost per acquisition
by 51%, increase its return on ad spend by 76%, and increase bookings by 86%.
•Hotel service: IBM has developed Connie, a robot concierge for Hilton, which draws on
domain knowledge from Watson and WayBlazer to help hotel guests figure out what to visit,
where to dine, and how to find anything at the property (Exhibit 2). Through its partnership
with Josh.ai, IHG is implementing AI that understands voice commands. This way, customers
can make a simple request to play music, stream a show, and take other measures to make
themselves comfortable during their stay. Hyatt is upgrading its services by combining AI and
travel in the form of an AI-powered bed. The Bryte Restorative Sleep Bed relies on sensors to
track a patient’s heart rate and breathing, adjusting the temperature and firmness of the bed
to help sleepers enjoy a deeper doze. Customers can even set preprogrammed movements
BERNSTEIN
22 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
so the bed nudges them awake in the morning.
•Expense management: Auto-Itemization from Navan is an AI-powered expense
management solution that automatically splits transactions and attributes each line item to
specific expense policies. The technology allows a user to upload a receipt for automatic
itemization; using AI, foreign language translation and fuzzy matching, each line item is
categorized and assigned to a specific policy with >90% accuracy. The use of machine
learning also ensures that accuracy will increase over time.
•And now…a generative AI chatbot: Although chatbots have existed in the travel sector,
Trip.com's is the first to use the new generative-AI technology and therefore in theory has a
broader scope. Although it is in early stages, it likely shows we have entered the AI era of travel
bookings, and the implications are likely to be huge in coming years.
EXHIBIT 1: Expedia launched an AI powered service agent
in 2021 more to deal with post-booking queries
Source: Company website
EXHIBIT 2: Connie the robot concierge
Source: Hilton.com
AI TO TACKLE THE PAIN
POINTS IN TRIP BOOKING
Although AI has made some inroads into travel, there is far more it can do. Travel is a consumer
segment with a huge amount of fragmented choice, but not much assistance in where to go or
when best to travel. Most travel platforms rely on a user to have already decided where they
want to go and when before they even start searching. Those companies that have made any
attempt to fix this (e.g., Hopper's price prediction tool) have been proven share gainers. Designing
a trip might be a joy for travel enthusiasts, but in many cases it is painful — from finding new
travel ideas to making reservations, travelers often fall into the "tyranny of the search bar" by
repeating the process of searching, comparing results, and modifying searches. This is also a
major source of discontent with current OTA models, as they are unable to provide effective
itinerary recommendations. Airbnb and Google have been the first to try an "anywhere/anytime"
BERNSTEIN
OTAS AND HOTELS: BEGINNING OF A RAPID ADOPTION CURVE? 23
search option and have demonstrated customer desire for more inspiration.
Although we are very much in the early stage, AI chat technology is poised to bring revolutionary
changes to future travel. With AI, platforms can, over time, become users' personal travel
assistant, leveraging their access to service providers and massive user data. A mature chat AI
can in theory generate a comprehensive itinerary with recommended flights, hotels, tourist spots,
restaurants, and in-city transportation, with actionable links all in one go — likely also helping
to find the most cost-effective time and method of travel. Real-time and accurate user data
also enables AI to tailor recommendations without further action from customers, enhancing the
overall travel experience.
TCOM is well aware of users' pain points, and even before the introduction of TripGen, it already
made several attempts to help travelers out of the "tyranny of the search bar." On Trip.com mobile
app, travel tips of major tourism cities are provided with recommended experiences, hotels, and
restaurants. Despite combining the offerings in one page, the feature is only provided on the
basis of each city for users' reference, and users still need to go through traditional planning and
booking procedures. On the Chinese app of Ctrip, users can order customized trips with simple
inputs of the destinations, days of stay, and must-visit landmarks. The service is outsourced by
Ctrip to designers from third-party travel agencies, and a single price quote is provided shortly
with the package trip solution, including tickets, hotels, and chartered car services.
The potential of AI chat lies in the existing resources and information owned by OTAs, and without
them there is limited value and actionable items that can be provided. Among Chinese OTAs, Ctrip
is the most resourceful with extended service offerings and a premium customer base. Ctrip has
over 1.7 million hotels under its global coverage, far above the ~1 million of Fliggy (owned by
Alibaba, which is covered by Robin Zhu) and Meituan (covered by Robin Zhu). TCOM group, which
includes Ctrip and Trip.com, has over 50% high-end room nights mix compared with the industry
level of 20%, indicating its established relationship with high-end hotels (chains) and high-profile
customers who prefer quality and experiences over cheap pricing (Exhibit 3 and Exhibit 4).
BERNSTEIN
24 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
EXHIBIT 3: Ctrip has the largest number of hotels under
coverage, well-positioned to offer tailored trip advice
1,700
1,000
900
0
200
400
600
800
1,000
1,200
1,400
1,600
1,800
Ctrip Fliggy Meituan
2022 Global Hotel Coverage (K)
Source: Company websites, Bernstein analysis and estimates
EXHIBIT 4: TCOM hotel offerings are more skewed to
high-end, targeting affluent customers who value
experiences over cheapest pricing
53%
20%
13%
0%
10%
20%
30%
40%
50%
60%
TCOM Industry Meituan
High-end roomnight mix
~2.7x
Source: Company websites, Bernstein analysis and estimates
AI EMPOWERS OTAS TO
DEFEND AGAINST DIRECT
BOOKINGS
TCOM made exploratory attempts in AI
In January, TCOM launched TripGen on its Trip.com mobile app, which provides instant travel tips
and itinerary suggestions in text format. TCOM management commented in its latest earnings
call that TripGen is still at the early exploratory stage. The feature will play three important roles
in the future: (1) enable users to find relevant information efficiently; (2) link search results to the
products; and (3) assist the customer service team to improve the quality of services.
Currently, TripGen has integrated the OpenAI API without any customization for the OTA or
travel industry, neither does it connect to Trip.com's product offerings or customer information
database. It generates answers based on sources that are not verified, and the error rate is still
high. In the internal trial, TripGen does not abide by the policies on cancellation and returns and
becomes too "nice" in some cases. It is more like a knowledgeable friend rather than a customer
service tool, and sometimes advises users to book the flights on the websites of airlines or
competing platforms. TCOM believes it won’t be used to replace customer service for at least
five years. The conversation might be different when the feature is integrated with Microsoft's
Cloud service with access to products and users' data, though the timeline is still unclear.
TripGen currently supports English, Japanese, Korean, and traditional Chinese. As OpenAI is
better developed in English, it provides more comprehensive answers in English, and that is why
the feature is rolled out in Trip.com and not on Ctrip first.
On the market upheaval over AI, other OTAs also commented on how they can benefit from the
technology. Expedia claimed to be "by far the forefront in our category in terms of using AI," while
Booking said its capabilities are "as good as anybody else's." We hold a conservative view toward
the benefits of AI at the current stage because when everyone claims to be uniquely benefited, it
is likely that no one is more advantaged than others (Exhibit 5).
BERNSTEIN
OTAS AND HOTELS: BEGINNING OF A RAPID ADOPTION CURVE? 25
EXHIBIT 5: Unique advantage for everyone... = commoditization?
Company Commentary
Airbnb
Will uniquely benefit from AI due to private rental's heterogeneous product nature (vs. hotels).
Booking.com
Our capabilities are as good as anybody else's; we will adapt and do very well in these new technologies.
Expedia
We are by far at the forefront in our category in terms of using AI.
TripAdvisor
This is a technology that puts us in a place of advantage.
Source: Company reports, Bernstein analysis
COULD AI BETTER EQUIP OTAS
AGAINST DIRECT BOOKING?
Although the full effects of chat AI are yet to materialize, we have a positive outlook on its
commercial prospects in the long term. Besides the basic needs of accommodations and
transportation, customers are focusing more on all-round experiences, especially for leisure
trips. A survey on Chinese travelers indicates the top purpose of their next outbound trip is to try
local food, followed by experiencing local life, and beach and sea (Exhibit 6). This can be fulfilled
through a guided package tour at expensive charges. An AI-powered OTA is better-positioned to
provide tailored and overarching trip planning at affordable pricing, if not for free.
Direct booking has been a threat to OTAs for their capability of providing exclusive services, the
best price, and other booking benefits (Exhibit 7). An overwhelming trip experience at bundled
(and cheaper) price might dwarf the benefits of direct booking. As early entrants in AI, OTAs such
as TCOM are likely to create a virtuous circle, where the most tech-savvy customers use and train
the AI to become smarter which, in turn, attracts more customers. Once this technology is scaled,
the trend of using AI-assisted booking platforms will become irreversible.
At this junction of China reopening, TCOM's introduction of TripGen reflects its endeavor to
enhance product offerings, especially in overseas markets. China is an under-penetrated market
in terms of outbound tourism, with outbound tourism expenditure at merely1.8% of GDP in 2019,
dwarfed to most developed countries of above 2.5%. After three-year Covid-19 restrictions,
penetration is deemed to be even lower, and there is huge pent-up demand for Chinese to travel
abroad.
During the past three years, TCOM has strategically invested more in R&D and less in sales and
marketing. Its R&D expense remained above 40% of revenue and reached 45% in 2021, far
exceeding global peers such as Booking (3.1%) and Expedia (8.8%). S&M expenses was low at
22% in 2022, compared with Expedia (50%) and Booking (35%). TCOM's GTV dropped below
Booking and Expedia during the past two years, and its take rate has been low at 4% (versus
14% of Booking and 12% of Expedia). China domestic travel has recovered and outbound is still
suppressed by limited flight capacity. We believe TCOM's continuous investment in products and
customer experiences will pay off on full-fledged travel resurgence.
BERNSTEIN
26 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
EXHIBIT 6: Local experiences (food, life) are the main
purpose of Chinese travelers' next outbound trip
Source: Dragon Trial International, Bernstein analysis
EXHIBIT 7: Extra services is the top reason for Chinese
travelers to book direct
28%
21%
21%
18%
12%
What is the top reason you book direct?
To ask for extras Best price
To negotiate a better rate Loyalty programs
Booking benefits
Source: Hostpitalitynet, Bernstein analysis
THE ONSET OF AI COULD
CONFLICT WITH OTAS' CORE
BUSINESS MODEL AND LEAD
TO NEW ENTRANTS
This section focuses on Western OTAs, which operate in a relatively slow growing and more
competitive market. Although we expect the OTAs to be a large part of the conversation on the
usage of AI, we also see some challenges:
1. Risk of top of funnel disintermediation. Trip.com's ChatBot is notably not a proprietary
tool; it is licensed from ChatGPT, which is partly owned by Microsoft. Google is also creating
its own AI product — Bard. The two tech leaders of AI within the Western markets, Microsoft
and Google, are making strides in travel even before AI is widely adopted. Google has been the
leader here — with D-Edge (private) (a channel manager) saying its hotels have seen a 95%
increase in revenue coming via Google Hotels ads from 2019 to 2022 (Exhibit 8). Notably,
Google seems to be focussed on growing the usage of its platform rather than profitability,
introducing free listing in 2021 and making this more visible on maps and searches in 2022.
These have lowered the average cost of booking, but the total revenue continues to increase
(Exhibit 8 shows this dynamic for D-Edge distributed hotels). Microsoft has been less active
in travel in recent decades since selling Expedia in 1999, but it started a partnership with
Amadeus (covered by Alex Irving) in 2022 and is speaking at the 2023 Phocuswright travel
conference — suggesting a deeper interest in the sector. If Google and Microsoft, proven
disintermediators even pre-AI, keep the best tech to themselves and continue to invest in
travel, it would suggest further risk to OTAs.
2. An effective AI OTA would send bookings direct. In our view, there are advantages to
a customer booking direct, both in price and in service: for hotels, points, digital check-in,
digital concierges, and room upgrades are largely reserved for direct bookers; for flights,
"choose your seat." Therefore, an effective AI system by Google, Bing, or any other should
likely highlight that there are advantages in booking direct. Even TripGen admits there are
advantages to booking direct. In fact, the main argument for using an OTA over booking direct
BERNSTEIN
OTAS AND HOTELS: BEGINNING OF A RAPID ADOPTION CURVE? 27
is likely to be the visibility, trustworthiness, and functionality of the direct website, which could
be fixed by AI.
3. Risk of new entrants. Recent years have shown that innovation can allow new entrants
to come into the OTA market. Our market share analysis (Exhibit 10) shows market share
gains for Hopper, eDreams, and Trip.com from 2019 to 2022 — all three have come up
with innovations and have used AI/machine learning to gain an edge. Hopper uses machine
learning in its price prediction algorithm and monetizes this via fintech products, eDreams
has been a pioneer in travel subscriptions, and trip.com has taken the lead on AI chatbots.
Notably, all these companies have also signed B2B relationships with Expedia and/or Booking
to bolster their supply. These B2B relationships (OTAs & Uber: Splitting the fare - what is a
B2B OTA partnership?) have lowered the barrier to entry for a new OTA and, therefore, we
expect we will see further new entrants that use generative AI, and B2B partnerships could
be disruptive.
4. Risk to OTA revenue model. One of the challenges for OTAs will be reconciling a useful
recommendation tool with their upsell revenue tool, i.e., providing a long list of hotels and
charging to be at the top. We would estimate that ~10% of Booking's revenue and ~20%
of EBITDA is from sponsored listings. In theory, the value of being the recommended hotel
on a chatbot would be enormous, but this would mean the platform was not using any
"intelligence," but just putting forward the highest commission paying hotel. In Western
markets, OTAs have faced disruption from Google, largely because Google has constructed a
result-agnostic search platform (it gets paid the same wherever you book), which has resulted
in better choice and price discovery for consumers (Exhibit 9). AI could further challenge this
model.
EXHIBIT 8: Google has become a meaningfully higher
source of revenue for hotels; D-Edge-controlled hotels
have grown their exposure by 95%
100
129
66
2019 2022
Indexed revenue from Google Hotel Ads
Booking revenue generated via Free links
Booking revenue generated via Paid links
+95%
Source: D-Edge, Bernstein analysis
EXHIBIT 9: On a per booking basis, Google has become
meaningfully better value, but its higher volume means
more revenue
9.0%
5.2%
100
113
92
94
96
98
100
102
104
106
108
110
112
114
0.0%
1.0%
2.0%
3.0%
4.0%
5.0%
6.0%
7.0%
8.0%
9.0%
10.0%
2019 2022
Google's commission and revenue
Average commission rate per booking on Google Hotel
Ads
GHA Indexed Revenue
Source: D-Edge, Bernstein analysis
BERNSTEIN
28 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
EXHIBIT 10: Hopper, eDreams, and Trip.com have gained share in OTAs in recent years; all three have used
innovation and AI (and B2B relationships) to fuel this performance
6.1%
2.2% 1.5% 1.3% 0.5% 0.0%
0.0% 0.0% -0.3% -0.4% -0.5% -0.8% -0.8% -0.9% -1.0%
-6.8%
Hopper
Booking.com
Trip.com
eDreams
Expedia
ebookers
ebookers Hotels & Flights
Priceline
HotelsCombined
Tripadvisor
Agoda
Kayak
Orbitz
Travelocity
Agoda
trivago
Global market share of app usage - 2022 vs 2019
Note: Hopper and eDreams are private companies.
Source: Apptopia, SimilarWeb, Bernstein analysis
WHO WINS? LIKELY DEPENDS
ON WHO HAS THE MOST
USEFUL DATA
Whether OTAs or other tech players win will ultimately depend on who has the best customer
data to feed into the AI tool and provide the best offer.
Expedia Group CEO Peter Kern said in 2020 that "Expedia Group will always have the advantage
when providing valuable data to partners because Google is at the top of the proverbial funnel,
offering data merely about clicks, while Expedia can offer more meaningful transaction data.
Expedia has more in-depth data about the dates of stay, and pricing."
The counterpoint is that the bigger tech firms can look across multiple channels, as Wolfgang
Krips, Senior Vice President, Corporate Strategy, Amadeus, said in a joint report with Microsoft:
"When a person travels, they go through several channels, generating data each time. Our goal is
to create a link between each of these data pools and to allow our customers to extract relevant
analysis, ensuring compliance at all times with applicable legislation and the highest security
standards."
Equally, hotels and airlines are likely to be more willing to share their data with Google or Microsoft
if those channels continue to send traffic to their website, more than with OTAs. As Shane
O’Flahery, Global Director of Travel, Transportation and Hospitality for Microsoft added: "These
technical elements, when combined, are the 'secret sauce' we are working to create in order to
increase the value of every trip for every traveler, which in turn can lead to incremental revenue for
the provider. While collating this variety of data does pose technical challenges, once overcome,
the rewards have the potential to revolutionize the entire sector. I think Amadeus and Microsoft
can create solutions that will add more value services, reduce cost and support the needs of
travelers."
BERNSTEIN
OTAS AND HOTELS: BEGINNING OF A RAPID ADOPTION CURVE? 29
Airbnb best placed
Airbnb is likely one company which can more seamlessly include AI in its offer. It has three
advantages over other platforms: (1) it is at the top of the funnel and has a unique product
set, which has made Google and other Big Tech less of a threat, (2) it has a flat commissions
structure, so will worry less about a change in model affecting upselling, and (3) its product is less
intrinsically filterable, i.e., it is harder for a consumer to know what it wants and where it wants to
go, which makes any offer that helps to suggest more valuable.
Airbnb has already started on this process: (1) It uses machine learning for ordering its properties
to best match the consumer and the host, i.e., in how it orders its listings (see here), and (2) It
has been a pioneer in making suggestions to guests, through its "categories"-based search. This
uses machine learning to categorize listing into various buckets (see here).
However, it is likely that Aibnb can go on further matching hosts and guests, as CEO Brian Chesky
said at the recent Q4 results:
"I'm actually very excited about the possibilities of AI. I think Airbnb will uniquely benefit from
this. And the reason why it's because Airbnb is a fairly difficult product challenge, which is unlike
hotels, we don't have SKUs. There's no representative inventory. Every single 1 of our 6.6 million
listings are unique guests left more than 100 million reviews last year. And to parse through all
these reviews is very laborious. And I think that AI is going to really benefit our long tail of data. And
the fact that our search problem isn't really a search problem, so much as a matching problem.
-- right? If there's like 50,000 homes in a city, what's the right 1 for you, that's less of a search
problem than a matching problem. And I think that AI is going to be a really good opportunity for
us. And just stay tuned for some developments there."
Airbnb management clearly sees matching of hosts and guests as its key USP, and will likely
continue to use AI to achieve this.
VALUATION METHODOLOGY See Disclosure Appendix for valuation methodology.
EXHIBIT 11: Ratings and target prices
25-May-2023 Target
Ticker Rating Currency Closing Price Price
TCOM.US O USD 31.21 43.00
BKNG.US U USD 2,597.37 2,100.00
EXPE.US M USD 95.79 96.00
ABNB.US O USD 104.42 135.00
MXAPJ 503.58
SPX 4,151.28
Source: Bloomberg, Bernstein analysis and estimates
RISKS See Disclosure Appendix for risks.
Cherry Leung cherry.leung@bernstein.com +852 2918 5756
Richard J. Clarke richard.clarke@bernstein.com +44 207 170 0536
Pearl Xu pearl.xu@bernstein.com +852 2918 7884
Kate Xiao kate.xiao@bernstein.com +44 207 170 5121
Niall Mitchelson niall.mitchelson@bernstein.com +44 207 170 0674
BERNSTEIN
30 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
OPPORTUNITIES FOR AI IN US
RESTAURANTS
HIGHLIGHTS The restaurant industry has lagged other industries on tech innovation due to low
margins, misalignment of incentives between franchisee and franchisors, and relatively
low labor costs. Historically, the restaurant industry has spent 2.5% of revenue in tech
investments versus an average of 8.2% for other industries, as the cheap labor cost (25%
cheaper than the closest industry) did not justify investments to improve (top line or bottom line)
productivity.
Profitability pressures and the reduced gap between labor costs and tech costs are
inducing restaurants to look for long-term solutions to restore their profitability; we
believe AI will finally find its place in the restaurant ecosystem. Multiple use cases are
arising with this increased focus. We believe self-service kiosks, drive-thru automation, and AI-
led order accuracy detection systems have the highest potential in the restaurant industry, as
we expect them to scale faster and generate more meaningful EBIT impact. Other noticeable
applications include AI chatbots, AI voice ordering, and AI-led personalizations of marketing and
customer engagement (in the front of the house) and inventory management, kitchen automation,
and staff management (in the back of the house).
We expect AI to augment the competitive advantages of restaurants with digital culture,
favoring Yum! Brands and Chipotle. Scale is an important factor in ensuring maximum return
from tech investments, while "operating" the stores offers more control over the speed of the
rollout of innovation. But that it is not the reason why we expect Yum and Chipotle to be the
winners of the digital (r)evolution in restaurants. We believe the secret sauce for digital success
is creating a digital culture that permeates all levels of the organization and is consistent over
time. We also have high expectations on Domino's, but look forward to further signs of restoring
their leadership innovation.
INVESTMENT IMPLICATIONS We rate four companies Outperform: Chipotle Mexican Grill (target price: US$2,300), Darden
Restaurants (target price US$180), Wendy's (target price US$27) and Yum! Brands (target
price US$144). We rate three companies Market-Perform: McDonald's (target price US$300),
Starbucks (target price US$102), and Restaurant Brands International (target price US$70). We
rate Domino's Underperform (target price US$275).
I. WHAT IS THE STATE OF THE
DIGITAL LANDSCAPE IN THE
RESTAURANT INDUSTRY?
Operating a restaurant is one of the most complicated tasks — a challenging endeavor that
needs a perfect dance from the back-end (purchasing the right vegetables, meats, etc.) to kitchen
operations (having the right cooks to cook the right meal with the right taste, every time) to front-
end experience (serving the right customers the right food, with the right service standards).
While one would imagine a high level of technology investment to streamline this multitude of
simultaneous tasks, technology investments in the restaurant industry have historically lagged
other industries (Exhibit 1).
BERNSTEIN
OPPORTUNITIES FOR AI IN US RESTAURANTS 31
We believe there are three main reasons for this underinvestment in technology:
1. Tight margin structure: Restaurant operators have historically witnessed a tight margin
structure of 5-10% (Exhibit 2), leaving very little wiggle room for investments in technology.
Consequently, restaurant operators have historically lagged other sectors in investing in
technology.
2. Low labor costs: Unlike other industries, the restaurant sector typically employs a higher
proportion of semiskilled workers on minimum wage. Consequently, restaurants employ
workers at a relatively low cost (Exhibit 3) and find it more cost feasible to drive solutions by
employing more people versus investing in technology.
3. Burden of investment is on franchisees: The restaurant industry predominantly operates
on the franchisor-franchise model, comprising numerous small, fragmented franchisees
running store operations with the branding, marketing, and support of large brands.
Consequently, the burden of investing in technology lies on franchisees who often are willing
to trade-off long-term technology investments in favor of near-term store profitability. As
a result, even though franchisors continue to invest in technology at a brand level and
demonstrate the operational effectiveness of new technologies at company-owned stores to
convince franchisees to invest, franchisees may decide not to invest in the technology, as they
often have full discretion on the investment decision.
Today, profitability pressures and the reduced gap between labor costs and tech costs (Exhibit 4)
are inducing restaurants to look for long-term solutions to restore their profitability; we believe
AI will finally find its place in the restaurant ecosystem.
EXHIBIT 1: Restaurants' technology investments are limited
Source: Flexera 2020 State of Tech Spend Report, Hospitality Technology Restaurant Technology Study (2016)
BERNSTEIN
32 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
EXHIBIT 2: Restaurants operate on a tight margin structure of 5-10%, further compressed by interest charges on a
usually levered balance sheet
US Franchisee (USD '000s) TB PLK MCD DPZ WEN BK KFC PH
AUV $1,941.0 $1,712.1 $3,595.5 $1,370.6 $1,874.6 $1,464.1 $1,300.3 $950.0
COGS $449.6 $566.6 $1,125.0 $424.7 $613.0 $452.4 $407.1 $270.5
Labor $559.5 $465.4 $1,075.7 $412.3 $601.7 $494.9 $394.1 $285.7
Royalties $119.3 $85.6 $300.1 $88.4 $75.0 $65.9 $50.6 $45.9
Ad funds $82.5 $83.5 $143.8 $82.2 $93.7 $58.6 $58.5 $45.1
Occupancy and Other expenses $353.5 $291.0 $502.0 $225.9 $307.4 $252.4 $266.3 $215.7
EBITDA / store $376.7 $220.0 $448.8 $137.0 $183.7 $140.0 $123.7 $87.1
% of Sales
COGS 23.2% 33.1% 31.3% 31.0% 32.7% 30.9% 31.3% 28.5%
Labor 28.8% 27.2% 29.9% 30.1% 32.1% 33.8% 30.3% 30.1%
Royalties 6.1% 5.0% 8.3% 6.4% 4.0% 4.5% 3.9% 4.8%
Ad funds 4.3% 4.9% 4.0% 6.0% 5.0% 4.0% 4.5% 4.8%
Occupancy and Other expenses 18.2% 17.0% 14.0% 16.5% 16.4% 17.2% 20.5% 22.7%
EBITDA / store 19.4% 12.8% 12.5% 10.0% 9.8% 9.6% 9.5% 9.2%
Source: FDDs, company disclosures, Bernstein analysis and estimates
EXHIBIT 3: Restaurant workers earn significantly less
compared with workers from other industries
47.7 42.1 39.0 36.5 35.1 32.5 31.1 28.9 27.9 26.1 19.4
Avg Hourly Earnings ($/hr)
Data as of September 2022
Source: BLS, Bernstein analysis
EXHIBIT 4: Restaurant wage growth has far exceeded
IT, hardware, and services inflation, making tech
investments more palatable
0
50
100
150
200
Jan-06
Sep-06
May-07
Jan-08
Sep-08
May-09
Jan-10
Sep-10
May-11
Jan-12
Sep-12
May-13
Jan-14
Sep-14
May-15
Jan-16
Sep-16
May-17
Jan-18
Sep-18
May-19
Jan-20
Sep-20
May-21
Jan-22
Sep-22
CPI Index Evolution: IT, Hardware and Services vs. Restaurant
Wages
CPI: IT, Hardware, Services CPI: FAFH Wage
Source: BLS, FRED, Bernstein analysis
II. HOW IS AI BEING USED IN
RESTAURANTS SO FAR?
Over the last few years, the restaurant industry was forced to adapt technology to the meet the
rapidly changing consumer preferences driven by the need for digital ordering during Covid-19,
the rise of aggregators, and the disruption from virtual brands and ghost kitchens. As a result,
many restaurants have upped their ante in technology investments and digital orientation with
larger chains leading the charge (Exhibit 5). Below are some key current and planned use cases
for AI.
Front of the house use cases for AI:
•AI-led drive-thrus: AI can significantly ease drive-thru chaos — from lane management,
order taking via voice recognition, auto product recommendation based on past order history
and orders from similar customers, and effectively communicating the order to the back of
the house. In fact, McDonald's acquired Apprente, a startup that uses conversational AI to
automate voice-based ordering across multiple languages in 2019, and has tied up with IBM
BERNSTEIN
OPPORTUNITIES FOR AI IN US RESTAURANTS 33
(covered by A.M. (Toni) Sacconaghi) to start rolling out the technology starting with ~10 drive-
thrus in Chicago (McDonald's highlighted its AI-led drive-thru goals here).
•For order accuracy, we believe AI could halve order inaccuracies, which represent ~15%
of customer orders. Since a store takes ~50 additional seconds on average to fill an
inaccurate order, we expect the impact for the industry to be in the order of US$9-US
$10Bn, assuming a US$7.6 average ticket on QSRs (per Euromonitor data) and ~200K
drive-thru stores in the US.
•For speed of service, we believe AI voice recognition and digital-only lanes could speed up
the average drive through service time by at least 20-30%, enabling the industry to reach
higher speed of service levels than pre-Covid-19 (Exhibit 6), an estimated impact of US
$5Bn for the industry.
•On upselling opportunities, we estimate that the lack of automation costs the industry US
$2Bn. Relying on staff to make recommendations leads to uneven results (Exhibit 7) and
leaves brands exposed to the risk that franchisees are not putting enough effort in training
and controlling the operations.
EXHIBIT 5: Digital sales are still relatively low — there is
significant scope for technology penetration
67%
50%
40% 37% 36% 33%
10% 10%
Domino's* Starbucks Yum* Chipotle McDonald's RBI Darden Wendy's
Digital Sales as % of Total Sales (4Q22)
*Most recent fiscal year
Source: Company reports, Bernstein analysis
EXHIBIT 6: Using AI could drastically improve speed of
service, reducing it beyond pre-Covid-19 levels
327
357
382
372
2019 2020 2021 2022
Speed of Service (in seconds)
Source: QSR Magazine, Intouch Insight, Bernstein analysis
EXHIBIT 7: Significant scope for AI-led upselling versus relying on uneven employee-based suggestive selling today
72%
62% 62% 54% 48% 44% 41% 32% 23%
10%
Suggestive Selling Penetration by Brands (%)
Suggestive Selling (%) Industry Average (%)
Industry Avg. = 42%
Source: QSR magazine, Intouch Insight, Bernstein analysis
•AI-led suggestive selling kiosks: Self-operated kiosk-based ordering and checkout has
significantly eased front of house operations, enhanced customer experience, and improved
systematic suggestive selling. With the rise of kiosk sales, restaurants are now utilizing
AI to analyze past customer orders, use order similarities with other customers, optimize
BERNSTEIN
34 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
current store inventory levels, and incorporate current weather patterns and time of day to
recommend items/product combinations to customers as they order. This usage of AI could
significantly uplift average checks in a frictionless upsell/cross-sell engagement process.
With kiosks rolled out across more than 40% of its units, McDonald's has collaborated with
Dynamic Yield (owned by Mastercard; Mastercard is covered by Harshita Rawat) to use AI to
prompt order recommendations (Exhibit 8).
EXHIBIT 8: McDonald's uses AI to recommend products based on past order data, similar orders from other
customers, weather/daypart data, and store-level dynamics
Source: The New York Times, Bernstein analysis
•Personalization of marketing and engagement: The restaurant industry has a very
peculiar advertising and marketing model, whereby individual, small franchisee operators
contribute a small percentage of their sales (usually at or below 5%) to their respective
brand-oriented ad fund to facilitate national and local advertising. These pooled ad funds can
significantly benefit from leveraging data-driven insights and AI-led predictions/allocations
of ad spends. For example, Yum! Brands recently acquired Kvantum, which uses AI to optimize
marketing. In fact, deploying AI at Pizza Hut UK led to ~40% return on ad spends, prompting
Yum to expand coverage to 60 brand-country combinations that cover 60%+ of system-wide
sales. Besides, AI is also being used to kick-start personalization of customer engagement.
For instance, the KFC MENA team uses AI to generate ~500 creatives/minute to personalize
outreach to its customers, which has resulted in 20% growth in digital transactions and ~24%
growth in average ticket.
•Chatbots replacing phone interactions: Chatbots have become one of the foremost
applications of AI in the restaurant sector, helping consumers order from the ease of their
home/website/app/messengers/social media and at their own convenience, while reducing
the burden on restaurant bandwidth and easing the biggest choke point in throughput
improvement. Domino's was the first to introduce its chatbot in 2016, making the complicated
pizza ordering process extremely simple for its customers (Exhibit 9). Recently, Yum! Brands
acquired Tinct, a conversational ecommerce AI platform to streamline ordering across all its
brands — it has since rolled out TicTuk's platform at 3,200+ stores across 49 markets (Exhibit
10). We expect that improvement in natural languages, scaling of capabilities of AI chatbots,
BERNSTEIN
OPPORTUNITIES FOR AI IN US RESTAURANTS 35
and integration with brand-specific digital assets can unlock further opportunities for the
industry.
EXHIBIT 9: Domino's chatbot has significantly eased
the often complicated and customized pizza ordering
process
Source: Domino's, Bernstein analysis
EXHIBIT 10: Yum acquired and rolled out its Tictuk
conversational ecommerce AI to streamline ordering
across all its brands
Source: Company reports, Bernstein analysis
•Voice ordering: What began as an AI phone call assistant has now spawned into a virtual
assistant for taking orders from any voice-led device (think Amazon's Alexa/Google Home).
Domino's launched the industry's first AI-led voice assistant, DOM, in 2014. However,
compute and network costs related to DOM's AI compute have led to low economic feasibility
for franchisees and limited rollout to pilots. With rapidly falling compute and network costs,
Domino's and many others across the industry are now expanding their investments in this
technology and scaled rollout could happen soon.
•AI-inspired menus: The next evolution is AI-led digital menu boards that will recommend
products on digital menu boards depending on product inventory, kitchen capacity at time
of ordering, weather patterns, and daypart rush — or as the industry depicts it — to conduct
"AI-driven suggestive selling." McDonald's rolled out its digital menu boards with AI-led
suggestive selling (Exhibit 11) and has witnessed significant increment in check sizes. AI-
generated menus (Exhibit 12) could also accelerate the scaling of virtual brands and reduce
the time to market for new menu introductions.
BERNSTEIN
36 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
EXHIBIT 11: McDonald's digital menu boards now double
up as upselling agents
Source: Company reports, nrn.com, Bernstein analysis
EXHIBIT 12: AI-generated menus are accelerating
digitization of menus
Source: ai.lunchbox.io, Bernstein analysis
While AI has significantly assisted in front of house activities, AI-led applications have and will
continue to deeply impact the operational and financial aspects of back of the house activities.
Back of the house use cases for AI:
•AI-led inventory management: Supply chain management is one of the key pillars of
good restaurant operations. Having the right vegetables/meats/condiments/oil in the right
quantity at the right time is essential for service. In fact, Chipotle's commentary highlighted
"how not having the right products" led to throughput issues, which eventually hampered
comp growth. Restaurants often over-order to cover-up for potential lack of products, so
much so that 4-10% of food purchased by restaurants never reaches consumers. With
food costs at ~30% of sales, restaurants can save ~1.2-3% in margins by streamlining
supply chain operations and order management. Recent advancements in AI have helped
link the back of the house to the front of the house. For example, Yum! Brands uses its AI
application "Recommended Ordering" that predicts and recommends the quantity of product
for a restaurant manager to order each week with the goal of reducing product waste and
intra-store transfers of inventory. Yum has rolled out this product to ~3K US stores across
Taco Bell and KFC.
•Kitchen automation: AI-led robotics and IoT sensors can significantly alter the experience
of cooks and line managers, taking the burden of mundane tasks and allowing them to
focus on their culinary expertise, and improving product throughput and order accuracy, thus
generating a win-win scenario for operators, employees, and customers. Restaurants are
already using AI-led Bluetooth temperature sensors that are placed both in chicken walk-in
coolers and hot holding cabinets to make sure that chicken is served at the right temperature
and meets safety standards. More recently, restaurants have started incorporating AI-led
advanced robotics, e.g., Chipotle launched a pilot for Chippy, a robot developed by Miso
Robotics (not covered) that can make tortilla chips (Exhibit 13).
BERNSTEIN
OPPORTUNITIES FOR AI IN US RESTAURANTS 37
EXHIBIT 13: Chipotle's Chippy for tortilla chips
Source: Engadget
EXHIBIT 14: Chipotle's Hyphen Makeline for digital orders
Source: Food on Demand
•AI-led throughput improvement initiatives: One of the key concerns for growing/
successful restaurant brands is throughput management/improvement. AI-led robotics/
makelines can help restaurants significantly improve their throughputs by identifying process
bottlenecks, monitoring throughput, and streamlining order preparation via technology. For
example, Chipotle's new Hyphen Makeline (Exhibit 14) continuously monitors the quality of
ingredients, automatically sends refill orders when quantities reach threshold levels, and also
prepares the dishes by serving and mixing the ingredients in pre-fed quantities. For a growing
company such as Chipotle, improving the throughput even by five entrées during peak hours
can contribute to an additional 1% in comp.
•Improvement of order accuracy: Order accuracy has been one of the biggest issues with an
average accuracy rate of ~85% (drive-thru order accuracy taken as a proxy, Exhibit 15). AI-led
vision analytics (Exhibit 16) can significantly improve order accuracy by doing a quick quality
check, image-check versus item ordered, and trigger human intervention for inaccuracies
before the order leaves the counter. Domino's-deployed DOM Pizza Checker (video here) uses
AI and vision scan technology to identify pizza type, topping distribution, product quality, and
order accuracy (correct toppings, pizza type, etc.).
EXHIBIT 15: Drive-thru order accuracy highlights one of the biggest service pain points in the restaurant industry
89% 89% 87% 86% 85% 84% 83% 83% 82% 79%
Order Accuracy (%)
Order Accuracy (%) Industry Average (%)
Industry Avg. = 85%
Source: QSR Magazine, Intouch Insight, Bernstein analysis
BERNSTEIN
38 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
EXHIBIT 16: Vision AI could significantly assist in order preparation and accuracy
Source: Precitaste, Bernstein analysis
•Staffing and optimized labor scheduling: Besides food costs, labor costs represent the
largest cost item for any restaurant. With increased labor challenges, high labor turnover,
and the continuous rise in the need for more flexible staffing, restaurants could benefit
significantly by leveraging AI tools to optimize labor scheduling, recruiting, and retention.
Recruiting is one of the low-hanging areas — in a labor market with significant issues in filling
roles (Exhibit 17), AI-led virtual assistants can accelerate the employment process by taking
over basic application processes such as indexing resumes, matching skills, and scheduling
interviews, making it easier to recruit talent. For example, Wendy's uses its virtual assistant
Lou to assist it in its recruitment process (Exhibit 18). Besides, AI tools are also being used
to optimize staff management. For example, Yum acquired [tracks] and uses its robust AI
platform to manage staffing and scheduling in its restaurants.
•Other upcoming AI applications — store automation and autonomous delivery: While
still in its infancy, AI can radically change the way restaurants operate with robotics and make
autonomous operations an integral part of restaurants. For example, Domino's was testing a
fully autonomous delivery vehicle Nuro (video here), while McDonald's just recently opened its
first fully-automated store in Fort Worth, Texas that could lay the blueprint of how technology
could reshape the restaurants of tomorrow (video here).
BERNSTEIN
OPPORTUNITIES FOR AI IN US RESTAURANTS 39
EXHIBIT 17: The increase in wages was driven by a persistent staffing challenge: Job openings for accommodation
and food services industries hit an all-time high
286
1,670
0
200
400
600
800
1,000
1,200
1,400
1,600
1,800
Accommodation and food services job openings
('000)
+107%
+82%
Source: BLS, Bernstein analysis
EXHIBIT 18: Recruitment AI assistant Lou deployed by Wendy's
Source: Wendy's, Paradox.ai
III. WHICH AI APPLICATION
HAS THE HIGHEST
POTENTIAL?
Assessing the many AI applications restaurants have adopted/could adopt, we believe there are
only a few "must haves" that will drive the maximum impact with a long-tail of numerous "good-to-
haves." Based on our estimates and expert interviews, we believe the AI innovations most likely to
scale in the next five years are: (1) inventory and supply chain management, (2) staff management,
(3) order management, (4) self-service kiosks, and (5) drive-thru automation. AI applications in
these areas will represent a significant leap forward in driving comparable sales (via enhanced
customer experience, ordering convenience, order accuracy, throughput improvements, etc.)
and reducing the major costs (labor costs, commodity costs, throughout improvements, etc.).
The restaurant industry lags in the use of AI in these critical areas and can benefit from greater
adoption of AI tools (Exhibit 19 and Exhibit 20). For example, inventory management AI tools can
BERNSTEIN
40 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
save food wastages (usually around 4-10% of food ordered by restaurants) that can at the very
least result in a margin uplift of ~1.2% (food costs are ~30% of restaurant sales, Exhibit 2).
EXHIBIT 19: AI application in top five critical areas is
significantly low — there is high scope for improvement
18.8% 18.0% 17.1% 17.0%
9.0%
Order
Management
Self-service
Kiosks
Staff
Management
Inventory
Management
Drive Thru
Automation
Automation/AI Adoption %
Source: Restaurant Readiness Index, PYMNTS.com, Bernstein analysis
EXHIBIT 20: Five critical areas drive maximum impact in
the shortest time
Source: Bernstein analysis and estimates
IV. IF AI SCALES, WHO WINS? a. Restaurants that have the largest scale?
Scale is a critical parameter to determine the success of AI: the high initial tech costs spread
across a large ecosystem of restaurants make it worthwhile to invest in innovation. Besides, with
scale the quantity of data a restaurant can collect also increases and, with that, the efficiency and
precision of machine learning programs that sustain the infrastructure. Scale alone should imply
that McDonald's would be the biggest beneficiary of advancements in AI, but the percentage
of digital sales is not too dissimilar from Chipotle, which has ~1/12th the number of stores
(Exhibit 21). We believe McDonald's strategy of being a fast follower, who can rapidly deploy
a scale innovation that has become widely accepted, is perfectly in line with its positioning of
targeting the "mass" market with repeatable and consistent experiences, but leaves less room
for innovation through AI.
b. Restaurants that directly operate their stores?
Convincing a franchisee to make technology investments that may or may not lead to expected
results is hard. Doing it across multiple countries, each with different use cases, tech partners,
and underlying tech stack of the existing architecture, is borderline mission impossible.
Identifying Chipotle as a potential winner should be a no-brainer then. It controls the operating
environment of its stores and can roll out innovation across its 3,000+ stores virtually overnight.
It has a rigorous stage-gate process to advance the tech to the next "approval" phase before the
rollout. It has the capital to invest in technology once it is proven it meets certain ROI thresholds
(not purely economic returns: the innovation needs to be easy for its existing operations and
provide great experience to customers). But then why did we point out Yum as a potential winner
in our front page?
c. Restaurants that have a digital culture!
We believe the continuous, relentless focus on innovation differentiates Yum and Chipotle from
the pack. From the creation of innovation centers, to the scaling of the digital team reporting
directly to the CEO, to the prominence of digital in their mission statements (Exhibit 22 and
Exhibit 23), Chipotle and Yum have shown consistently that digital is not just an enabler, but a key
BERNSTEIN
OPPORTUNITIES FOR AI IN US RESTAURANTS 41
differentiating point that can unlock their next leg of growth.
We conclude with a special mention to Domino's digital leadership. For years, Domino's has
been regarded as a "tech company that happens to make pizza," and its digital investments were
unrivaled for a decade. The digital assets built over 15+ years and the data collected are still
incredibly valuable. Yet recent performance challenges, the softness in demand, and the tough
operating environment have redirected management's focus from innovation to execution, and
we look forward to a trend reversal (Exhibit 24).
EXHIBIT 21: Scale can be a significant enabler of
technology revolution
23%
10%
0%
10%
20%
0%
10%
20%
Carl's Jr. KFC Hardee's Chick-fil-A Burger King Arby's Wendy's McDonald's Taco Bell Dunkin'
Suggestive Selling (%) Industry Average (%)
55,362
40,275
30,722
19,880
7,095
3,187
YUM McDonald's RBI Domino's Wendy's Chipotle
System-wide Units
Source: Company reports, Bernstein analysis
EXHIBIT 22: Yum has put digital and technology at the
center of its growth strategy for the next two decades
Source: Yum Investor Day Presentation, Bernstein analysis
EXHIBIT 23: Chipotle has emphasized its focus on
technology and innovation to drive growth and
productivity
Source: Company reports, Bernstein analysis
EXHIBIT 24: The recent tough operating environment has
redirected Domino's focus from innovation to execution
Chart Title
20
25
30
35
40
45
50
# times digital
0
10
20
30
40
50
Sep-14
Jan-15
May-15
Sep-15
Jan-16
May-16
Sep-16
Jan-17
May-17
Sep-17
Jan-18
May-18
Sep-18
Jan-19
May-19
Sep-19
Jan-20
May-20
Sep-20
Jan-21
May-21
Sep-21
Jan-22
May-22
Sep-22
No. of times "Technology" or "Digital" was mentioned in an
earning call
Source: Company reports, Bernstein analysis
BERNSTEIN
42 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
VALUATION METHODOLOGY See Disclosure Appendix for valuation methodology.
EXHIBIT 25: Ratings and target prices
25-May-2023 Target
Ticker Rating Currency Closing Price Price
CMG O USD 2,057.90 2,300.00
DRI O USD 161.03 180.00
DPZ U USD 304.50 275.00
MCD M USD 285.52 300.00
QSR M USD 72.43 70.00
SBUX M USD 98.44 102.00
WEN O USD 22.22 27.00
YUM O USD 130.26 144.00
SPX 4,151.28
Source: Bloomberg, Bernstein analysis and estimates
RISKS See Disclosure Appendix for risks.
Danilo Gargiulo danilo.gargiulo@bernstein.com +1 212 969 1232
Bill He bill.he@bernstein.com +1 212 969 2536
BERNSTEIN
OPPORTUNITIES FOR AI IN US RESTAURANTS 43
BERNSTEIN
44 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
HAIER SMART HOME: OPPORTUNITIES
FOR AI IN HOME APPLIANCES
HIGHLIGHTS •Integration of AI is inevitable in the appliance sector. Consumers are no longer
satisfied with home appliances operating in silos. Companies must offer complete smart
home solutions that can address users' fundamental pain points to attract upgrades. The
integration of AI could potentially transform the business model from a one-time purchase to
a subscription model, under which users would pay for the smart home solution empowered
by AI. It will also increase customer stickiness and create a self-enhancing loop for a better
ecosystem. For Haier and AI developers that connect to the Haier platform, the vast amount
of user data acquired can also be used for continuous AI iteration and to provide more
customized services for users.
•ChatGPT can empower smart homes from being "smart" to being "intelligent." With
recent technological advancements in AI research, we are seeing increased potential for
further penetration of smart appliances through the adoption of AI. The bottleneck for a smart
home is the "question-answer" mode, and current user scenarios are not broad enough. Now,
with ChatGPT's better natural language processing capability, smart appliances can be fine-
tuned to respond to user commands in real time based on the conversation and context,
making the experience more similar to human interaction.
•Haier is well-positioned for AI adoption. Haier has developed a wide product matrix and
partnerships that make it well-positioned for advanced AI integration. Haier has established
various partnerships (e.g., Google (covered by Mark Shmulik), Baidu (covered by Boris Van)
and consumer electronic companies such as Huawei and Xiaomi (not covered) to complement
its capabilities in the AI and other home service fields, a one-stop-shop smart home solution
for customers.
INVESTMENT IMPLICATIONS We rate Haier Market-Perform.
INTRODUCTION As early as 1995, Bill Gates illustrated the idea of a smart home in his book The Road Ahead:
"First thing, as you come in, you’ll be presented with an electronic pin to clip to your clothes," wrote
Gates. "The electronic pin you wear will tell the house who and where you are, and the house
will use this information to try to meet and even anticipate your needs — all as unobtrusively as
possible." "As you walk down a hallway, you might not notice the lights ahead of you gradually
coming up to full brightness and the lights behind you fading." He also predicted that in the not-
too-distant future a home without a smart home system will be as out of fashion as a house
without internet access. With the internet thoroughly integrated into our lives, what is the current
integration progress of the once highly anticipated smart home? Two decades later, the smart
home Gates envisioned has gradually stepped into common households, thanks to equipment
manufacturers and software developers. Our appliances can now automatically adjust/turn on/
turn off based on pre-set instructions or upon our request.
BERNSTEIN
HAIER SMART HOME: OPPORTUNITIES FOR AI IN HOME APPLIANCES 45
In 2021, the smart home market reached US$107.3Bn, and is projected to more than
double to ~US$223Bn by 2027 (Exhibit 1). As a cornerstone of the smart home offering, smart
appliances take up the lion's share and are growing faster than the total smart home market. They
contribute ~40% of annual growth in the smart home industry and are expected to grow at a
~20% CAGR from 2021 to 2027. By 2027, the smart appliances segment will likely reach US
$88Bn, representing 40% of the whole smart home market. Hence, growth potential is huge for
smart appliance manufacturers. Recent technological breakthroughs in ChatGPT heralded the
dawn of AI and will revolutionize the way businesses are conducted. Businesses that can harness
new AI capabilities are best-positioned to stay relevant in the new era, and some may emerge
from the transformation as new winners. However, bottlenecks still remain for widespread
adoption to take place. With recent technological advancements in AI research, we are seeing
increased potential for further penetration in smart households through the adoption of AI.
In this chapter, we discuss Haier's Smart Home solutions as a case in point, look at its AI
applications in home appliances, and how the company is positioning itself in the long term to
ride the AI trend.
EXHIBIT 1: Smart home market size reached ~US$118Bn in 2022
41.87
55.71
68.12 78.7
107.3 117.6
139.3
160.3
181.4
202.5
222.9
0
50
100
150
200
250
2017 2018 2019 2020 2021 2022 2023 2024 2025 2026 2027
US$bn
Global Smart Home market size
Comfort & Lighting Control & Connectivity Energy Management Home Entertainment Security Smart Appliances
Note: 2022-27 data is Statista projection
Source: Statista, Bernstein analysis
HOW WILL AI IMPACT HAIER
AND THE INDUSTRY?
Financially, the integration of AI should be able to drive up ASP with the offering of smart home IoT
solutions. Currently, Haier is leading in the premium appliances sector. To further increase ASP
and strengthen its position in the premium segment, Haier could use AI as an enabler to deliver
more intelligent smart home solutions. Margins and returns on investments would follow.
Strategically, the integration of AI is an unstoppable trend in the appliances sector. The growth
of the sector has shifted from first-time buyers to upgrades in China. Today, users are no longer
satisfied with the separation of home appliances. Companies must offer complete smart home
solutions that can address users' fundamental pain points to attract upgrades. Failing to do
so would lead to missed opportunities and a weakened market position. More importantly, the
integration of AI could potentially transform the business model from a one-time purchase to a
subscription model, under which users would pay for the smart home solution empowered by AI.
It will also increase customer stickiness and create a self-enhancing loop for a better ecosystem.
BERNSTEIN
46 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
For Haier and AI developers that connect to the Haier platform, the vast amount of user data
acquired can also be used for continuous AI iteration and to provide more customized services
for users.
The integration of AI will set off a huge change in smart home technology, opening up new
possibilities for user experience improvement and reshaping the entire industry in the process.
Industry competition is expected to intensify as more players enter the field. On balance, this is a
once-in-a-lifetime opportunity for smart home manufacturers to refresh their brand and expand
into new markets.
HOME APPLIANCES AI As an appliance manufacturer, Haier's core strength is in creating smart home appliances (i.e.,
the third element) (Exhibit 2) as opposed to AI development and content curation. When it comes
to AI itself, internet giants (e.g., Google, Baidu, and Alibaba) and consumer electronics companies
(e.g., Xiaomi and Huawei (not listed)) have strengths in developing core AI speech recognition,
speech synthesis, natural language processing (NLP), and content recommendation algorithms,
all of which necessitate extensive technological know-how. Haier should avoid competing in this
market. As for content, Haier can gain access through partnerships with content providers or
with users' own subscriptions or open-source information. However, no matter how smart AI-
generated home solutions become, they can only be implemented in conjunction with suitable
appliances. Therefore, we see Haier as an essential contributor to the smart home solutions
ecosystem.
Three elements to realizing a smart home solution:
1. AI development. To understand user commands, generate solutions, and control linked
smart appliances. How "smart" the AI is determines the development stage of the AI. In
general, smarter AI requires less input or instructions provided by the user and can generate
more solutions independently.
2. Content curation. Relevant under different user scenarios for searching and entertainment.
A smart home should be able to provide users with suggestions and satisfy users'
entertainment needs with TV series, movies, songs, etc.
3. Smart home appliance products. The appliance product should understand instructions
from AI or users. Various smart home appliances should be interconnected to a network and
allow for interoperability between devices so that the entire home smart solution can be
enabled.
BERNSTEIN
HAIER SMART HOME: OPPORTUNITIES FOR AI IN HOME APPLIANCES 47
EXHIBIT 2: Haier's core strength lies in creating smart home appliances, which plays an essential role for smart
home solutions
Smart home solution key
elements
Haier's
presence Haier's accessibility
Limited Through partnership
No Through partnership
/user's own subscription
Smart home solution
Strong Self manufacturing
AI
Content
Smart
home
products
Source: CNN (Picture credit: Leanza Abucayan/CNN), Bernstein analysis
THE FUTURE OF HAIER'S
SMART HOME SOLUTION
ChatGPT can empower the smart home industry from being "smart" to being
"intelligent"
The fundamental bottleneck for a smart home is the immaturity of AI technology.
Although smart home appliances have made great progress over the last decade, people often
find the current solutions not "smart" enough. As Haier's CTO Xue Wei has pointed out, there are
still many pain points in the smart home space. From a user's perspective, there are difficulties
in cross-brand interconnection, the low value-add of human-machine interaction (many home
appliances are simply controlled via APP), and the single scenario experience (e.g., user scenarios
are usually limited to leaving home and coming home only). In addressing difficulties in cross-
brand interconnection, Haier has built the AI of Things (AIoT) platform and made most of its smart
appliances connectable to mainstream AI assistants. However, the fundamental bottleneck for
a smart home remains the immaturity of AI technology, which leads to unsatisfactory interaction
experiences and limited user scenarios. Currently, most smart home products interact with
users through hardware buttons, voice, gestures, auto-sensing, mobile apps, etc., and the
basic interaction form is command-based question-and-answer. However, the questions and
instructions are usually limited in the scenario and, from time to time, inaccurate and lagged.
Therefore, the next key bottleneck for the industry is to upgrade the intelligence level
of the operating system and the AI speaker (which plays the role of central control).
Haier ranks the intelligence level of a smart home into five stages: single-product intelligence,
synergy intelligence, decision-making intelligence, highly proactive intelligence, and ubiquitous
intelligence.1 Most businesses are still in the first two phases (i.e., single-product intelligence and
synergy intelligence), while Haier is leading in the third phase of "decision-making intelligence,"
1 Ubiquitous intelligence is defined as the development of software that can be used in multiple platforms instead of just one
device (source: Ubiquitous Intelligence and Computing by Braylen Stevenson (editor)).
BERNSTEIN
48 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
where devices are connected and upgraded to active perception, learning, decision-making, and
proactively providing services for humans. For example, adjusting the temperature automatically
based on the weather is an application of decision-making intelligence.
Now, with the recent disruptive development in OpenAI technology represented by
ChatGPT, smart home appliances can be upgraded to the next level. As ChatGPT has
better natural language processing and content recommendation algorithms, not only can it
execute the system's inherent commands and operations, it can also be fine-tuned to respond
to user commands in real-time based on the current conversation and context. For example,
Mate Marschalko, a web developer, shared his smart home experience of using ChatGPT in
combination with Siri shortcuts to make the process "smarter." He told Siri, which was connected
with ChatGPT, that he was "recording the video in the dark in the office" and asked, "Can you do
something about that?". He got the response "turning on the light for you" (the original video).
Despite the complexity of the commands given by Marschalko, the voice assistant combined
with ChatGPT was able to understand and quickly respond. In this way, the traditional "question-
answer" interaction in the smart home will be gone.
ChatGPT makes the interaction more human. Meanwhile, ChatGPT is able to communicate
more efficiently and anthropomorphically based on the autonomous learning capabilities of
previous interactions, making the interaction experience more similar to real human interaction.
In particular, its powerful contextualization and learning capabilities allow ChatGPT to provide
personalized advice and services based on a user's personal preferences and usage habits.
Ultimately, its application will greatly enhance the user experience by improving the accuracy of
feedback, understanding user intent better, and offering proactive service solutions based on
user data and behavior. In time, "decision-making intelligence" will evolve to "highly proactive
intelligence" or even "ubiquitous intelligence."
We expect other AI speakers will follow soon. Baidu has announced the integration of its AI
speaker Xiaodu with Wenxin Yiyan, Baidu's ChatGPT-like project, to train an upgraded version
"Xiaodu Linji" for smart home solutions.
HOW IS AI BEING USED
IN HAIER'S SMART HOME
SOLUTION?
Haier's way of integrating AI into its smart home solution
Haier's prototype smart home solution can be dated back to 2006. In July 2006, Haier released
a new generation of the home network platform, U-home. In the following years, Haier made
considerable progress. By 2012, Haier officially launched its fifth strategic transformation
— networked strategic transformation, i.e., transforming from traditional home appliance
manufacturing into a platform-based enterprise incubating creators. As Haier CEO Zhang Ruimin
pointed out, the sector's growth driver will come from platform building, and Haier has to shift its
focus from scaling to becoming a platform enterprise with its own ecosystem.
In 2014, major platforms including Google, Amazon, Xiaomi, Midea, etc., started building their
smart home platforms and launched ecosystems around the platforms. Seeing the disruptive
nature of the Internet model, Haier built its smart home platform over the next few years.
Haier's U+ smart life platform was officially released in 2014, and the U+ industry alliance was
established. U+ is a software platform to accommodate all the connected appliances, comprising
the U+ Smart Home Interconnection Platform, the U+ Cloud Service Platform, and the U+ Big
Data Analysis Platform. In 2016, the operating system for Haier U+ was officially launched. The
following year, Haier released U+ Smart Home IoT cloud solution, U+ Cloud Chip, along with
BERNSTEIN
HAIER SMART HOME: OPPORTUNITIES FOR AI IN HOME APPLIANCES 49
the integration of micro-kernel cross-terminal UhomeOS 2.0 and became the first brand in the
industry to implement complete sets of sales.
By 2021, Haier Smart Home had developed the Smart Home Brain, which integrates AI, IoT, big
data, and other advanced technologies to connect "terminals" with the "cloud." The "terminals"
refer to hardware products such as refrigerators, washing machines, and air conditioners in the
home, while the "cloud" is composed of software such as Smart Home Brain and IoT technology.
For example, when a user uses the air conditioner, the air conditioner will record the time
and temperature when it is switched on, and transmit this data to the software, which will
continuously learn the user's habits, so that when the user turns on the AC again, it will actively
adjust the temperature to the right level, and if the weather shows that the temperature is low
today, it will also ask the user if the temperature needs to be turned up.
Building an AIoT platform through partnerships
Although Haier has developed its own smart speaker with AI assistant "Xiaoyou," it does not
have much market share. The leading players in smart speakers in China are Xiaomi, Baidu, and
Tmall (Exhibit 3). Globally, the market is dominated by Amazon Alexa, Google Home, and Apple
HomePod. Haier's solution to this problem is to partner with a group of companies that have
differentiated strengths to create the AIoT smart home platform. Haier has made its smart home
appliances connectable to the major smart AI speakers in the market (e.g., Baidu's Xiao Du,
Alibaba's Tmall Genie, Xiaomi's Xiao Ai, Apple's Homekit, Amazon Alexa, etc.). We believe such
a strategy will keep Haier's resources on its core strength in appliances while empowering the
products with smarter software by leveraging the expertise of platform companies.
EXHIBIT 3: Leading players in smart speakers in China are Xiaomi, Baidu, and Tmall; Haier smart speaker Xiaoyou's
share is limited
Baidu, 35%
Xiaomi, 31%
Tmall, 27%
Huawei, 4%
Others, 3%
Major AI speaker manufacturer market share in China (2022)
Source: Runto Technology, Bernstein analysis
Haier has entered into multiple partnerships with ecosystem players (Exhibit 4) to strengthen its
Smart Home development:
•Haier launched Haiji.com, an AIoT open platform focused on smart homes, to enable
developers to build smart home applications faster. Haiji.com supports individual or enterprise
developers who create, develop and test, and launch products or applications on the website,
BERNSTEIN
50 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
making it the fastest and most convenient one-stop development platform for developers
to enter the smart home ecosystem. The "Escort Program" was launched with it, providing
developers with development tools, Haiji.com courses, documentation and other support
tools, as well as free access to scenario labs, a professional consulting team, and a 10-year
warranty on all categories of modules.
•For ecosystem partners in the home appliances sector, including the furniture and
consumer electronics industry, Haier Smart Home has also launched the Tenglong Plan,
which will help enterprises upgrade their AIoT intelligent products and give them the greatest
support in terms of hardware, software, and ecological service resources, such as sharing
Haier's 183 million smart home users data and sharing all channels of the Three Winged Bird
Experience Store (mainly selling smart home products).
•For hardware developers, Haier can provide smart chips, module software development
kits (SDKs2), and a one-stop hardware access platform. For application developers, Haier
can provide more than 550 core application programming interfaces (APIs3), a library of
more than 200 device standard models, a library of more than 400 device user interface (UI)
components, more than 80 scenario templates, and a corpus of home appliance knowledge.
•For AI development, Haier has cooperated with Sogou (not covered) and Baidu to strengthen
its capability in this area. In 2017, Haier U+ reached strategic cooperation with Sogou
Zhiyin (a search engine that focuses on voice interaction technology), focusing on two main
areas: (1) Sogou Zhiyin provides U+ with technical support for speech recognition and
speech synthesis, and also empowers smart appliances with full semantic understanding
capabilities. Sogou will also provide back-end content for Haier's U+ platform, including, but
not limited to, encyclopedia, yellow pages, weather, news, movie tickets, recipes, and other
content resources, and (2) On the hardware level, Sogou Zhiyin will provide a complete set
of far-field speech recognition solutions, including hardware microphone arrays designed to
achieve the best far-field speech recognition. In order to accelerate the deep integration
of AI technology with smart homes and to bring the value of AI technology into play,
Haier, together with Sogou, the Chinese Academy of Sciences, Xiangsheng Internet (private),
Alibaba, Linglong Technology (private), and GateQ (private), launched the Haier U+ Smart
Home "+AI Family" program in 2017. In the "+AI Family" plan, Haier creates a distributed,
multi-entry interaction platform through multi-modal interaction, and opens it up to third-
party services, including hard and soft voice solutions, smart device interaction services,
hardware data entry, and skill development interfaces. In 2018, Haier reached strategic
cooperation with Baidu in the field of smart homes, IoT, big data, and AI. Among those fields,
the two sides plan to jointly explore smart living scenarios (e.g., smart living room, smart
kitchen, smart bedroom, and other scenarios of technical solutions) and make other hardware
devices equipped with Baidu's DuerOS (Xiaodu's operating system) and U+ platform.
2 A collection of software tools and programs made available by hardware and software providers for developers to utilize in
creating applications for certain platforms.
3 A software component that enables two disconnected software programs to communicate.
BERNSTEIN
HAIER SMART HOME: OPPORTUNITIES FOR AI IN HOME APPLIANCES 51
EXHIBIT 4: Haier works with a group of companies (including Google, Amazon, etc.) to accomplish its smart home
products
Source: Haier website
Haier's current smart home solutions
Overall, Haier now has more than 56 categories and 4,000 models of smart appliances
(intelligent home appliances with IoT capability), and can provide users with products ranging
from individual products to complete sets, as well as intelligent and connected products for
different spaces and different needs.
User scenarios demonstrating how Haier brings AI assistant into everyday life
Smart Living Room: With just one word, Haier Smart Living Room can automatically turn on the
lights, turn on the music, turn on the air conditioning with customized exclusive comfort breeze,
adjust the temperature, play soothing background music, while security is disarmed, and these
operations can be customized freely according to the user's needs.
Smart Bedroom: Haier Smart Bedroom not only allows users to set the light switch,
temperature, as well as light color, but also customizes a variety of light modes, which helps
cultivate sleepiness even more. After the reading is done and fatigue sets in, the curtains close
automatically, the security system is activated, the lights are turned out, and the air conditioning
enters sleep mode. In the morning, you can also ask the AI speaker to open the curtain, adjust the
AC temperature, and set the water heater to a certain temperature (Exhibit 6).
BERNSTEIN
52 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
EXHIBIT 5: Haier fridge's food management
Source: Haier website
EXHIBIT 6: When you wake up in the morning, you can ask the AI speaker to open the curtain, adjust AC
temperature, set the water heater, and set the temperature
Chinese translation: You tell the speaker: 'I want to take a shower, please prepare the hot water for me.'
Source: Haier website, Bernstein analysis
BERNSTEIN
HAIER SMART HOME: OPPORTUNITIES FOR AI IN HOME APPLIANCES 53
Smart Kitchen: Haier Smart Fridge has a powerful function in food management (Exhibit 5).
After storing items in the fridge, the platform uses radio frequency identification (RFID) to provide
intelligent management, allowing users to trace the source of an ingredient, connect the fridge
to other smart appliances, remind users about the freshness of food, and set a reminder to
automatically replenish supplies when low. If there is a shortage of ingredients, they can be
purchased directly with one click through the big screen, and fresh ingredients will be delivered.
Another scenario is cooking. When you are thinking about your dinner, open the fridge and pick
up the ingredients, and the fridge will automatically suggest the recipe for the oven, then display
the recipe on the oven's screen (Exhibit 7). Then Haier's Chef@Home oven, which has a presence
sensor, will switch on immediately, providing the user access to the full touch control panel on
a big screen. Additionally, with Haier's PreciTaste technology (utilizing Bluetooth), the oven can
recognize the sorts of food placed inside the oven, check the temperature inside with one-degree
accuracy, and even let the oven adjust cooking cycles by itself, automatically, and notify the
user when the meal is done. Meanwhile, its fridge also has facial and voice recognition function,
so it can remember the face and habits of different family members and then provide tailored
recommendations.
EXHIBIT 7: Example of Smart Kitchen user scenario
User opens smart fridge Smart fridge sugguests a recipe Fridge sends the receipe to smart oven The oven will notify the user
based the ingredients user picks and set temperature automatically when the food is ready
Source: Flaticon, Bernstein analysis
Smart Bathroom (Exhibit 8): Users can ask the Xiaoyou AI speaker to turn on the water heater,
set it to healthy anti-bacterial mode, set the water temperature, and turn on the dehumidifier, etc.
Based on user data, Haier's water heater can also adjust bath water temperature based on the
different needs and physical conditions of different family members.
Smart Balcony (Exhibit 9): Haier's intelligent washing machine can automatically identify the
water quality and temperature, fabric category, and laundry detergent composition, and control
the washing process by voice through U+ APP. After the washing program is completed, the
hanger will automatically lower down for users to hang up the clothes and elevate with a
single touch. In this way, the washing and drying experience is upgraded, and iterated with
the integration of smart home appliances. Haier also cooperates with thousands of ecosystem
partners to create customized home balconies for users, including pet balconies, leisure-
oriented balconies, and fitness-oriented balconies.
BERNSTEIN
54 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
EXHIBIT 8: Haier Smart Bathroom
Source: Haier website
EXHIBIT 9: Poster of Haier's Three Winged Bird smart
balcony
Source: Haier website
Will Haier face threats from internet giants, given their strength in AI?
A smart home solution requires a system smart enough to be "intelligent" and a product portfolio
rich enough to realize the instructions made by AI under different scenarios. Haier has a rich
product portfolio and efforts in building a smart home solution platform. Therefore, we think such
threats should be limited.
The relationship internet giants have with appliance manufacturers should be more
cooperative than competitive. While internet giants such as Microsoft, Google, and Baidu have
the advanced technology to develop a smarter AI assistant (the interface usually in the form of
a smart speaker), they do not possess the rich smart appliance portfolio to meet users' various
needs. Moreover, appliance manufacturing is a low value-add and would lower their investment
returns if they decide to enter the appliance industry themselves.
While consumer electronics firms such as Huawei and Xiaomi have been developing ecosystems
with a variety of small appliances, they lack a foothold in the major appliances market, and large
appliances are more closely related to user experience improvement under AI. That said, Haier
has a limited market presence in smart TV, which is an area where Xiaomi has a strong position.
Haier's product offerings are focused on daily care (e.g., washing, cooking, cleaning, etc.) instead
of entertaining.
Haier is also better-positioned than other appliance companies to integrate AI with its
smart appliance sets. As set products are usually designed to achieve synergies, AI could
more actively and easily instruct appliances to jointly provide services for users with a complete
smart home appliance set. Haier has cultivated its premium smart home solution brand Three
Winged Bird, offering smart home solutions in the aforementioned five living spaces (living room,
kitchen, bathroom, bedroom, balcony). There are more than 20,000 interior designers and 2,000
eco-partners on the Three Wings Bird platform, working together to provide users with full-
process visibility of home appliances, home decoration, and home and home life services. Since
the launch of Three Winged Bird in September 2020, Three Winged Bird Smart Life Experience
Centers have expanded to 1,605 nationwide, covering 238 prefecture-level cities as of 3Q22
and, according to Haier Smart Home's plan, will expand to 3,316 stores by 2023 in 304 cities.
BERNSTEIN
HAIER SMART HOME: OPPORTUNITIES FOR AI IN HOME APPLIANCES 55
Haier's major domestic competitor Midea also offers appliance sets, but it isn't as established as
Haier in the premium appliance set segment. Gree, another major appliance company, is focused
on air conditioners and has fewer SKUs and a limited market share in fridges and washing
machines, thus lacking the fundamental product matrix to provide a full solution for users.
VALUATION METHODOLOGY See Disclosure Appendix for valuation methodology.
EXHIBIT 10: Ratings and target prices
Haier Haier
600690.CN 6690.HK
Rating M M
Prices as of May 25, 2023 22 23
Currency CNY HKD
Target Price 25 27
52-Week Range 20.71 - 27.88 19.52 - 31.55
Market Capitalization (US$ billion) 28,549 28,549
TTM Performance -9.3% -10.4%
TTM Relative Performance -15.4% -8.2%
Bernstein EPS Forecast
2022A 1.58 1.75
2023E 1.76 1.95
2024E 1.99 2.21
2025E 2.22 2.47
EPS Annual Change
2022A-2023E 11.4% 11.4%
2023E-2024E 13.1% 13.1%
2024E-2025E 11.7% 11.7%
Consensus EPS
2023E 1.8 2.0
2024E 2.1 2.3
2025E 2.3 2.6
P/E on Bernstein EPS Forecast
2023E 12.1x 11.3x
2024E 10.6x 9.8x
2025E 9.4x 8.7x
Shares Outstanding (mil.) 6,309 2,867
Yield 2.10% 2.37%
Dividend per Share 0.57 0.57
Benchmarks, with closing prices as of May 25, 2023, are:
- Stocks trading in Greater China are benchmarked against the MSCI Asia Pacific ex Japan Index, which had a closing price of 507.4905
Source: Bloomberg L.P., company reports, Bernstein analysis and estimates
RISKS See Disclosure Appendix for risks.
Melinda Hu melinda.hu@bernstein.com +852 2918 5727
Charles Gou Charles.Gou@bernstein.com +852-2918-5789
Shirley Yang shirley.yang@bernstein.com +852 2918 5303
BERNSTEIN
56 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
US APPAREL & SPECIALTY RETAIL: TO
BUILD OR TO BUY? OPPORTUNITIES FOR
AI IN RETAIL
HIGHLIGHTS •Predictive recommendations: When done well, predictive recommendations can resolve
the infinite shelf dilemma and paradox of choice, driving higher conversion and oftentimes
greater satisfaction for retailers — however, most retailers only offer a far more primitive
system of recommendations that aren't fully leveraging their vast amounts of data.
•Product design: Any item of clothing can be broken down into a set of attributes that can
each be enhanced based on customer feedback, and larger brands are testing this approach
of incorporating direct feedback into new product designs (e.g., Nike and Adidas).
•Warehouse matching: Using RFID, both small-scale (e.g., Stitch Fix) and large-scale (e.g.,
Inditex) retailers have built real-time matching algorithms that make e-commerce fulfillment
and shipping decisions in real time to minimize last-mile distribution costs.
INVESTMENT IMPLICATIONS We rate SFIX Market-Perform.
WHERE'S AI IN RETAIL? Within Apparel & Footwear retail is a gold mine of data. With almost US$2Tn of annual
sales and transactions (Exhibit 1) involving millions of dollars of products across thousands of
SKUs daily, there is an immense amount of consumer data generated every day on consumer
sales patterns and behaviors. There is also a huge amount of supplier data, with products
going through several layers of producers and intermediaries across multiple countries in the
manufacturing process. And in the middle, there is a vast amount of data for the distribution,
delivery, pricing, and merchandising of the products. As the consumer experience becomes more
digital (e-commerce is now 30% of sales — Exhibit 2), there is also a vast tranche of app, loyalty,
and web traffic data that can describe and explain consumer behavior.
As such, there is an ever-increasing potential for what data can be collected, what can be done
with it, and what it can do to help retail businesses. AI and machine learning can potentially
transform how retailers interact with their customers, suppliers, and products if the data is
captured and used properly.
Retailers face three questions when dealing with all their data: What data to collect?
How to convert the data into insights and make decisions? And how to scale up and
monetize the data solutions? Companies are far along on the first question, identifying what
data they want from their customers, suppliers, and products. They are making some progress
on the second question by leveraging AI and machine learning tools to store and analyze their
data in a manageable way. But they are at very early stages on the last question — some of
the companies that have the most advanced AI capabilities are still not translating those into
revenues and profits, while the ones that have the most to gain aren't yet fully leveraging AI.
BERNSTEIN
US APPAREL & SPECIALTY RETAIL: TO BUILD OR TO BUY? OPPORTUNITIES FOR AI IN RETAIL 57
In going through these examples, we focus on a retailer that has already ingrained AI-
driven solutions in each area: Stitch Fix. Uniquely positioned in the space between fashion
and data science through a data-rich, subscription-based buying model and e-commerce
pure play channel structure, Stitch Fix has focused on data collection and analysis as a core
competency since its founding in 2011. We look at Stitch Fix's data solutions in predictive
recommendations, warehouse matching, and product design as examples for other retailers.
We also look at the other end of the spectrum — retailers that are big enough to justify an
investment into AI, but haven't made enough progress yet to drive meaningful results in
one or more of these areas. This is where the opportunity lies for AI.
We wonder what the implications of this uneven progress between retailers will be. If
data-rich smaller retailers such as Stitch Fix can build advanced AI capabilities, but are unable to
scale them up due to their own size and limited scale, while larger retailers and brands have the
scale but not the technology, should they build or buy their way there? We expect to see a sort
of VC model evolve, where startups build and experiment, and giants buy and scale up — either
buying the smaller companies outright or buying their data or technology as a service. The losers
will be the retailers/brands that do neither or the ones that try to do too much at once and fail to
integrate the technology into the core operations of the business.
Given the uneven progress across the sector, not correlated with size or data availability, we
expect to see a sort of VC model evolve, where emerging data-rich retailers make AI foundational
to their business model (e.g., Stitch Fix) and build these capabilities, and mature retailers/brands
that have the scale but not the technology opt to buy, either by buying the start-ups altogether
or buying the data (or the technology as a service) — both the builders and the buyers can
differentiate and win.
EXHIBIT 1: Global A&F market is nearly US$2Tn in annual
sales…
1.7 1.7 1.6 1.6 1.7 1.8 1.8
1.5
1.7 1.7
2013 2014 2015 2016 2017 2018 2019 2020 2021 2022
A&F Market Size ($ trillion)
Source: Euromonitor, Bernstein analysis
EXHIBIT 2: …with e-commerce now 30% of sales and
growing
9% 10% 12% 14% 16%
18% 20%
30% 31% 30%
2013 2014 2015 2016 2017 2018 2019 2020 2021 2022
A&F Ecommerce Penetration
Source: Euromonitor, Bernstein analysis
PREDICTIVE
RECOMMENDATIONS
The infinite shelf and the paradox of choice. Want a mop dog crossbody? You can get one
from Kate Spade (owned by Tapestry). Want an official FIFA World Cup ball? It's available on
adidas.com and wholesalers such as Dick's Sporting Goods (not covered). Want the same shoe
that marathon world record holder Eliud Kipchoge wore? You can get the Air Zoom Alphafly Next
% 2 from Nike. In the infinite shelf of e-commerce, having too many options at our fingertips
leads to the paradox of choice. Shoppers are inundated with choices — on size, color, fit, price,
brand, delivery time, delivery cost, return policy, and more — which can overwhelm shoppers and
BERNSTEIN
58 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
prevent them from actually purchasing anything. Take women's jeans for an example — there
are billions of search results on Google, tens of thousands of styles on aggregator sites, and
thousands of styles on even retailer platforms (Exhibit 3 and Exhibit 4).
With a narrower choice set comes higher conversion and oftentimes greater satisfaction.
Research has also shown repeatedly that having too many options can result in lower satisfaction
even after a decision is finally made. Version 1.0 of this in the world of e-commerce is putting a
series of filters on the selection — this somewhat helps if a customer has a pretty good idea of
what they are looking for, but less so when they are just browsing through tens (or hundreds) of
thousands of options looking for inspiration. Anyone who's been frustrated by the inability of the
standard filter options to help them find what they want will agree that there is a clear opportunity
for something more personalized.
Stitch Fix is one of the few retailers today that is doing AI-driven predictive
recommendations well. With the entire original business model based on curation and
recommendations, Stitch Fix ingrained personalization into the core business model right from
the start when the company was founded in 2011 (Exhibit 5). Stitch Fix curates a personalized
box, i.e., a "Fix," for its clients, which contains five recommended items that Stitch Fix anticipates
a client will like, based on data points they have collected (Exhibit 7). We describe below how they
collect this data and how they use it.
•The first source of personalization data is a questionnaire called a Style Profile. The
survey starts by asking the typical questions — your stats (height and weight), typical sizing for
tops and bottoms, and whether your usual sizing tends to run too small, just right, or too big.
The survey then goes further, inquiring deeper into your fit considerations (e.g., if you prefer
your tops slimmer or looser) and asks about any fit challenges you may have experienced (e.g.,
pant thighs too tight or sleeve lengths too long) with accompanying photos (Exhibit 6). Next,
the survey goes into how often you wear each category of clothing (e.g., leisure wear versus
business casual), and style preferences. Finally, the survey inquires your price preference for
a variety of categories, brands you typically wear, and has a 500 open-ended text box for any
final thoughts.
•The second source is keep versus return behavior. Understandably, having customers fill
out such a detailed questionnaire as above isn't feasible for most retailers — Stitch Fix has
the advantage there of being a subscription-based service. But the second source of data is
available to all — data on customer behavior in the purchase journey. Not only what customers
buy versus return, which is the most telling indicator, but retailers can also collect aggregated
data on what products customers tend to click on, what they add to their carts, and how
these behaviors vary by demographics, cohort, purchase history, basket size, etc. For Stitch
Fix's newer business model, Freestyle, which is basically just a personalized version of regular
multi-brand e-commerce, this second source is more valuable because the assortment is
newer and non-subscription customers are less likely to want to fill out a long questionnaire.
•A variety of algorithms are used to produce rank-ordered lists of the inventory — a
collaborative filtering problem. The job of these algorithms is to filter out fits with a client's
Style Profile and then rank them based on which would be most appealing. To start, a filtering
step removes styles that the client has received before (if they are a returning customer) or
have attributes the client has asked to avoid (e.g., customers can explicitly note items they
do not want to receive in the curated "fixes"). Next, the remaining items are fed into several
BERNSTEIN
US APPAREL & SPECIALTY RETAIL: TO BUILD OR TO BUY? OPPORTUNITIES FOR AI IN RETAIL 59
algorithms, which output match scores for each item, based on the match between the client's
preferences and the item's features, which take into account feedback from other clients on
each item. With the data they have at hand, Stitch Fix must fill in the sparse matrix to predict
the result of sending a style to a client who hasn't received it yet.
•The task is a longitudinal one. Stitch Fix tracks client preferences over time individually
and as a whole. What makes the task all the more challenging is that, in additional to explicit
features at their disposal (i.e., explicitly mentioned preferences), there are also latent (i.e.,
unstated) features that can be inferred. As each product and each client become more mature,
with a richer data history on each, this process becomes easier — the algorithms learn over
time and get better (Exhibit 8).
•Neural networks and natural language processing process images and text. Neural
networks are trained to look at photos of clothing customers like (e.g., from a client's Pinterest
board that they share with SFIX), and look for visually similar images. Similarly, a client's
request note and textual feedback is processed by AI.
•Humans are still generally involved in the ultimate decision, but the heavy lifting is
done by AI. Further, even the matchmaking between stylist and client is done by algorithm as
well. Match score is calculated based on each available stylist and each client who's requested
a shipment during the current period, considering affinities between client's and stylist's
stated (and unstated) style preferences as well as variables like fix history, feedback, style
profile alignment, and stylist queue size (Exhibit 9). The stylist finalizes selections from the
inventory list and writes a personal note with tips on accessorizing/making outfits, though
they are not always a part of the process — and, increasingly, as the model scales and the AI
becomes smart, the process has evolved to be more AI-driven.
Personalization is a big buzzword, but most retailers still aren't getting it right. Who else
is doing personalization well? The answer is — not many! Further, while Stitch Fix does predictive
recommendations well, the company operates at a small scale with only 3.6 million active clients.
Though personalization has been a big buzz word in 2023 across retailers of all sizes, large
players who are doing it have only dipped their toe into the waters (e.g., Gap "These Look Like
You" — Exhibit 10).
There currently isn't any A&F retailer that does predictive recommendations well and at
scale. Retailers generally offer product recommendations based on the browsing history and
item currently viewed, which is often quite rudimentary, and the recommendations aren't tailored
enough. Though brands such as Nike and Coach (owned by Tapestry) offer more of a lookbook
feel to their recommendations (Exhibit 11 and Exhibit 12), ultimately it is just a more elegant
execution of the same rudimentary recommendation system, based more on the product being
viewed than on each shopper's personal preferences and unique taste. Moreover, shoppers have
come to expect recommendations for similar products on product listing pages, but they are not
always helpful or even relevant — e.g., Macy's displays kid's track suits under "Similar Items" for
a Adidas women's track set (Exhibit 13). Others have offered some human interaction, but this
is not scalable and has a much higher variance of whether it will be satisfactory or not — e.g.,
Macy's stylist services.
BERNSTEIN
60 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
EXHIBIT 3: Google displays billions of results for
women’s jeans, and tens of thousands can be found on
aggregator sites
Women's Jeans # of Items:
ShopStyle 90,000
Lyst 80,000
Stylight 20,000
Source: Google, Company websites, Bernstein analysis
EXHIBIT 4: Retailers offer thousands of styles for
women's jeans
4,950
2,700
1,500
400 300 200
Nordstrom Asos Macy's H&M Gap Levi's
Women's Jeans Styles Available (US)
Note: Nordstrom, Asos, Macy's, H&M, Gap, Levi's not covered
Source: Company websites, Bernstein analysis and estimates
EXHIBIT 5: Stitch Fix is uniquely positioned in the space between fashion and data science
Source: Company disclosures
BERNSTEIN
US APPAREL & SPECIALTY RETAIL: TO BUILD OR TO BUY? OPPORTUNITIES FOR AI IN RETAIL 61
EXHIBIT 6: The style quiz is extensive, with questions
about size and fit, and also preferred styles
Source: Company website
EXHIBIT 7: The output is a personalized box, curated by
the interaction between algorithms and human stylists
Source: Company website
EXHIBIT 8: Stitch Fix finds the best match between their clients and available merchandise, leveraging their wealth
of data
Source: Company disclosures, Bernstein analysis
BERNSTEIN
62 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
EXHIBIT 9: Algorithms see customers and clothing as
vectors based on their various features, trying to find
the best match between the two
Source: Company website
EXHIBIT 10: Gap offers 12 product recommendations on
their homepage based on prior browsing history and
other customer data
Source: Company website, Bernstein analysis
EXHIBIT 11: Nike provides outfit recommendations on
select product listing pages
Source: Company website, Bernstein analysis
EXHIBIT 12: Coach also offers shoppable looks on product
listing pages
Source: Company website, Bernstein analysis
EXHIBIT 13: "Similar items" suggested by Macy's for an Adidas women's track set show similar-looking products,
but in a completely different category (kid's department)
Source: Company website, Bernstein analysis
BERNSTEIN
US APPAREL & SPECIALTY RETAIL: TO BUILD OR TO BUY? OPPORTUNITIES FOR AI IN RETAIL 63
PRODUCT DESIGN An item of clothing can be broken down into a set of attributes. Considering all
permutations for each attribute (e.g., long versus short sleeves, stripes versus plain, cotton
versus polyester, loose versus tight fit, 30" versus 32" inseam) makes for a very large number of
possible combinations — therefore, in theory, any product could have been designed hundreds
of thousands of different ways.
Brands and retailers get constant real-time feedback from customers on each product.
The feedback can be explicit, e.g., customers filling out reviews mentioning exactly what they
liked and disliked. Or it can be implicit, e.g., reading into differences in buy rates or return rates
across very similar SKUs with one or two major differences. For example, brands often have
excess stock in unpopular colors which are sold off on clearance at the end of a season — this is
a great example of real-time feedback on color preferences that the brand can use next season.
AI can help filter down the combinations to a list of potential new designs, based on
the feedback received, as well as filtering through the feedback itself. If personalized
recommendation is one side of the coin, product design is the other. The two are very much
related and dependent on an understanding of what the target audience wants and delivering.
Using data on buy rates, return rates, even click-through behavior such as dwell times and cart
adds, as well as the explicit feedback in customer reviews, AI can help narrow down viable options
on a given product to exclude those unlikely to resonate, and can pinpoint specific product
attributes that should be reviewed or changed.
Stitch Fix does this well for its own products and brand partner products. For the 18
brands offered exclusively on Stitch Fix, a mix of private label and brand partner made products,
the company uses AI to filter through explicit customer feedback, Style Shuffle, as well as keep
rates versus return rates, to pull out key insights for how the products can be edited. Billions of
data points fuel innovation in Stitch Fix's business and beyond to make more personal, appealing
products for their customers.
•Stitch Fix uses a picky genetic algorithm for new product design ideas. The algorithm is
inspired by genetics and natural selection, with the designs with the greatest "fitness"coming
out on top. There is both a vast amount of potential new design combinations and also a vast
amount of feedback from previous Stitch Fix customers. Feedback from users about what
they liked and disliked about their products is fed into a picky genetic algorithm that highlights
sets of attributes that are likely to suit the target client(s), which are then vetted and refined
by human designers (Exhibit 14).
•Feedback loop and real-time updates cause continuous improvement. Management
has commented on how popular Stitch Fix's own brands are with customers. A likely driver
of this popularity is the machine learning that goes into their design. As feedback-inspired
products go out to more customers, who then produce more feedback, it essentially creates
a continuously improving feedback loop that hones the designs over time.
Who could ideally benefit from this? Ideal candidates are brands that have a narrow core
assortment of recurring products and collections — for example, Lululemon. Each round of
feedback-inspired changes can guide the subsequent season of orders and, over time, the AI
driving that core assortment should become smarter in quickly discarding options that don't work
(e.g., certain colorways or silhouettes) and ordering bigger runs of the items that do work.
BERNSTEIN
64 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
Who else is doing this? It's early stages, but we are seeing some larger brands begin
to incorporate direct consumer feedback into product designs — e.g., Nike and Adidas.
Nike welcomes design ideas from anyone at any time and, in the past, held competitions with
fans submitting and voting on new design(s) to be released on Air Max Day (Exhibit 15). On a
similar note, in February 2023, Adidas announced its inaugural Consortium Cup competition
— a four-week-long footwear design tournament with 16 designs from the world's most iconic
sneaker stores (e.g., Offspring, Overkill, Nice Kicks, Xhibition, Extra Butter) going head-to-head,
with confirmed members voting on their favorite designs (Exhibit 16 and Exhibit 17). This makes
for a limited collection on a specific product that also has a very strong marketing component to
it — but one could imagine that the success of an initiative like this could lead to more crowd-
sourced idea generation and feedback collection across the consumer base to inspire either new
products or tweaks to existing collections.
EXHIBIT 14: AI helps highlight a variety of attribute sets
that are likely to be well-loved for human designers to
review
Source: Company website
EXHIBIT 15: The six winning sneakers from Nike's 2018
"On Air" design competition
Source: Nike
EXHIBIT 16: Adidas' Inaugural Consortium Cup features 16
teams, 30 designs, four rounds, and one champion
Source: Company disclosures
EXHIBIT 17: Offspring's design came out on top
Source: Confirmed app
WAREHOUSE MATCHING Most retailers have predetermined paths for distribution and delivery. If a retailer is
organized regionally with the same assortment in every regional distribution center (DC), then
customers in a region will always be served by that region's DC. If a retailer is organized by
specialty, for instance, using different warehouses for different categories (e.g., bulky home
furnishings require a very different warehouse infrastructure than delicate leather handbags,
small jewelry, apparel, or boxes of shoes), then depending on what a customer orders, that order
BERNSTEIN
US APPAREL & SPECIALTY RETAIL: TO BUILD OR TO BUY? OPPORTUNITIES FOR AI IN RETAIL 65
is fulfilled by one or more of the specialty DCs. A third format is a hub and spoke model, which
helps consolidate orders into a single DC or pool before the last mile, thereby reducing last-mile
costs, but adds to the complexity of the DC network and adds further warehousing costs.
AI helps optimize the solution in real time. Each of these predetermined DC pathways
is inherently inefficient at times, depending on the capacity of each DC, the most direct
transportation route, how many other orders are also going from point A to B on that particular
day, etc. Using AI, companies can make each fulfillment decision in real time right after the order
is placed, optimizing the pathway of each individual order.
Stitch Fix uses AI to route each client to a specific warehouse based on what they
ordered. When a Stitch Fix shipment needs to be sent, an algorithm solves the binary
optimization problem to assign a client to a specific warehouse. The algorithm finds the global
optimum, i.e., the customer's warehouse assignment, based on a variety of variables — not just
shipping cost, but also shipping time and inventory match. Once a warehouse is selected, the
items the customer receives in their shipment all come from the same warehouse, which is most
convenient to the customer and helps avoid multiple shipments to the same address.
•This works best for Stitch Fix, which has a limited assortment and standard box size.
A Stitch Fix shipment generally only consists of five items in a standard-sized box, and across
a limited set of assortment (not including odd-sized items such as shoes or bags or hats), so
picking a DC in real time works. For a larger retailer that has different assortments in each DC,
employing a similar strategy of assigning a single warehouse to each client might not work.
•Freestyle also uses AI, though the mechanics of the shipment are different. Even with
a broader assortment and standard retail purchase model, AI can be used to solve a different
version of the same problem — i.e., what combination of warehouses to use to fulfill the order?
Or how best to send items to a single distribution pool (hub and spoke) to then package up
and send to the customer? Stitch Fix's Freestyle model (i.e., normal multi-brand retail across
a much wider assortment, without predetermined basket sizes) will often send customers
multiple packages if they have ordered items that are spread out across warehouses, which
are generally aligned by product line (e.g., women's, men's, kid's).
•Predictive recommendations help too. Stitch Fix is rather uniquely positioned in the
symbiosis between its warehouse selection AI and its recommendation engine. While
recommending products to add to the basket, AI can also optimize on warehouse selection,
recommending ones that might be easiest to ship or ones that ship from the same warehouses
as other items in the basket, thereby reducing the cost per item.
Another very different example of successful last-mile optimization is Inditex (not
covered). The key to Inditex's warehouse selection AI is its ability to use the company's vast store
network as a potential warehouse, thereby opening up thousands of options of how to fulfill
each order. Therefore, it has a smaller network of omnichannel distribution centers, plus a much
larger network of omnichannel stores acting as mini distribution hubs (including staff trained
to package up orders, label printing capability, and packaging on site). As each order is placed,
Inditex's stock integration system uses RFID to identify all the possible "warehouses" (including
stores) that have the stock available to fulfill that order, and then picks the optimal fulfillment
option based on delivery time, cost, and existing shipments that are also scheduled to follow a
similar route that can share truck capacity.
BERNSTEIN
66 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
As more omnichannel brands and retailers have RFID-tagged SKUs, this is a viable
solution that could considerably optimize last-mile deliveries and trim distribution costs.
Within our coverage, Nike, Adidas, and Lululemon have RFID-tagged products and a decent
store footprint, the two minimum requirements to make a store-based warehouse optimization
process work. Nike's and Adidas's ongoing investments into logistics buildout suggest that we
may see a similar dynamic real-time fulfillment process from either or both brands in the near
future.
EXHIBIT 18: Stitch Fix assigns a warehouse for each
shipment request based on metrics such as cost, time,
and inventory match
Source: Company website
EXHIBIT 19: Even Stitch Fix's Freestyle offering makes
recommendations for clients
Source: Company website
BUILD? OR BUY? Increasingly, players in the Apparel & Footwear sector are recognizing the importance of
integrating AI capabilities into their business model to remain competitive, whether that be via
predictive recommendations, warehouse matching, or product design, as we've explored in this
chapter, or a wide range of other applications (e.g., pricing, inventory management, etc.).
What does this mean for the sector? Given the uneven progress across the sector, not
correlated with size or data availability, we expect to see a sort of VC model evolve, where
emerging data-rich retailers make AI foundational to their business model (e.g., Stitch Fix) and
build these capabilities, and mature retailers/brands that have the scale, but not the technology,
opt to buy, either by buying the start-ups altogether or buying the data (or the technology as a
service) — both the builders and the buyers can differentiate and win.
As a builder, Stitch Fix's potential lies in selling its data or as a potential acquisition
target. Currently, Stitch Fix's main limitation comes from the limited TAM of its main product, the
Fix, and the lack of differentiation in its Freestyle offering. However, Stitch Fix's strengths in data
make for a great add to large retailers who could implement Stitch Fix's technology on a larger
scale, whether that be through acquisition or buying its technology as a service, which Stitch Fix
currently does not sell.
On the other end of the spectrum, Nike has bought its way into some of these capabilities
via acquisitions. Over the last five years, Nike has made several acquisitions to beef up its AI
capabilities — Datalogue (2021), a data-integration startup; Celect (2019), a predictive analytics
platform; Zodiac (2018), a consumer data analytics company; and Invertex (2018), a computer
vision company. With its large scale and deep pockets, Nike is much better able to scale them up
and maximize ROIC across its near US$20Bn e-commerce business.
BERNSTEIN
US APPAREL & SPECIALTY RETAIL: TO BUILD OR TO BUY? OPPORTUNITIES FOR AI IN RETAIL 67
But companies with scale can also opt to buy the data (i.e., tech-as-a-service). Many
retailers in the A&F space have opted to partner with AI tech companies instead — for
example, TJ Maxx (owned by TJX Companies), Kate Spade (owned by Tapestry), and Famous
Footwear (private) have partnered with Coveo (not covered), a software-as-a-service search
engine that optimizes e-commerce websites for product discovery through a variety of tools
such as personalized recommendations, relevant product suggestions during search, prompts,
and social proof badges. Another example comes from Farfetch, which has partnered with
Aurora Mobile, leveraging its machine learning-based push notification services and intelligent
operational analytics to provide a smarter and more personalized retail experience for customers.
Who are the losers? The losers will be retailers/brands that do neither, and those who try to do
too much at once and fail to integrate new technology investments into the core operations of the
business. Brands that are late to investing into these new strategies will likely find themselves left
in the dust as consumer expectations for personalization, convenience, and appealing designs
grow.
For the latter, we can take the end of Stitch Fix's cousin Trunk Club, acquired by
Nordstrom (not covered), as an example. Though Nordstrom had purchased Trunk Club, it
was largely for the purpose of simply adding a subscription-based offering, rather than for the
company's AI capabilities. As "box fatigue" hit, so came the eventual demise of Trunk Club in
2022. In this case, the builder of Trunk Club came out as a winner, purchased in 2014 for US
$350Mn, whereas Nordstrom came out as a loser, taking a US$197Mn write-down on the service
three years after buying it. While buyers can buy start-ups with novel new AI technologies, the
onus is on them to then scale up and integrate the technology into the rest of the business.
VALUATION METHODOLOGY See Disclosure Appendix for valuation methodology.
BERNSTEIN
68 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
EXHIBIT 20: Rating and price target
Stitch Fix
SFIX.US
Rating M
Prices as of May 25, 2023 3.46
Currency USD
Target Price 4
52-Week Range 2.63 - 9.13
Market Capitalization (million) 389
TTM Performance -54.2%
TTM Relative Performance -60.4%
Bernstein EPS Forecast
2021A (0.08)$
2022A (1.90)$
2023E (1.54)$
2024E (0.77)$
2025E (0.51)$
EPS Annual Change
2021A-2022A
2022A-2023E
2023E-2024E
2024E-2025E
Consensus EPS
2023E -1.53
2024E -1.00
2025E -0.78
P/E on Bernstein EPS Forecast
2023E -2.3x
2024E -4.5x
2025E -6.9x
Shares Outstanding (mil.) 87
Yield n/a
Dividend per Share 0.00
Benchmark TTM relative performance SPX Index
TTM 6.1%
Price 4151.28
Stocks trading in the US are benchmarked against the S&P 500 Index, which had a
closing price of USD 4151.28 as of May 25, 2023.
Source: Bloomberg L.P., corporate reports, and Bernstein analysis and estimates
Source: Bloomberg, company disclosures, Bernstein analysis and estimates
RISKS See Disclosure Appendix for risks.
Aneesha Sherman aneesha.sherman@bernstein.com +1 212 969 1469
Shradha Mani shradha.mani@bernstein.com +1 212 969 1475
Jessica Tian jessica.tian@bernstein.com +1 212 969 1451
BERNSTEIN
US APPAREL & SPECIALTY RETAIL: TO BUILD OR TO BUY? OPPORTUNITIES FOR AI IN RETAIL 69
BERNSTEIN
70 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
GLOBAL LUXURY GOODS: LVMH AND
FARFETCH LEAD THE AI REVOLUTION
HIGHLIGHTS •The "Digital Revolution" has unlocked unprecedented amounts of information that luxury
brands can leverage to enhance consumer experiences and optimize decision-making.
This information can be collected at multiple checkpoints: (1) consumer navigation and
purchase online, and (2) in-store visits and purchase. AI can help luxury brands analyze
the data and make informed decisions on several fronts, such as: (1) store assortment
optimization and inventory management, (2) customer relationship management, and
(3) collection briefing and architecture. We are seeing numerous solutions and service
providers coming to the market with ad hoc applications, offering a wide array of solutions
that cater to the luxury industry's unique requirements. The most relevant AI applications
we have identified range from personalization and recommendation engines or trend
forecasting and predictive analytics to sustainability optimization (Exhibit 1).
•Major luxury companies have started embracing AI. As AI applications continue to evolve,
luxury brands that effectively integrate these innovations into their business strategies will
be well-positioned to succeed in the digital era. That said, it is essential to remember that
brand desirability remains the cornerstone of success in the luxury industry. Factors such
as product quality, design, heritage, and exclusivity continue to play a significant role in
driving customer loyalty and sales. In short, AI could enhance brand desirability by improving
customer experiences and optimizing various aspects of luxury brands' operations — but not
replace it.
•LVMH and Farfetch are leading the way. LVMH has established the LVMH Innovation Award
to recognize and promote cutting-edge AI solutions. It has partnered with several AI start-
ups, including Heuritech (private)for trend forecasting, Marqvision (private) for counterfeit
detection, Data&Data (private) for distribution channel management, Replika Software
(private) for social selling, OneStock (private) for omnichannel and inventory optimization,
and Kuaizi (private) for marketing strategy optimization. These partnerships reflect LVMH's
commitment to leveraging AI technologies across various applications. As Farfetch seeks
to transition from a marketplace to a pioneering "tech enabler" within the luxury sector, it
has been at the forefront of adopting AI and digital technologies to enhance its services.
For instance,it partnered with VIPER and Syte (private)for image recognition, and Aurora
Mobile (private) for personalized services to its customers. While we acknowledge Farfetch's
commendable efforts in embracing AI, we believe the primary challenge for digital luxury
platforms is not technology but generating traffic. As luxury brands prioritize their own
websites, multi-brand digital luxury distributors continue to face an uphill battle for survival.
INVESTMENT IMPLICATIONS We would continue to stay exposed to high-quality names. Our preference in this group continues
to be LVMH. A benign demand environment should be conducive to higher beta names yielding
relative outperformance. Kering, Prada, and Swatch Group top our list here and appear as good
candidates for fringe portfolio positions. Clearly, high Chinese exposure is a plus, as the sector
is to benefit from a high-end Chinese spend rebound — in China and abroad. This has propelled
BERNSTEIN
GLOBAL LUXURY GOODS: LVMH AND FARFETCH LEAD THE AI REVOLUTION 71
Richemont, Burberry, and Moncler as well. Yet, for different reasons, we view these as a second
priority at this point.
THE DATA REVOLUTION The "Digital Revolution" has unlocked unprecedented amounts of information that luxury brands
can leverage to enhance customer experiences and optimize decision-making. This information
can be collected at multiple checkpoints: (1) consumer navigation and purchase online, and
(2) in-store visits and purchase (Exhibit 2).
EXHIBIT 1: The most relevant AI applications in the luxury industry include improving "consumer dialog" and
optimizing "management processes"
New Media
Social media and
influencers are replacing
glossy magazines
Channel conflict
Off-price, wholesale and
grey market (once
tolerated) become a
liability and a hurdle
New Distribution
E-commerce and m-
commerce are growing
double digits
New Products
Smart watches are
killing entry price Swiss
made products
New Brands
lower distribution fixed
costs = lower barriers to
entry. New entrants in
eyewear, footwear,
fashion, watches, etc.
Consumer dialogue
Brands now know
virtually all of their
customers by name:
what will they tell them
(CRM)?
New Retailers
YNAP, Farfetch, Alibaba,
Amazon, Chrono24, and
many other
Management
Processes
New data allows
more “scientific”
decision making,
previously based on
experience and gut feel
Source: Bernstein analysis
EXHIBIT 2: Relevant information can be collected at multiple touchpoints: consumer navigation and purchase
online, and consumer in-store visits and purchase
PR
Outdoor / Print / TV / …
Word of Mouth
Online Ads
Viral Email
Digital Billboard
Search
Landing Page
Social media
3rd Party Sites
Direct Mail
Store
Call Center
Brand.com
Mobile app
IM/Chat
Call Center Customer Services Promotion on Invoice
Social media
Email
Newsletter
0 5 10 15 20 25
Customer Journey Map
Customer navigation
Purchases Purchases
Online touchpoints
Physical touchpoints
Awareness Consideration Purchase Service Loyalty
Source: Wizaly, Bernstein analysis
BERNSTEIN
72 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
APPLICATIONS OF AI IN THE
LUXURY INDUSTRY
AI can help luxury brands analyze this data and make informed decisions on several fronts, such
as:
•Store assortment optimization and inventory management,
•Customer relationship management, and
•Collection briefing and architecture (Exhibit 3).
EXHIBIT 3: AI can help luxury brands analyze vast amounts of data and make informed decisions on several fronts
Store
assortment
optimization
and inventory
management
Customer relationship
management
Collection
briefing and
architecture
Source: Bernstein analysis
We are seeing numerous solutions and service providers coming to the market with ad
hoc applications, offering a wide array of applications that cater to the luxury industry's
unique requirements (Exhibit 5 and Exhibit 6). We have identified the 10 most relevant AI
applications (Exhibit 4).
EXHIBIT 4: Here are the top 10 most relevant AI-powered applications we have identified
Source: Bernstein analysis
BERNSTEIN
GLOBAL LUXURY GOODS: LVMH AND FARFETCH LEAD THE AI REVOLUTION 73
EXHIBIT 5: We see a growing AI ecosystem in the retail and luxury industries
Source: Business of Fashion
EXHIBIT 6: Personalization and customer analysis are the two most powerful AI tools in terms of financial impact,
according to "ChatGPT"
AI-generated impact scores on selective AI-powered applications in luxury industry
Supply Chain
Optimization
Customer Analysis
Sustainable Design
Customer Engagement
Social Responsibility
Personalized
Recommendations
Virtual Try-ons
(20)
-
20
40
60
80
100
120
(20) - 20 40 60 80 100 120
ESG
Financial
AI-supported initiative impact score
Customer appeal
Source: SlackGPT, Bernstein analysis
BERNSTEIN
74 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
THE MAJOR COMPANIES
IN OUR COVERAGE HAVE
STARTED EMBRACING AI
As AI applications continue to evolve, luxury brands that effectively integrate these innovations
into their business strategies will be well-positioned to succeed in the digital era. That said, it is
essential to remember that brand desirability remains the cornerstone of success in the luxury
industry. Factors such as product quality, design, heritage, and exclusivity continue to play a
significant role in driving customer loyalty and sales. In short, AI could enhance brand desirability
by improving customer experiences and optimizing various aspects of luxury brands' operations.
LVMH and Farfetch are leading the way.
LVMH has established the LVMH Innovation Award to recognize and promote cutting-edge AI
solutions (Exhibit 7). The company has partnered with several AI startups, including Heuritech
for trend forecasting, Marqvision for counterfeit detection, Data&Data for distribution
channel management, Replika Software for social selling, OneStock for omnichannel and
inventory optimization, and Kuaizi for marketing strategy optimization (Exhibit 8 to Exhibit 14).
These partnerships reflect LVMH's commitment to leveraging AI technologies across various
applications. They also demonstrate how LVMH's own success is due in part to the ongoing
dialogue between its 75 Maisons and the world of startups, a constant source of creativity.
EXHIBIT 7: LVMH has established the LVMH Innovation Award to recognize and promote cutting-edge AI solutions
Source: Viva Tech, Bernstein analysis
EXHIBIT 8: Three of the six LVMH Innovation Award winners are AI-powered applications
2022 2021 2020 2019 2018 2017
Immersive Digital Experiences Bitski Data&Data Evrythng Oyst
Employee Experience, Diversity & Inclusion Gamino Each One
Image & Media for Brand Desirability SeenThis Aglet Euveka
Omnichannel & Retail The ShowCase Bambuser Onestock Slyce Kronos Care
Operations Excellence Toshi Hipli DigitalGenius
Sustainability & GreenTech WeTurn Galy Desserto
Data and AI Special Mention MarqVision Data&Data Crobox 3DLook VeChain Heuritech
AI-powered Applications Winner
LVMH InnovationAward Finalists
Source: Company websites, Bernstein analysis
BERNSTEIN
GLOBAL LUXURY GOODS: LVMH AND FARFETCH LEAD THE AI REVOLUTION 75
EXHIBIT 9: The company has partnered with several
AI startups, including Marqvision for counterfeit
detection…
Company: Marqvision
Year founded: 2020
HQ: Los Angeles, the USA
Description: Marqvision helps global brands identify and remove
counterfeits from more than 1,500 online marketplaces across the
world. Counterfeiting is a massive and growing threat worldwide, and
MarqVision is on a mission to protect creativity and innovation with
technology that allows brands to automatically monitor and protect their
IPs. Harnessing image recognition and natural language processing, this
AI-powered SaaS makes it faster than ever before to take down
counterfeits.
'MarqVision’s technology comes at a time when the global counterfeit
market is exploding, as it is projected to grow another 50% this year to
reach nearly $3 trillion in 2023. The company’s technology enables
efficient removal of counterfeits end-to-end by automating the traditional
anti-counterfeiting process. Its proprietary AI models detect counterfeits
with 95%+ accuracy and remove counterfeit sales at scale.
Signature product / technology:
Image Recognition
Our deep-learning-based image recognition model scans through
millions of product listings in order to quickly and accurately find
potential counterfeits with images of authentic products.
Machine Learning
Our machine learning model detects infringements based on listing
information such as price, product description and customer reviews,
while at the same time self-improving its accuracy based on patterns it
finds among confirmed infringements.
Bot-Powered Reporting
Our bot-powered reporting system automatically classifies different
infringement patterns and automates the process of submitting
takedown requests across thousands of marketplaces in 115
countries.
Partnerships:
LVMH:
Three of the LVMH Maisons have already selected MarqVision as
their brand protection provider.
Source: Company website, Bernstein analysis
EXHIBIT 10: …Replika Software for social selling…
Company: Replika (Founded in 2016, New York, the USA)
Description: Replika’s suite of tools for social sellers is designed to
help brands activate their sales associates and brand ambassadors to
sell more effectively online.
Signature product / technology:
Turnkey Social Selling Solution
Sell online, Inspire on Social Media & Connect with Consumers
Curate
Curate directly on brand’s website and generate trackable links
nspire
MyReplika includes a content suite to create inspiring
communications, social media posts and newsletters
Connect and Share
MyReplika is compatible with every social media platform
Partners
Source: Company website, Bernstein analysis
BERNSTEIN
76 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
EXHIBIT 11: …Heuritech for trend forecasting…
Company: Heuritech (Founded in 2013, France)
Heuritech, founded by two PhDs in Machine Learning, provides analysis
of insights drawn from the consumer throught the analysis of 3 million
images daily shared on social media. Heuritech can recognize over 2000
apparel details.
Signature product / technology:
The AI forecasting methodology
1 - Define representative panels - From hand-picked fashion-
forward influencers to AI-built mainstream segments
2 - Apply computer vision technology - To millions of social media
images to provide market insights. More than 3000 fashion details can
be spotted like shapes or fabrics.
3 - Use machine learning forecasting algorithms - 90% accuracy.
The forecasting toolkit includes a master algorithm that uses optimal
ensembling
4 - Insert Heuritech’s data into the market intelligence platform -
Creative and analytical profiles can access images, analysis, market
forecasts
Partners:
Source: Company website, Bernstein analysis
EXHIBIT 12: …Data&Data for distribution channel
management…
Company: Data&Data (Founded in 2012, France)
Description: Data&Data analyzes large volumes of transactions
carried out on international marketplaces to help luxury companies fine-
tune their marketing strategies. Data&Data participated in the fourth
season of LVMH’s La Maison des Startups accelerator program at
Station F.
Signature product / technology:
Grey market monitoring
Understand the extent of grey market exposure in respect with your
top competitors, identify grey market feeders
Pre-owned market monitoring
Capture a global overview of the online pre-owned market. Identify
trends impacting the positioning, pricing, and partnership strategies
Competitive intelligence
Determine your competitors’ standing and moves. Gain a clear picture
of your brand’s online visibility
Partners
Source: Company website, Bernstein analysis
EXHIBIT 13: …OneStock for omnichannel and inventory
optimization…
Company: OneStock (Founded in 2015, France)
OneStock is a stock unification system, that eradicates stock-out by
making products available on all sale channels.
(1) For brands : Icrease order volumes, sell stock faster and improve
profitability. (2) For consumers: Optimized and trackable delivery service,
with ability to return the product anywhere.
Signature product / technology:
Order Management Solution (OMS)
Optimise operations with OneStock’s powerful OMS business
intelligence. The OMS traces activities across every department, from
the moment an order is placed through to final delivery.
The SiSense BI
A solution based on SiSense technology. Options include easy
filtering and sorting and analyses.
Partnerships:
Source: Company website, Bernstein analysis
EXHIBIT 14: …and Kuaizi for marketing strategy
optimization
Company: Kuaizi (Founded in 2013, China)
AI-enabled automatic content production platform that also analyzes
performance of previous content.
Kuaizi in the service of science and technology content of commercial
ecological intelligent creative technology provider, AI methodology
based on the content element deconstruction, cloud computing, creative
content data such as core technology, through the creative intelligence
production, operation optimization, labels, insight, collaborative
management one-stop SaaS solutions, links to the global ecological
whole digital content business link, speed up the tens of millions of
brands, Internet companies and content providers to achieve business
growth.
Signature product / technology:
Intelligent production tools
Intelligent image and video splitting: AI fragment video mirroring,
based on point mirroring logic and script strategy, production and
recommendation of high-quality videos, recombination of image
elements, idea generation etc.
Script, mirroring, and dimensional analysis
Real-time feedback of creative data, dynamic visualization display,
clear creative effect data
Partners
Source: Company website, Bernstein analysis
BERNSTEIN
GLOBAL LUXURY GOODS: LVMH AND FARFETCH LEAD THE AI REVOLUTION 77
Kering has invested in AI startups and is focused on "striving for operational excellence
through AI." Kering's efforts in AI adoption, though less publicized than LVMH's, demonstrate its
dedication to innovation and enhancing customer experiences (Exhibit 15 and Exhibit 16).
Hermès has been equally discreet about its AI initiatives, but is known to be leveraging AI to
improve its online customer experience and optimize supply chain management.
EXHIBIT 15: Kering focuses on "striving for operational excellence through AI"
Source: Company CMD
EXHIBIT 16: Kering's efforts in AI adoption demonstrate its dedication to innovation and enhancing customer
experiences
Source: Company CMD
BERNSTEIN
78 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
ABILITY TO GENERATE
TRAFFIC — NOT TECHNOLOGY
— IS KEY TO DIGITAL RETAIL
SUCCESS
As Farfetch aims to transition from a traditional marketplace to a trailblazing "tech enabler" in the
luxury industry, it has taken the lead in adopting AI and digital technologies to enhance its services
(see reports: Farfetch: CMD kisses the prince into a frog and Farfetch: Follow the money!) (Exhibit
17). For example, Farfetch has implemented Semantic Search to better comprehend customer
search intentions and clarify ambiguous queries. The company also introduced VIPER, an AI-
powered platform for visual information extraction, which aimed to generate visually high-quality
item descriptions. In addition, Farfetch has partnered with Syte, a visual AI startup, to enable
its in-app visual search feature on iOS (Exhibit 18). This feature allows users to upload images,
which the app then analyzes and presents the corresponding product or a similar one, seamlessly
bridging online and offline inspiration. Moreover, Farfetch has collaborated with Aurora Mobile
(Exhibit 19) to provide an "unparalleled shopping experience to global customers." By leveraging
Aurora Mobile's AI-driven technology, intelligent operational analytics, and machine learning-
based push notification services, Farfetch can craft intelligent retail experiences and offer more
effective and personalized services to its customers.
While we acknowledge Farfetch's significant efforts in adopting technology, we believe the
primary challenge for digital luxury platforms is not technology, but generating traffic. As luxury
brands prioritize their own websites, multi-brand digital luxury distributors such as Farfetch
continue to struggle for survival (Exhibit 20 to Exhibit 22) (also see the Blackbook Global Luxury
Goods: The Devil Is in the Retail).
EXHIBIT 17: Farfetch has been at the forefront of adopting AI and digital technologies to enhance its services
Source: Company reports, Bernstein analysis
BERNSTEIN
GLOBAL LUXURY GOODS: LVMH AND FARFETCH LEAD THE AI REVOLUTION 79
EXHIBIT 18: Farfetch has partnered with Syte, a visual
AI startup, which enables users to upload images, and
then presents the corresponding product or a similar
one
Company: Syte (Founded in 2015, Israel)
Connecting Shoppers With Products They’ll Love. Founded in 2015 as an
image search company, Syte empowers shoppers to instantly find items
they’ll love with inspiring, visual product discovery journeys that drive
conversion and increase lifetime value.
As the world’s first Product Discovery Platform, we use visual AI to create
intuitive search and discovery experiences for all types of shoppers. Our
solutions include visual search, automated product tagging, advanced
personalized recommendations, shoppable social curation, and more.
Signature product / technology:
Visual Discovery
Help shoppers instantly find items they want with an inspiring visual
search experience, powered by AI.
AI Tagging & Merchandising
Drive revenue, boost text search, and optimize merchandising with AI-
enriched automatic product tagging.
Hyper-Personalization
Provide unique recommendations based on shoppers’ visual
preferences and current journey on your site.
The World’s First Product Discovery Platform
Syte is pioneering the future of search and discovery with a unified
platform that drives revenue through hyper-personalized, intuitive
customer experiences.
Partners
Source: Company websites, Bernstein analysis
EXHIBIT 19: Farfetch has collaborated with Aurora Mobile
to deliver tailored intelligent retail experiences, and
provide more effective and personalized services to its
customers
Company: Aurora Mobile (NASDAQ: JG, not covered, founded in 2011,
China)
Aurora provides a comprehensive suite of services to mobile app
developers in China. The Company utilizes Artificial intelligence (AI) and
machine learning to gain actionable and effective insights from its data
and to develop and refine its data solutions. Leveraging these
technologies built upon its massive and quality data foundation, the
Company has developed a variety of data solutions that offer industry-
specific, actionable insights for customers in a number of different areas.
Aurora is also in the process of developing and launching new data
solutions that will further leverage its data and insights to increase
productivity for additional industries and customers.
Signature product / technology:
The Company's core data solutions include:
Targeted marketing ("XiaoGuoTong")
Help advertisers improve their marketing effectiveness by enabling
them to target the right audience with the right content at the right
time.
Financial risk management: Assist financial institutions and financial
technology companies in making informed lending and credit
decisions.
Market intelligence
Provide investment funds and corporations with real-time market
intelligence solutions, such as the Company's product iApp, which
provides analysis and statistical results on the usage and trends of
mobile apps in China.
Location-based intelligence ("iZone")
Help retailers and those from other traditional brick-and-mortar
industries, such as real estate developers, track analyze foot traffic
and, conduct targeted marketing and make more informed and
impactful operating decisions, such as site selection.
Partners
Enable New luxury retail through AI and digital technologies
Source: Xueqiu, company reports, Bernstein analysis
BERNSTEIN
80 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
EXHIBIT 20: Burberry's social store tried to combine in-store and online data for improved AI appliations; however,
this effort faced issues due to a complex data collection process and a lack of clear value for consumers
Source: Company press release, Bernstein analysis
EXHIBIT 21: Connected mirrors at Browns store are an
entertaining add-on, but we struggle to find their real
benefit to customers
Source: Company reports, Bernstein photos and analysis
EXHIBIT 22: Likewise, the launch of Smart Tags in 3Q21 in
Browns stores aimed to boost CRM and personalization,
but it remains uncertain how these innovations truly
benefited customers
Source: Company reports, Bernstein analysis
BERNSTEIN
GLOBAL LUXURY GOODS: LVMH AND FARFETCH LEAD THE AI REVOLUTION 81
VALUATION METHODOLOGY See Disclosure Appendix for valuation methodology.
EXHIBIT 23: Ratings and target prices
Ticker Rating Currency Closing Price Target Price Ticker Rating Currency Closing Price Target Price
BRBY.LN M GBp 2,192.00 2,347.00 1913.HK O HKD 52.45 71.00
CFR.SW M CHF 146.65 152.00 UHR.SW O CHF 278.30 374.00
EL.FP M EUR 174.40 193.00 KER.FP O EUR 510.40 681.00
FTCH.US M USD 4.78 6.00 MC.FP O EUR 823.40 999.00
RMS.FP M EUR 1,914.80 1,819.00 MSDLE15 1,848.02
MONC.IM O EUR 61.98 64.00 SPX 4,151.28
Source: Bloomberg (pricing as of May 25, 2023)
Source: Bloomberg, Bernstein analysis and estimates
RISKS See Disclosure Appendix for risks.
Luca Solca luca.solca@bernstein.com +44 207 959 4884
Renny Shao renny.shao@bernstein.com +44 207 170 0614
Clementine Flinois clementine.flinois@bernstein.com +44 207 170 0653
BERNSTEIN
82 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
OCADO: AI IN ONLINE GROCERY
AUTOMATION AND OPTIMIZATION
HIGHLIGHTS •Ocado isn't a traditional automation hardware reseller, but instead enters into
partnerships with grocers, offering its full-technology stack from front-end to
delivery routing to their partners. We think most of the value of Ocado's technology sits
within its differentiated software offering (rather than the hardware) and think the most
differentiated areas are around inventory management and picking (including robotic
arms).
•Ocado's technology has a single view of stock throughout its warehouse that
enables it to improve forecasting, reduce markdowns through intelligent discounting,
improve availability for consumer orders, and optimize the grid for more efficient
picking (reacting in real time to new orders and external factors such as weather). On the
picking side, Ocado's on-grid robots collaborate in real time (moving at 4m/s in herds of
1,000s) to work together to pick at a rate of >200 units per hour (UPH) (well above the store
pick rate of 60-100). Its new robotic arms are also game-changing, raising pick rates to
>300 UPH with the use of machine vision to enable single-item picking without the arm
knowing what an item is and, more importantly, packing it efficiently to avoid disappointed
customers.
•Elsewhere, there are significant benefits on the front-end (to enable greater
personalization on the website, to grow basket sizes through recommendations, and
through personalized advertising by suppliers) and in the delivery and routing area
with real-time optimized routes, flexible order allocation (to maximize drop densities and
enable shorter lead time orders), and eventually autonomous vehicles (although most of that
benefit in the medium term will accrue to automated guided vehicles (AGVs) in the warehouse
to move stock around).
INVESTMENT IMPLICATIONS We rate Ocado Outperform with a target price of £14.50. It is a transformative structural winner
with differentiated technology, impressive unit economics, and a strong pipeline. It is riding the
wave of online grocery penetration globally. Its technology is market-leading and differentiated,
with years of experience running an online business.
BERNSTEIN
OCADO: AI IN ONLINE GROCERY AUTOMATION AND OPTIMIZATION 83
EXHIBIT 1: Impact of AI on online grocery and Ocado
Source: Bernstein analysis
IMPACT OF AI ON ONLINE
GROCERY
Introduction
Ocado isn't a traditional automation hardware reseller, but instead enters into partnerships
with grocers, offering its full-technology stack from front-end to delivery routing to their
partners. We think most of the value of Ocado's technology sits within its differentiated software
offering (rather than the hardware) and think the most differentiated areas are around inventory
management and picking (including robotic arms). This leads to significant benefits such
as -150-250bps lower product wastage, -300-500bps lower picking costs, and an overall
enhanced customer experience (with better availability, more on-time in full orders, etc.).
You might think that a chapter about grocery shopping and AI is odd, but we think there
are significant enhancements throughout the value chain driven by applications of AI. Our
definition of AI here is not a purist one, and we certainly include elements of machine learning
and predictive learning. However, we think the whole suite of data and analytics helps power
Ocado's technology. As we outline in Exhibit 1, we think the areas of the highest impact are
inventory management and picking. Ocado's technology has a single view of stock throughout
its warehouse that enables it to improve forecasting, reduce markdowns through intelligent
discounting, improve availability for consumer orders, and optimize the grid for more efficient
picking (reacting in real time to new orders and external factors such as weather). On the picking
side, Ocado's on-grid robots collaborate in real time (moving at 4m/s in herds of 1,000s) to work
together to pick at a rate of >200 UPH (well above the store pick rate of 60-100). Its new robotic
BERNSTEIN
84 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
arms are also game-changing, raising pick rates to >300 UPH with the use of machine vision
to enable single-item picking without the arm knowing what an item is and, more importantly,
packing it efficiently to avoid disappointed customers.
Elsewhere, there are significant benefits on the front-end (to enable greater personalization
on the website, to grow basket sizes through recommendations, and through personalized
advertising by suppliers) and in the delivery and routing area with real-time optimized routes,
flexible order allocation (to maximize drop densities and enable shorter lead time orders), and
eventually autonomous vehicles (although most of that benefit in the medium term will accrue to
AGVs in the warehouse to move stock around).
•Videos: See an introductory video on Ocado's use of AI Video (and many more interesting
videos) here, Ocado's 2022 Re:Imagined product launch found here, and Ocado's technology
blog with lots of tangible examples of AI usage here.
Front-end website
There are massive opportunities to optimize websites using AI to create more personalized
consumer experiences and improve on some of the consumer pain points (e.g., out of stock items,
missing items from baskets, etc.). AI can also help improve basket sizes by recommending new
products to consumers that they might decide to add to their baskets, and also through interstitial
pages (i.e., the pages you go through before you ultimately checkout). Many websites are in their
early days of personalization, but Ocado is well-positioned, having optimized its grocery website
for many years. Plus, it has an advantage versus other grocers in having an integrated full-stack
with front-end well-connected to inventory and order management.
•Personalization in grocery can be vastly improved by the application of AI by recommending
products that you have bought before (easy) to the more difficult task of recommending items
that they think you might want or recommending items you've previously bought (based on
your history and frequency of buying the items before — e.g., you typically buy dishwasher
tablets every 2.5 months, therefore, the company should proactively suggest these based on
historical usage). Ocado is in the early days of optimizing its personalized experience, but has
a long way to go with a significant opportunity. See blog posts here.
•This can also be manifested in the interstitial pages (Exhibit 5) before checkout, which can
boost basket sizes significantly. These interstitials use AI-based analytics to: (1) promote
your favorites; (2) promote items that you might have forgotten or that you've missed out
on offers for; and (3) promote items that are on flash sales (e.g., about to expire).
•Out of stock and availability status are vastly improved through full-stack integration of
the inventory in the CFC and the website. This means that the Ocado Smart Platform (OSP)
can update in real time the inventory on the website based on items sold, items ordered, and
the expiry dates of existing stock. One of the big problems historically with online grocery
ordering was the weak availability — i.e., you placed an order and only got 70-90% of your
items because the picker couldn't find them in store or they weren't available. With Ocado's
deep understanding of availability and the predictive analytics around orders and inventory, it
can improve availability significantly and reach availability of 99.9%.
•Advertising opportunity (Exhibit 2 to Exhibit 4): Ocado already earns one of the highest
rates of advertising as a percentage of sales versus global peers. Part of this is due to its
focus on prioritizing the opportunity as a pure play grocer. It has brand home pages and
BERNSTEIN
OCADO: AI IN ONLINE GROCERY AUTOMATION AND OPTIMIZATION 85
banner advertising. Ocado also enables brands to buy sponsored ads on specific products and
deliver personalized ads to specific customer types and/or demographics. This now makes
up around 4% of sales and is high-margin revenue.
EXHIBIT 2: Brands advertising in the catagory
Source: Company websites
BERNSTEIN
86 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
EXHIBIT 3: Large advertising banner for Quorn
Source: Company websites
EXHIBIT 4: Promoted products are listed as first option from search results
Source: Company websites
BERNSTEIN
OCADO: AI IN ONLINE GROCERY AUTOMATION AND OPTIMIZATION 87
EXHIBIT 5: Interstitial pages before check out recommending items that you bought before boost basket size and
improve economics
Source: Company websites
Inventory management
One of the key things that advanced machine learning and artificial intelligence can do for Ocado
and its partners is help optimize inventory within their automated warehouses. This is critical
because many retailers are losing 2-3% sales in shrink and product wastage at the moment,
whereas an Ocado CFC only loses 0.04% sales to product wastage. This is supported by AI
across a number of different areas.
Firstly, Ocado has full inventory visibility of all products stored within the warehouse. As they are
decanted (i.e., unboxed from suppliers) and put into totes (i.e., crates) that sit within the 3D grid,
the OSP registers both the inflow and outflow of product as well as key information such as when
a product may go out of date or expire. This is a vast improvement to a traditional store-based
supermarket where the retailers don't know which products are in the store and rely on store staff
to walk around the store checking availability and expiry dates (and marking down accordingly).
The role AI can play in inventory management is: (1) better forecasting of inventory requirements;
(2) management of availability across warehousing; (3) managing markdowns and product
wastage; and (4) optimizing the grid for the most efficient pick based on orders and other external
factors such as weather.
•Forecasting: OSP, supported by AI and machine learning, can also more accurately predict
when products are going to be needed in stock or when they're going to run out. It has lots
of data across 20 years of customer behavior and ordering patterns to understand the sell-
through of certain SKUs at certain times of year and with certain customers to understand
what to order and when. Also, as a result of full inventory visibility, it can only order what is
needed versus the stock held — this is very different to traditional store-based grocery, which
has algorithms to forecast demand but still relies heavily on human intervention to validate
BERNSTEIN
88 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
new product orders (i.e., the person you have seen wandering around a store in the evening,
scanning shelf-edge labels for out-of-stock items).
•Availability across warehouses: One of the key things that Ocado's new system called
Orbit can do is manage inventory for sell-through across the entire supply chain. Typically,
suppliers drop off products at one warehouse for the products to be sold at that warehouse
to customers. However, Ocado now has a single view of stock through Orbit (Exhibit 6),
which enables its different warehouses to feed off each other (using supply in one warehouse
to support another when it runs low or when it receives a large number of orders for a
specific product). The role that AI has here is critical — there is no human intervention in the
management of stock levels across warehouses. OSP can organize the shipment of certain
products and SKUs to other warehouses or request stock from another warehouse when it
thinks it is going to need it. This reduces the level of safety stock required in the system, helps
improve product availability and, ultimately, improves customer experience.
•Markdowns and product wastage (Exhibit 7): One of the huge benefits to grocers is that
the OSP also knows when products are going out of date and therefore can mark them down
in real time at the right level to optimize sell-through at the highest margin. This reduces
product wastage from 2-3% of sales to 0.04% of sales. It can do this by seeing all the orders
coming in, seeing all the stock in the system, and predicting any future orders it might get
on a shorter lead time. The platform can accurately markdown products (due to full-stack
integration) to reasonable markdowns (i.e., 10-20%) in order to generate sell-through. Again,
this is different to traditional store-based grocery where a shop worker walks around at the
end of the evening and selects all the items that are going out of date and marks them down
to clearance (i.e., 50%+) in order to get rid of them. Lots of food in stores is still thrown away
due to the mismatch of supply and demand.
•Grid optimization: Another way AI and machine learning can optimize the warehouse for
picking is based on orders already placed and for external factors (such as the weather).
Typically, Ocado's CFCs store the fastest-moving SKUs toward the top of the grid for easy
picking and the slower-moving SKUs toward the bottom of the grid. OSP can see the orders
that are coming as they are placed and therefore can start optimizing the grid for future
demand (before it actually needs to pick the items). For example, if one slow-moving SKU has
been picked already and the grid knows that it needs to be picked again soon, it can store it
toward the top of the grid to save it digging to the bottom again. Equally, if the OSP knows that
it is going to be hot on a particular weekend, it can bring the burgers and BBQ items toward the
top of the grid because it will expect sell-through to be higher. This helps the overall efficiency
of the pick and the time taken to pick an order.
BERNSTEIN
OCADO: AI IN ONLINE GROCERY AUTOMATION AND OPTIMIZATION 89
EXHIBIT 6: Ocado Orbit — one warehouse shipping different products to the other to share inventory
Source: Company reports, Bernstein analysis
EXHIBIT 7: Ocado flash sales in the pages before you check out with markdown items that are going out of date
soon
Source: Company website
BERNSTEIN
90 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
Picking
Ocado's platforms already pick significantly faster than many store-based picking methods.
Ocado currently delivers throughput of >200 UPH by 60-100 in-store and will be able to deliver
>300 UPH with robotic arms. It is only through AI and machine learning in the OSP that bots
work at rapid speeds on the grid, and much of the additional future benefit will come through
AI-powered robotic arms, which will remove another few points of labor out of P&L. With the
automated picking interface as is, we think Ocado can save partners 300-500bps in labor costs.
This will increase considerably with robotic arms, which could save another 100bps+ of labor.
•On-grid robots (Exhibit 8) collaborate with each other to pick grocery orders. There are
swarms of thousands of them, and they move within 5mm of each other on the grid at speeds
of up to 4m/s (covering up to 60km per day). They are not directed by humans, but instead
work together and autonomously to pick 70-100 items per order. They work together by
digging through the grid (sometimes 10-12 totes deep) in order to find the right tote that
contains the items for orders. They are also able to communicate with each other to tell
each other where each robot is going to avoid collisions. This predictive communication is
patented by Ocado, and protected in the UK and the US. This R-Hive patent (Exhibit 10)
covers a software/analytical system for controlling robotic devices operational on a Rainbow
warehouse system. Compared with other types of traffic routing software, this technology is
unique because it allows the hundreds of bots on top of the grid to operate autonomously and
find routes around each other (reacting to behavior, unforeseen stoppages, and automated
tasks). It uses a cloud-based AI air traffic control to communicate with thousands of bots 10x
a second to enable greater efficiency. For example, a bot can find the best route from one
area of the grid to another, predict where other bots are going to be, and inform them that it is
going to cross their path. The aim is to increase speed and reduce the risk of a collision. See
a YouTube video of on-grid robots here.
•Robotic arms (Exhibit 9) are the next big frontier for picking efficiencies, taking pick rates
up from 200 UPH to 300+ UPH. There are still three manual processes in the warehouse: (1)
decant — unloading items into the CFC; (2) picking items; and (3) loading vans. Robotic arms
help solve the manual item pick. These arms can sit on the top of the grid and sort items from
storage totes into delivery totes. AI is important here as the arm needs to not only recognize
an item ordered by a customer, but understand how to pick it up and then pack it in the most
efficient way possible. This is no easy challenge, given the complexity and variability of grocery
— e.g., the arms need to understand how to pick eggs versus a bottle of Coca-Cola. Ocado
says this technology uses machine vision, reinforcement learning and advanced sensing to
pick items without any prior knowledge of what they are. The big difficulty is knowing how to
pack the items — i.e., knowing not to place something that is dense on top of a loaf of bread.
This is game-changing in that it alters the ultimate productivity of a site and enables much
greater throughput. See a YouTube video of on-grid robotic arms here.
BERNSTEIN
OCADO: AI IN ONLINE GROCERY AUTOMATION AND OPTIMIZATION 91
EXHIBIT 8: Ocado's on-grid robots
Source: Ocado YouTube channel
EXHIBIT 9: Ocado's on-grid robotic pick
Source: Ocado YouTube channel
BERNSTEIN
92 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
EXHIBIT 10: Part of R-Hive technology
Source: US District Court Document (District of New Hampshire)
Routing and delivery
One of the final areas where Ocado could apply more AI and machine learning, but is probably
in the earlier stages of optimizing, is routing and delivery. This can cost up to 10% of sales for a
retailer as it requires a person to drive a van and deliver orders to customers, often only managing
two to four orders per hour (see YouTube video here). Places where AI can be important are:
•Simple optimized routing (Exhibit 11): Ocado uses AI to optimize routing algorithms to
enable the most efficient delivery routes and maximize drops per hour. This factors in elements
such as real-time traffic, weather, and time of day to ensure that the routes are built most
effectively for drops. They can also recalibrate in real time to react to any delays. Ocado is also
leading in pushing environment-friendly slots to consumers (i.e., when a van is already in your
area), which also helps reduce costs.
•Flexible order allocation (Exhibit 12): One of Ocado's Re:Imagined technologies is Ocado
Swift Router, which enables last-minute, short lead time orders to be picked, packed, and
delivered alongside larger, longer lead time orders utilizing excess capacity in the same van.
This is good for consumers and for Ocado's partners as they can offer immediacy services.
However, it is also important for delivery densities and increasing drops per hour. Using AI,
Ocado can work out where there is excess space in vans and then create short lead time
delivery slots and even proactively promote these to consumers at reduced costs to reduce
overall delivery costs.
•Autonomous vehicles powered by AI would be the next frontier for Ocado. If they could
remove all labor from deliveries, it would be game-changing. However, it is years away. The
more interesting place for Ocado is with AGVs in warehouses to enable the shuttling around
of products from inbound (from suppliers) to the decanting bays, and from the automated
frameload stations to the van loading bays. It is still early days on this, but Ocado has invested
in Oxbotica (private), which is building software to support some of these use cases. See a
YouTube video here.
BERNSTEIN
OCADO: AI IN ONLINE GROCERY AUTOMATION AND OPTIMIZATION 93
EXHIBIT 11: Optimizing delivery drops per hour can
further reduce costs as a percentage of sales
Source: Bernstein analysis and estimates
EXHIBIT 12: Ocado Swift Router optimizes van utilization
and offers shorter lead time, smaller orders to be
fulfilled
Source: Company website
VALUATION METHODOLOGY See Disclosure Appendix for valuation methodology.
EXHIBIT 13: Ratings and target prices
25-May-2023 Target
Ticker Rating Currency Closing Price Price
OCDO.LN O GBp 402.90 1,450.00
EDM USD 1,123.01
Source: Bloomberg, Bernstein analysis and estimates
RISKS See Disclosure Appendix for risks.
William Woods william.woods@bernstein.com +44 207 959 4525
Eric Chen eric.chen@bernstein.com +44 207 170 0635
Alexander Nielsen alexander.nielsen@bernstein.com +44 207 170 0671
BERNSTEIN
94 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
BERNSTEIN
Industrials
INDUSTRIALS 95
BERNSTEIN
96 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
AI IN AGRICULTURE: RISE OF THE
GROWBOTS
HIGHLIGHTS •AI is already present in large agriculture, but adoption is in early stages. The latest
generation of precision ag uses AI to automate decisions throughout all parts of the growing
process (sense and act): using computer vision/machine learning to identify crops versus
weeds and only spray the weeds (reducing herbicide usage by ~80%), automated combine
harvesters that sense real-time conditions and adjust settings to optimize crop yield, and
smart planters that sense real-time soil conditions to deposit seeds in the optimal field
location (reducing seed usage and increasing yield). Greater adoption results in structurally
higher farmer profits and OEM pricing power.
•The next big opportunity to leverage AI to create value on the farm are autonomous
production systems. Precision ag represents a ~US$55Bn market opportunity for OEMs
(doubles the large agriculture OEM TAM). Adoption of autonomous production systems will
unlock a greater portion of this potential. Part will come from labor cost savings, but a greater
share will come from hitting the key agronomic windows during the growing process (i.e., for
each day farmers plant outside the optimal window they lose 1% yield). John Deere and CNH
lead the market in the commercial introduction of autonomy (tillage, spreader, etc.), but the big
value unlock will be fully autonomous production systems covering the full grow cycle (tillage,
planting, spraying, and harvesting), which Deere plans to launch by 2030.
•The journey from automation to autonomy will transform agriculture OEMs in three
ways: (1) The revenue model will shift to a greater share of recurring/subscription
business. This change will reduce cyclicality, increase gross margins, and could structurally
increase the mid-cycle multiple by 2-turns to 18x. Moreover, the usage-based model aligns
more with how farmers think about their operating costs and expands the addressable market,
enabling smaller scale farms to access this technology. (2) Barriers to entry will be higher,
favoring the largest OEMs. AI gets better as it trains on more data, so OEMs covering
the most acreage (John Deere has a +50% share in North America) will likely build more
productive autonomous products. Moreover, as equipment become more connected, farmers
will change how they measure ROI from individual equipment to production system, again
favoring the largest OEMs. (3) Density of technology stack will become a big competitive
differentiator. As discrete automated production steps become more integrated into a single
production system, OEMs will need to fully control each of the eight layers of the agriculture
OEM technology stack. Computer vision is emerging as one of the more important, and we
expect greater M&A focus on these layers and others. The large ag OEMs (John Deere, CNH
Industrial, and AGCO), who are well-capitalized, are best-positioned on this front.
INVESTMENT IMPLICATIONS We maintain our Market-Perform rating for AGCO Corporation and John Deere, and their
respective target prices of US$135 and US$398. We maintain our Outperform rating for CNH
Industrial and its target price of €17.
BERNSTEIN
AI IN AGRICULTURE: RISE OF THE GROWBOTS 97
WHERE DOES AI FIT INTO THE
PRECISION AG TECHNOLOGY
ROADMAP?
The most pressing issue facing the agriculture sector over the next 30 years is how to deliver the
70% increase in food production needed to support the anticipated 30% growth in population
at a reasonable price. Precision ag is emerging as the leading technology to deliver such
productivity growth, which is mainly focused on increasing crop yields (potential of a 3x increase)
and reducing input costs (40% of a farmer's cost structure). Precision ag began with GPS
(precision ag 1.0), has progressed to variable rate technology (precision ag 2.0), and the ultimate
destination is autonomy (precision ag 3.0). AI currently underpins much of the variable rate
precision ag technology in the market and will play a critical enabling role as the industry adopts
autonomous crop production systems (Exhibit 1 to Exhibit 3).
EXHIBIT 1: AI is the enabling technology that underpins the present and future of precision ag
2000 2010 2020 2030
GPS (Precision Ag 1.0)
Variable Rate Tech. (Precision Ag 2.0)
Autonomy (Precision Ag 3.0)
AI: The Enabling Technology
Source: Bernstein Analysis
EXHIBIT 2: Adoption of AI in agriculture could help unlock
a 3-4x increase in crop yields
27% 39%
73% 61%
Corn Soybeans
% of Potential Yield Realized
Realized Yield Unrealized Yield Potential
Max =
616 bu/
acre
Max =
120
bu/
acre
Current =
167 bu/acre
Current =
47 bu/acre
Source: Bernstein analysis
EXHIBIT 3: AI could enable disruption of ~40% of farmer
cost structure
23%
19% 17%
14%
5% 5% 5% 4%
0%
5%
10%
15%
20%
25% 40% of the Production
Cost is Susceptible to
Disruption
Source: USDA, Bernstein analysis
Conditions are right for broader adoption of AI in farming. Adoption of AI in a particular
industry is driven by several factors, and farming has many of these in place already. Availability of
large amounts of data is essential for the development and training of AI models. The complexity
of tasks, the need for automation, and the desire to improve operational efficiency are other
important drivers of AI adoption. Emergence of advanced AI technologies such as machine
BERNSTEIN
98 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
learning, natural language processing, and computer vision also drives AI adoption in addition to
the increasing demand for personalized experiences (Exhibit 4).
Problems AI can solve in farming. One of the primary benefits of using AI in farming is its ability
to help farmers make more informed decisions about crop management, pest control, and other
critical aspects of agricultural production. With AI-powered sensors and predictive analytics,
farmers can gather data on soil moisture levels, weather patterns, and other environmental
factors to optimize crop yields and reduce waste. Additionally, AI can help farmers identify and
respond to crop diseases and pest infestations in real time, preventing significant losses and
reducing the need for harmful pesticides. AI-powered drones and autonomous machinery can
also help farmers reduce labor costs, increase precision in planting and harvesting, and optimize
fertilizer and water usage. The use of AI in farming can help reduce the environmental impact of
agricultural practices by reducing waste and optimizing resource utilization. Finally, AI will reduce
labor intensity at the farms as autonomous farm equipment usage expands.
AI enables Sense & Act in farm equipment. AI is the underlying technology that will allow
farm equipment to sense the surroundings and take actions to optimize crop yields and input
costs. The ultimate end-game of deploying AI in agricultural equipment is the full deployment
of autonomous production systems. Reaching this goal begins with breaking down the different
steps of the crop production process and automating them. To that end, John Deere and AGCO
have been building their technology stacks scale using this piecemeal approach into a unified
system. As this technology matures, these decisions will increasingly occur at the plant-level
(versus field-level historically). Large agricultural equipment are best-positioned to deploy this
technology, given they physically touch the soil and are among the crops on a regular basis.
EXHIBIT 4: Conditions are right for further adoption of AI in farming
Factors that Drive AI Adoption Within Industries Present in Farming?
Large Amounts of Data
Yes - A single 2K farm can contain 66M plants.
Complex Tasks
Somewhat - While the individual tasks of farming are simple, aggregating
each step of the production process at the right time and in the right amount
creates complexity.
High Labor Intensity/Scarcity
Yes - Rural labor scarcity has reached acute levels.
The Need for Personalization
Yes - Precision ag is all about scaling actions down from the farm level to the
plant level - individualized treatment for each plant optimizes outcome.
The Need to Improve Operational Efficiency
Yes - Arable land growth is limited and the world will need to grow food
production by 70% to support a 30% increase in global population.
Existing Use of Enabling Technologies
Yes - OEMs have been adding such technology as machine learning and
computer to equipment over the last few years.
Source: Bernstein analysis
TECHNOLOGY STACK OF
LARGE AG OEMS
Deploying AI in the agriculture sector requires multiple layers of technology, which agricultural
equipment OEMs have built up over time through a combination of organic R&D and M&A. In
aggregate, these technologies enable the equipment to sense the environment and eventually
make growing decisions on the plant level. The first layer of the technology stack begins with
farm equipment (tractors, planters, sprayers, combines, etc). The next layer is the digital
operating platform, which allows the farmer to comprehensively manage the farm (builds a
digital twin of the farm, connects the equipment together and to farm management software).
John Deere's version of this technology is the John Deere Operations Center, AGCO's version is
BERNSTEIN
AI IN AGRICULTURE: RISE OF THE GROWBOTS 99
Field View, and CNH's version is AFS Connect. CNH has also partnered with Monarch Tractor
(private), which has its own agronomic and machine data management platform. Guidance,
aka GPS, creates spatial awareness for equipment in the field, which in many cases needs
precision down to less than an inch. John Deere has owned this layer of the tech stack for 20
years, but other OEMs are increasingly realizing they need to bring it in-house (AGCO, CNH).
Connectivity includes telematics, but the real value creation opportunity is connecting each
piece of equipment with another, so they can seamlessly share information across time and space
to make better growing decisions (i.e., moving data collected from tractor during fall tillage to
the planter during spring planting). CNH acquired NX9 software for its ISOBUS connectivity
technology. Computer vision begins with adding to farm equipment high resolution cameras,
which "sense" the environment through capturing images, which are then passed through a GPU
for processing and identification (is it a crop or a weed, is the obstacle ahead of the tractor
something that needs to be avoided). An area of further investment for agricultural equipment
OEMs will be improving depth and distance perception — John Deere's investment in Light solves
for this issue. CNH acquired Augmenta in 2021 and EarthOptics in 2023, which will provide
cameras and sensors. Machine learning is used in conjunction with the "sensing components"
such as computer vision — this part of the technology stack deals with identifying objects using
data pulled from the farm. Edge processing helps solve the problem of lack of rural connectivity
and is the area of the technology stack where collection, analysis, and reaction happens (on the
machine versus the cloud). Robotics is the final link in the AI chain. It physically implements the
growing decisions (i.e., the "spray" part of See & Spray). OEMs have been building this area of the
technology stack through M&A (John Deere bought Bear Flag, AGCO bought Precision Planting,
and CNH bought Raven) (Exhibit 5 to Exhibit 11).
EXHIBIT 5: AGCO's tech stack
Source: AGCO website
BERNSTEIN
100 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
EXHIBIT 6: John Deere's tech stack
Source: John Deere website
EXHIBIT 7: CNH Industrial's tech stack
Source: CNH Industrial presentation
EXHIBIT 8: John Deere's key acquisitions
Date Target Name Payment Target's Activity
2017
Blue River Technology
Inc
$305M
Developed farm robotic technology using computer vision, helping farmers
see, diagnose and execute actions like spraying herbicide on weeds
2020 Harvest Profit LLC Undisclosed Provider of farm profitability software
2021 Bear Flag Robotics Inc $250M Develops autonomous driving technology compatible with existing
machines
2022 InnerPlant Inc* $16M
Collects data from plant through its sensors and provides data on crops to
farmers
2022 Light Undisclosed
Uses computer vision approach, develops depth sensins and camera
perception for autonomous vehicles. DE will integrate Light's platform to its
autonomous tractors.
2023 Spark AI TBA
Develops human in the loop technology, which helps robots resolve edge
cases in real-time.
*reported as investment
Source: Bloomberg, Bernstein analysis
BERNSTEIN
AI IN AGRICULTURE: RISE OF THE GROWBOTS 101
EXHIBIT 9: AGCO's key acquisitions
Date Target Name Payment Target's Activity
2017
Precision Planting
LLC
$188M
Develops planting technology. Offers monitoring systems,
germination tools and singulation tools (software, hardware, and
after market production)
2019 Tecsoil Inc* Undisclosed
Designs and develops software solutions. Offers monitoring and
machine status, production traceability and logistics optimization,
climate sensing, among others.
2021 151 Research* Undisclosed
Provides custom data analytics solutions and allows customers to
gain actionable business insights.
2021 Headsight Inc* Undisclosed
Manufactures height sensors for tractor headers (for both corn and
grain harvesting). Also offers electrical adapters and automatic
steering devices.
2021 Faromatics SL Undisclosed
The company had the first ceiling-suspended robot that monitors
broiler chickens, and helps increasing animal welfare and farm
productivity. The product uses a set of sensot to measure thermal
sensation, air quality, light and sound. It uses AI to identify risks to
health, welfare and farm equipment.
2021 Apex.AI Inc $56M
Safety-certified software for mobility and driverless vehicles as it
proves a strong development infrastructure
2021 Centure Applications $22M
Greeneye Technology has a proprietary AI-enabled precision
spraying system that detects and sprays individual weeds.
2021
Appareo Systems
LLC*
Undisclosed
Specialized in the research, development, design and manufacture
of tangible technology that utilizes AI, mechatronics and innovative
electronics.
2022 JCA Industries Inc Undisclosed
Developer of autonomous software for agricultural machines,
implement controls and electronic system components.
*pending completion
Source: AGCO, Bloomberg, Bernstein analysis
EXHIBIT 10: CNH Industrial's key acquisitions
Date Target Name Payment Target's Activity
2019 AgDNA Undisclosed
Provides single point data integration, mapping and analytical tools which
paired with CNH Industrial's fleet management telematics allows farmers to
consolidate a range of agronomic data streams into a single platform to
expedite decision making
2021
Zimeno Inc (dba
Monarch)
$20M
Offers electric tractor platform of electric powertrain with autonomous
technologies
2021
Augmenta Agriculture
Technologies
$110M
Designs and develops software solutions, offers real-time camera-based
system that retrofits farm equpment and automates farm operations like
fertilizer and chemical applications
2021 Raven Industries $2B Manufactures precision ag products
2021 NX9 Software Undisclosed
Small Software house specialising in ISOBUS technology for agricultural
equipment and industry-standard communication protocol which allows
machines and implementse to talk to each other.
2022
Stout Industrial
Technology
Undisclosed
Offers an AI based ag machinery in helping growers automate labor-
intensive field work like weeding.
2023
GroundTruth Ag Inc
(dba EarthOptics)
Undisclosed
Offers sensor technology which precisely measures the health and
structure of soil through a combination of ground-based sensors, satellites,
physical soil samples, machine learning models and agronomic expertise.
2023 Hemisphere GNSS $175M
Leader in high-performance satellite positioning technology, which will help
CNH to advance their automated and autonomous solutions
Source: Bloomberg, CNH Industrial Reports, Bernstein analysis
BERNSTEIN
102 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
EXHIBIT 11: Ag tech universe is robust and full of potential acqusitions
Source: University of California
PRODUCT AUTOMATION
COMES BEFORE AUTONOMY
AI is increasingly becoming the engine behind precision ag products, as they are
increasing yields and lowering inputs and, hence, driving higher farmer profits. Ag OEMs'
precision ag products automate discrete parts of the crop growing cycle (tillage, planting,
spraying, and harvesting). For example, during the planting cycle, AGCO's Fendt Momentum
Planter senses field conditions in order to achieve the optimum depth and spacing, reducing
inputs and compaction, which improves yields up to 8bu/acre. With the Vertical Contouring
Toolbar, it compensates for height differences, which helps move around the field without
sacrificing accuracy and speed. The planter also has the Conceal product that senses and applies
nutrition in the soil three inches away from the seed. vSet helps achieve a high singulation
accuracy; it feeds the seeds into a disc, which passes over to a delivery tube and only a single
seed falls down the center, which helps farmers achieve higher yields. The planter allows farmers
an increase in productivity, enabling planting speed at 10mph. To improve the spraying process,
John Deere's See & Spray has a set of cameras that capture the information, which is analyzed
in real time, distinguishing the weed and the crop, and the equipment automatically sprays
on the weed. During harvest, John Deere's Combine Advisor automatically adjusts to changing
conditions and reduces losses. Once the performance settings are set, The Combine Advisor
makes adjustments on the fly. It is equipped with several sensors and cameras, which allow
adjustments on the rotors, for example, and reduce grain loss. (Exhibit 12).
BERNSTEIN
AI IN AGRICULTURE: RISE OF THE GROWBOTS 103
EXHIBIT 12: Key products
Crop Cycle Company Product Productivity Target Cost Reduction Yield
Fendt Momentum Planter Optimum depth and spacing Input Reduction
Reduces compaction: center section
rows improve yield up to 8 bu/acre
vDrive
Allows seed and insecticide meter
to turn off at a boundary or where it
has already been planted
Input Saving --
DeltaForce
Increases or reduced weight on
each row
-- --
SmartDepth
Removes row-to-row variability
through a calibration
--
Yield benefit of planting at the right
depth is up to 9%
FurrowForce
Adapts to the planting conditions to
remove air pockets and firm soil to
keep moisture
--
Yield increase of 7.8 bu/acre in a no-
till challenging field, 5.2 bu/acre in a
no-till mildly challenging field and a
6.7 bu/acre advantage in a
conventional till field
SeederForce
Automated downforce control
system
-- --
RowFlow
Allows to plant variable rate
prescriptions
Minimize wasted seed --
SpeedTube High-speed planting system Time reduction --
Precision Planting Ready
Row Unit*
Solution to keep an existing planter
frame
-- --
Conceal
Planter fertility attachment that
places nutrition in the soil 3 inches
away from the seed.
-- --
FurrowJet
Planter mounted device that places
starter fertilizer near the furrow.
-- Yield increase is up to 9.2 bu/acre
ExactEmerge
Increase planting speed, optimum
planting window; improve seed
placement and singulation
Lower seed usage by 2-8%. --
ExactRate
Monitors and controls liquid
fertilizer application during planting
Up to 12% input reduction --
ExactShot Robotics based fertilizer system
60% reduction of the amount of
starter fertilizer
--
AGCO VApplyHD
Pump control module to manage
and measure flow and applies
liquid fertilizer at the correct rate
-- --
CNH
Case IH Trident 5550
applicator**
Spreader combines driverless
technology with Raven's
perception system
-- --
ExactApply
Application system with variable
rate fertilization depending on the
field donditions.
2-5% reduction of herbicides
and pestices vs standard.
--
See & Spray
Detects and sprays chemicals on
weeds
80-90% chemical input savings
vs conventional methods
--
Ideal Combine
Makes automatic on-the-go-
adjustments, improving the harvest
process
--
Grain output of 89 tonnes at a loss of
1% vs competitor loss of 3%
YieldSense
Yield monitoring system that
provides easy calibration and
spatial accuracy throughout
harvest.
-- --
CNH Omnidrive**
Gives the combine operator the
ability to autonomously call a
driverless grain cart tractor directly
to the harvester to offload without a
second operator
Labor hours reduction --
DE Combine Advisor
Helps automatically maintain the
equipment and also adjust settings
of the combine
--
5% better grain quality and 5% better
yield
Several cycles DE AutoTrac Vision
Software tied into the auto
guidance system
8.5% reduced overlap (meaning
reduced inputs)
--
AGCO
DE
Planting
AGCO
Spraying/Fertilizing
DE
Harvesting
*Includes vSet, Vdrive, DeltaForce, SpeedTube, and CleanSweep; **limited commercial availability
Source: John Deere, AGCO, CNH Industrial, Bernstein analysis
BERNSTEIN
104 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
AUTONOMOUS PRODUCTION
SYSTEM
Large agricultural equipment OEMs are targeting 2030 for the first commercially available
fully autonomous production system (tillage, planting, spraying/fertilizing, and harvesting). The
strategy is to incorporate the individual precision ag products (See & Spray, Precision Planting,
Combine Advisor, etc.) onto farm equipment that autonomously moves through the fields. John
Deere has led the charge on this front, first introducing the Autonomous 8R Tractor for tillage
applications only. This product took three years of development to reach commercial viability,
but the product cycle time will accelerate as training the first autonomous platform takes the
longest time. The next part of the crop cycle to become autonomous will likely be spraying/
fertilizing in 2024, which is when AGCO will likely announce its product. The grain cart in the
harvest process will likely follow in 2025 (another AGCO target). Planting, the most critical and
complex component of the crop cycle, will likely be the last process to go autonomous. Row crops
such as corn, soybeans, wheat, and cotton are most likely to see the first autonomous production
systems (Exhibit 13).
CNH has various autonomous products under development across the crop cycle (Exhibit 14),
including the Raven OMNiDRIVE system, which allows a combine operator to autonomously
call a driverless grain tractor to the harvester to offload during tillage, and the Case IH Trident
5550 applicator, which is an autonomous spreader. Both these solutions are commercially
available, but there are others still undergoing field testing, including the Driver Assist Harvest
Solution (synchronizes combine and tractor functionality during unload on the go operations) and
Driverless Tillage Solution (allows the user to operate the machine remotely from a mobile tablet
device).
EXHIBIT 13: Autonomous product roadmap
Earliest Key
Commercial Variables
Availability to Consider OEM Commentary
Tillage 2022
Soil conditions, contour of farm and
elevation are key variables in tillage.
Value add is the use of AI to adjust
implement (otherwise a manual
process).
DE introduced the the autonomous 8R tractor, which only focuses on fall
tillage, in 2021, with commercial availability in 2022. Tech is backward
compatible until '20 model-year. All 8Rs will be autonomous-capable
starting with '23 model year. AGCO is targeting 2025 for a tillage retrofit
option
Spraying/
Fertilizing 2024 Limited data disclosed
Harvest - Grain Cart 2025
Sensing grain cart fullness, speed of
combine
AGCO targeting early 2025 for product release, making grain cart smart
enough to follow combine and know where to unload then reconnect with
the combine afterwards.
Planting TBD
Soil temperature, moisture, amount
of organic matter, seed
depth/spacing are the key variables.
No comment from OEMs about when autonomy will be available, but it will
likely incorporate existing technology such as variable rate fertilizer
deployment (ExactShot for DE, Precision Planting for AGCO)
Source: Company data, Bernstein analysis
BERNSTEIN
AI IN AGRICULTURE: RISE OF THE GROWBOTS 105
EXHIBIT 14: CNH Industrial shared this guide on the levels of autonomy in its products during its Tech Day
Presentation in 2022
Level 1 Level 2 Level 3 Level 4 Level 5
Guidance
Coordinate &
Optimize
Operator Assited
Autonomy
Supervised Autonomy Full Autonomy
Equipment controlled
by operator, but some
driving assist features
may be included
Operator remains alert,
engaged in the
environment and
control. Equipment
performs one or more
tasks simultaneously.
Equipment monitors the
environment and
performs operational
tasks. Operator is not
required to be alert, but
must be ready to take
control
Equipment performs all
operational tasks and
monitors environment
under limited
conditions. Operator
may still need to
intervene
Equipment perform all
driving and operation
functions. Operator not
required.
Till
Tillage Automation Autonomous Tillage
Plant
Planting Automation
Apply
Spraying Automation
Autonomous Spreading
(Case IH Trident)
Harvest
Harvest Automation Autonomous Grain Cart
Hay & Forage
Baler Automation
Orchard & Vineyard
Vineyard Automation
Autoguidance
Technologies (like AFS
AccuGuide AFS
AccuTurn, or PLM
IntelliCruise)
V2V Technologies
(AFS AccuSync, PLM
Intellifield)
Source: CNH Industrial Tech Day Presentation 2022, Bernstein analysis
HOW WILL THE BUSINESS
MODEL CHANGE?
AI-enabled automation and autonomy will change the OEM business model in several
ways. It will:
•Expand the ag OEM TAM. John Deere has estimated that precision ag will expand the
ag OEM TAM by as much as US$55Bn to ~US$110Bn, with a large portion of the value
capture coming from AI-enabled automation and autonomy. AI will enable growing decisions
to be made increasingly at the plant level (versus farm level historically) and increasingly
autonomously. On the autonomous front, value creation begins with the labor savings (5%
of farmer cost structure) and other input costs, but the true value lies in increasing yields —
missing the optimal planting window can cost farmers as much as 1% of yield per day (Exhibit
15).
EXHIBIT 15: Precision ag could unlock as much as US$55Bn of value for ag OEMs (2x higher market size)
$108,828
$89,884
$64,254
$53,110 $53,110 $53,110
$55,718
$36,774
$11,144
$0
$20,000
$40,000
$60,000
$80,000
$100,000
$120,000
50% 33% 10%
$ Ag OEM TAM (Millions)
Share of Total Precision Ag Value Captured by OEMs
Large Ag Equipment Market Size Today Precision Ag Value Unlock Opportunity for OEMs
Source: Company data, Bernstein analysis
BERNSTEIN
106 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
•Shift OEM business models from point of sale to recurring subscription. Farmers will
increasingly pay for precision ag technology via a subscription model. As more AI-enabled
equipment hits the field, the pool of data grows, which will yield better algorithms that can
be pushed to connected equipment in the field. Since productivity growth will be driven more
by what happens in the digital world, rather than the physical world (new hardware), it will be
easier for OEMs to charge a continuous fee for such updates. Also, the scale of investment
required to bring automation/autonomy to farming is becoming too large to monetize through
an upfront point of sale. In this type of model, revenue will be generated on a per-acre covered
basis, which more closely aligns with how farmers think about their production costs (dollars
per acre) and makes the cost more variable in nature, rather than fixed. The progress of this
shift will take time. John Deere is forecasting that only 10% of its revenues will come from
recurring sources by 2030 (we estimate the current revenue contribution is in low single
digits). Such a shift will require OEMs to rethink their KPIs — instead of "take rate," we expect
OEMs to increasingly measure the utilization rate of their technologies.
•Raise barriers to entry. The greater the access to data, the better the algorithm, the better
the decisions that AI makes during the growing process. Going forward, camera vision will
play a much bigger role in capturing data on the farm, and OEMs with the greatest installed
base/global fleet are best-positioned to capture the greatest amount of data. The end result
is likely to be greater market consolidation, favoring the biggest OEMs. On this front, John
Deere is very well-positioned, given it has a 50-60% market share, versus 25-30% for CNH
and 5-10% for AGCO.
AI-ENABLED AUTOMATION/
AUTONOMY WILL CATALYZE
MULTIPLE EXPANSION
The path to autonomy is also a path to multiple expansion. Greater automation of production
processes will result in better growing decisions (lower input usage) and higher yields, enhancing
farmer productivity and creating value. OEMs share this value with the farmers and keep
some for themselves (price realization). Higher prices translate to higher gross margins, which
translate to a structurally higher multiple. As a greater share of John Deere's value comes from
recurring revenues tied to subscriptions to AI-enabled precision ag products, margins will grow.
We estimate subscriptions command a ~75% gross margin versus 30% for equipment. As
companies such as John Deere grow their mix of subscription business (it is targeting 10% of
total sales by 2030), we see 400bps of margin expansion, which could drive John Deere's mid-
cycle multiple from ~16x today to 18x (Exhibit 16 to Exhibit 18).
BERNSTEIN
AI IN AGRICULTURE: RISE OF THE GROWBOTS 107
EXHIBIT 16: Higher productivity growth drives pricing power
Source: Bernstein analysis
EXHIBIT 17: Structurally Higher Gross Margins Drive
Structural Multiple Expansion
TRMB
DE
AGCO
HON
PLOW
PCAR
CAT
GBX
CMI CNHI
ALG WAB
OSK
ROP
TEX
WNC
ADSK
ADBE
ANSS
R² = 0.79
0x
5x
10x
15x
20x
25x
30x
35x
40x
0% 20% 40% 60% 80% 100%
FY2 P/E Multiple
Gross Margin
Source: Bloomberg, Bernstein analysis and estimates
EXHIBIT 18: John Deere's shift to a higher mix of recurring
revenues will drive a 400bps margin expansion
29%
33%
27%
28%
29%
30%
31%
32%
33%
34%
0%
10%
20%
30%
40%
50%
60%
70%
80%
90%
100%
Today 2030
Total Gross Margin
Revenue Mix
Deere: Mix Shift Creates a 400bps Gross Margin Uplift
Recurring Revenue Equipment Revenue Total Gross Margin
Source: Company data, Bernstein analysis
VALUATION METHODOLOGY See Disclosure Appendix for valuation methodology.
EXHIBIT 19: Ratings and target prices
25-May-2023 Target
Ticker Rating Currency Closing Price Price
AGCO M USD 113.60 135.00
DE M USD 354.88 367.00
CNHI.IM O EUR 12.22 16.00
CNHI O USD 13.30 17.31
SPX USD 4,151.28
EDM EUR 1,123.00
Source: Bloomberg, Bernstein analysis and estimates
RISKS See Disclosure Appendix for risks.
Chad Dillard chad.dillard@bernstein.com +1 212 969 1017
Nicholas J. Green nicholas.green@bernstein.com +44 207 170 5055
Miguel Marques miguel.marques@bernstein.com +1 212 823 3907
Ellen Lundstrom ellen.lundstrom@bernstein.com +44 207 170 0695
BERNSTEIN
108 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
EUROPEAN AIRLINES & AIR TECH:
OPPORTUNITIES FOR AI IN FREIGHT
FORWARDING
HIGHLIGHTS •The forwarding triangle: what does a winning hand look like? Freight forwarding is
a service business that consists in solving customers' logistics problems and helping their
supply chains to run more efficiently. To be a successful, profitable, growing freight forwarder
demands three core success factors. First, logistics expertise. Clients pay forwarders for
the logistics value-add they can provide, and without this the product will not be compelling.
Second, technology. Forwarding consists of vast numbers of transactions, underpinned by
enormous quantities of data. Manual processes will see too-high unit costs and threaten
profitability. Third, relationships with customers. This should help to grow volume and limit
churn.
•Digital forwarders: good technology, but less strong on other factors. The digital
forwarders' strategic bet is that technology will lead to greater efficiency, enabling them to win
business from incumbent forwarders. We agree with the first point, but not the second. Start-
ups lack the logistics expertise and relationships of the incumbents: without them, they will
likely be condemned to low-margin business and will struggle to make money. It is possible to
build or buy these capabilities, but the road to relevance is long and steep. Large forwarders
have over 1,000 locations globally: the largest digital forwarder, Flexport, has 29.
•Incumbents: odds-on to remain dominant, with an easier road ahead. Freight
forwarding, even at the more technologically capable of the incumbents, remains shockingly
inefficient. DSV disclosed in 2022 that 68% of its Air & Sea bookings come manually.
However, the large forwarders in our coverage are not standing still — they have developed
self-serve online portals and are nudging their booking mix toward more automated channels.
Kuehne+Nagel continues to work on automating manual processes under its internal project
"eTouch." Improving technology at existing forwarders will be much easier than the expertise
and relationships build-out required at digital challengers. Furthermore, already responsible
for more freight moves than anyone else on the planet, the largest forwarders should have
access to the most comprehensive datasets — and can go further, applying them to derive
insights for customers' supply chains and executing to improve service quality.
INVESTMENT IMPLICATIONS Large forwarders are best-placed to capitalize on advances in technology. We rate DSV and
Kuehne+Nagel Outperform, and Deutsche Post Market-Perform.
BERNSTEIN
EUROPEAN AIRLINES & AIR TECH: OPPORTUNITIES FOR AI IN FREIGHT FORWARDING 109
THE FORWARDING TRIANGLE:
THREE THINGS FORWARDERS
NEED TO WIN
In this chapter, we investigate the rise of digital forwarders in the freight forwarding sector.
Several companies, of which the best-known is Softbank (not covered)-backed Flexport (private),
have been seeking to use technology to disrupt incumbent profit pools. However, technology
is a necessary, but not sufficient condition for success, and replicating the logistics capabilities
of incumbents will be a long, hard, labor-intensive road. The best use of productivity-enhancing
technology will be in combination with logistics expertise and a broad network of relationships
with customers and carriers. Advantage: incumbents.
Freight forwarders are in the business of solving customers' logistics problems. To do that, and
make money in the process, they need three things. First, logistics expertise to solve supply
chain problems. Second, technology, both to integrate with third parties such as customers and
carriers, and to improve the efficiency of internal processes — ultimately making money for the
company. Third, relationships with customers, so that business volumes continue to flow. These
three constitute our forwarding triangle: only those that combine all three factors will experience
sustainable commercial success (Exhibit 1).
•Logistics expertise is at the core of customer demand. Freight forwarders exist to
solve supply chain problems for their customers: ensuring that goods flow smoothly and
finding the most effective routing for goods. There are several degrees of competence, from
the table stakes such as booking capacity on vessels, up to more complex tasks such as
consolidating and de-consolidating loads, clearing customs, moving goods that need special
handing, and physically intervening when things go wrong…which, in logistics, they do all
the time. By providing more advanced services, forwarders can increase their GP margins
and earnings. Already, the large forwarders in our coverage — DHL, Kuehne+Nagel, and DSV
— have highly advanced logistics capabilities. They are joined by many others ranging from
large players down to niche operators that are extremely strong in certain combinations of
products and geographies. The greatest challenge for digital forwarders, if they wish to
be profitable on a sustained basis, is to grow their logistics capabilities. This requires
a major expansion of office footprint, headcount, and capital investment.
•Technology competence will determine conversion margins. For a forwarder to generate
attractive margins on a sustainable basis, it needs to have an efficiently-run operation.
Flexport CEO Ryan Petersen's favorite joke — that this is not the freight forwarding business,
it's the freight email forwarding business — has more than a hint of truth. When around two-
thirds of opex below the gross profit line is staff cost, redundant manual processes create an
inexcusable drag on productivity and push down margins. Worse yet, human error will create
avoidable problems in execution of transport or invoicing, impacting service quality. We think
of technology as having two angles: internal and external. The internal part improves the
efficiency of the forwarder's own operations. The external part includes booking and tracking
interfaces for customers, and integration with carriers for direct booking onto their systems.
All have merit, and none can be overlooked. In general, digital forwarders are ahead of
traditional forwarders on technological competence.
•The human factor: relationships are still critical. Forwarding is a business built on
relationships, on trust. Customers need to have confidence that the person responsible for
moving their freight understands how important it is, and will ensure goods arrive on time and
in full as often as possible. A story at DSV's 2022 CMD was illuminating: there was a time when
they tried to consolidate their offices in Denmark, but that resulted in losing business, so they
reopened them. The human factor still matters! However, we expect this to become relatively
BERNSTEIN
110 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
less important over time compared with the other two success factors we identify above. As
millennials increasingly move into managerial positions and take decisions on procurement,
we expect more of a focus on digital interaction and self-service than in the past. Forwarders
whose sole competitive advantage is existing customer relationships will find it harder and
harder to make money. Digital forwarders lag traditional forwarders on relationships.
EXHIBIT 1: The forwarding triangle: structural winners must combine logistics expertise, technology, and
relationships
Source: Bernstein analysis
HOW DOES FREIGHT
FORWARDING WORK, AND
HOW CAN AI HELP?
Freight forwarding is historically heavily manual, paper-based, and inefficient. A standard freight
forwarding transaction consists of three main processes, each with many sub-processes:
quotation and booking, execution, and invoicing. At each stage, better digital tools and insightful
use of data, potentially augmented by the use of AI, can improve efficiency and/or customer
service (Exhibit 2).
•Quotation. The forwarding process starts with the need for goods to be moved. A shipper
contacts the freight forwarder, or several freight forwarders, to ask for a price to move
certain goods with a certain frequency from A to B, with any special handling requirements
communicated. The forwarder then contacts several carriers in the appropriate transport
mode(s) for pricing and availability. Not only are many forwarders and carriers not set up to
quote immediately, some requests may not even be responded to at all: lurking at the bottom
of an email inbox or on a post-it note until it is too late to win the business. Automated quoting
through pricing engines can both improve the shipper's experience and eliminate manual, low
value-added work at the freight forwarder.
BERNSTEIN
EUROPEAN AIRLINES & AIR TECH: OPPORTUNITIES FOR AI IN FREIGHT FORWARDING 111
•Booking. This goes hand in hand with quotation. When customers ask for goods to be moved,
forwarders will often contact several carriers in the appropriate transport mode(s) for pricing
and availability. When the best option is determined and the quote accepted by the customer,
the forwarder books with the carrier(s) for the goods to be moved. In the most common form,
this involves emails and phone calls between carrier and forwarder: easily thousands of hours
per year between businesses of scale. A more automated process, where forwarders can
book directly onto carrier systems, saves time and reduces errors. Some of the more advanced
forwarders and asset-owning transport companies are already doing this.
•Execution: physical movement of goods. Items are picked up, loaded, moved, unloaded,
and dropped off. If borders are crossed, customs declarations must be filled and processed.
For smaller shipments, these will often be combined with other freight to minimize empty
miles in a less-than-truckload (LTL) or less-than-container load (LCL) shipment. For the legacy
forwarder, the status of the shipment is only known on request, such as when the agent calls
a driver. This puts shippers and forwarders on the back foot to begin with, and in firefighting
mode: "the driver didn't show," rather than, say, observing a dashboard and seeing a truck
stuck in traffic, or a container not loaded on the right vessel. Underpinning all this you have
manual handoffs, paper forms, text messages, emails, and phone calls: unstructured data, not
visible to all stakeholders that need it.
•Execution: information and documentation flows. The international movement of goods
requires significant documentation. What is being moved? What industry code is that? Where
is it going? How much is it worth? Who is insuring it? How heavy is it? What volume does it take
up? And so on. Carriers, forwarders, terminals, customs agents, shippers and consignees all
either have or require information for the goods to flow.
•Invoicing. Getting the bills right is surprisingly difficult. How heavy was the shipment really?
Did everything that was supposed to ship, actually ship? Were there any damages or service
failures and who pays for those? In a legacy forwarder, the office staff do their best with the
information they have been given, but bills often have to be revised: an overcharged customer
will call out errors, an undercharged one may not always do so in this highly transactional
industry. Automated invoicing can help reduce errors.
•And then we go again. One shipment done, on to the next. What worked well, or didn't?
How did the various actors perform? If it is possible to learn anything from the last shipment
to improve the performance of the next one, that should be taken into account. If we are
relying on individual people bringing up pain points, things can be missed. With better data,
automatically captured, the ability to identify areas for improvement will be enhanced. AI
can potentially bring all of this together, finding patterns and deriving insights for logistics
professionals to act upon.
BERNSTEIN
112 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
EXHIBIT 2: How digital capabilities can improve the forwarding process
Source: DSV capital markets day presentation
TECHNOLOGY CAN BRING
IMMENSE IMPROVEMENTS TO
FREIGHT FORWARDING
Technology can bring improvements to the forwarding process, enabling better service provision
at a lower cost. Forwarders have three primary pathways to use technology: with customers,
within the forwarding organization, and with other stakeholders. Digitizing customer interactions
can streamline the quotation and booking process, simplify form filling, and improve visibility
and invoice quality. Digitizing the forwarding organization can automate routine tasks and
improve employee productivity while reducing human error. Digitizing interactions with other
stakeholders can smoothe the flow of goods and improve the overall performance of the actual
execution of transport activities (Exhibit 3).
•Digital with the customers: technology is a route to better service. It says much about
the freight forwarding business that today's most well-known disruptor, Flexport, was started
in the mid-2010s by an importer who was frustrated by the slow, inefficient service he
received from his forwarder. Customers often need to wait days for a quote from a forwarder,
who might not even get back to them at all. To track a shipment, a customer may need to call
the forwarder, who then needs to call the carrier to find out — and may not even know. Real-
time tracking can be all but impossible, meaning developing problems cannot be seen and
interventions made until they are already impacting the next link in the supply chain. Filling
forms is time-consuming and manual. Bills can be wrong and take phone calls to sort out.
With better technology, customers can enjoy a far superior freight forwarding service: rapid
quotation, automated or even real-time tracking, digitized form-filling, automated invoicing
with fewer errors…or all of these. This improves the overall service a customer can receive.
•Digital within the organization: technology can bring significant operating
efficiencies. As manual as interactions are with customers, internal processes are often just
as manual, inefficient, unstructured, and error-prone. Internal emails, phone calls, and paper-
based handoffs are common, increasing the amount of work hours spent on a shipment.
A forwarder that uses a robot to perform routine, low-value adding tasks should be more
profitable than one where this is done by a human. As software improves and develops, we
would expect the amount that can be done by automated tools to increase, further increasing
the value-add of the average labor hour in the organization. Kuehne+Nagel is best-in-class
here, with its eTouch project focusing on precisely this area. Better data capture can also
create a virtuous circle. For example, humans can validate the output of a tool, with the results
used to improve the future performance of the tool. Alternatively, using big data to optimize
forecasting and freight routing can bring stronger logistics performance.
BERNSTEIN
EUROPEAN AIRLINES & AIR TECH: OPPORTUNITIES FOR AI IN FREIGHT FORWARDING 113
•Digital with other stakeholders: a win-win. The very nature of supply chain management
is coordinating activities outside the boundary of the firm. Seen through this lens, the benefit
of technology can extend beyond the buyer and seller of freight forwarding services to other
stakeholders. With carriers such as airlines and shipping lines, technology can be used to
streamline the booking process, even going so far as to enable forwarders to book directly
onto carriers' systems. During transport, carriers that are integrated digitally can coordinate
much more easily with other supply chain participants to make handovers run more smoothly:
say, the dock knows when the truck is coming and can prepare for it. Customs declarations
can be filled in and filed manually, but automated filling can speed up the process, minimize
errors, and make for smoother Customs clearance.
EXHIBIT 3: Three types of improvement: digital inside the forwarder, digital with customers, and digital with
carriers and other stakeholders
Source: Bernstein analysis
DIGITAL FORWARDERS ARE
AIMING TO DISRUPT EXISTING
BUSINESS MODELS, BUT ARE
STRUGGLING
Digital forwarders have been growing rapidly over the last few years. These companies aim
to offer a freight forwarding product by relying on software, rather than humans, to do much
of the underlying process, and are starting to take real market share. Flexport is already the
sixth-largest forwarder on the Transpacific and its last funding round, in early 2022, valued the
company at US$8Bn. We expect this growth to continue, supported by a software-first culture
and likely a lower price point. Whether these companies can be sustainably profitable on the
current business model is still a matter for debate, and they appear to be struggling (Exhibit 4).
•The latest in a long line of industries facing digital disruption. Over the past decade,
technology-led companies have disrupted the businesses and business models of incumbent
players: think Uber versus the minicab industry or Netflix versus Blockbuster. Digital
forwarders are now trying to do the same to freight forwarding, using technology to improve a
heavily manual, slow-to-change industry. Chief among these is Flexport, founded in 2014. The
pitch is simple: there is so much data in the industry that goes unused, and so many manual
processes that drive up costs for customers and forwarders alike. A digital-first approach has
its advantages.
BERNSTEIN
114 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
•Forwarding cannot be disintermediated: disruption here means more efficient
operations. Forwarding is a logistically complex operation. It is not like being a travel
agent, where low-cost carriers disrupted business models decades ago by selling direct to
customers. Managing Customs, consolidation, deconsolidation, and firefighting on shipments
across 200 countries in real time is simply not an option for many companies to do in-house.
The game digital forwarders are playing is one of radical process change: the core demand
requirements by the customer remain the same.
•Digital forwarders are growing like wildfire. Customers and volumes continue to grow
rapidly for digital forwarders. While still relative minnows next to the incumbents in this highly
fragmented industry, they are attracting the business and valuations to make themselves felt.
Flexport is now the #6 forwarder on the Transpacific trade lane, and at its last valuation round
in 2021, achieved US$8Bn valuation — around 25% of that of a DSV or a Kuehne+Nagel at
the time.
•One core advantage over many forwarders is culture. As an industry, forwarding is
resistant to change. Processes and established ways of working are deeply embedded in
organizations. One particularly salient example is the aborted rollout of an SAP Transport
Management System at DHL in the mid-2010s: the software worked fine, but the culture
rejected it, and the division took a €300Mn write-off. At a new company, there is no
established culture to resist a piece of software doing a job that a human might otherwise
carry out, no legacy processes to eliminate. In our view, this confers an advantage on digital
forwarders from a back-end process perspective.
•What's the pitch: better or cheaper? Customers basically choose a supplier because the
offer is the most attractive. In forwarding, that means a cheaper quote or higher service quality.
Which are the digital forwarders going to be? Technology should help them to at least be
cheaper, but we have heard suggestions that part of the rapid growth is due to subsidizing
freight. As these companies are not public, we cannot comment on whether this is true. If it is?
Fine! If the goal is scale, credibility, and market share, and access to new capital at ever-higher
valuations is predicated on revenue growth, then why not? But eventually the companies will
need to start putting up prices, and from that point on, the competitive landscape will look
quite different.
•For a business scaling as rapidly as Flexport, 20% headcount reductions suggest
demand problems. Digital forwarders have been growing at ferocious rates in recent years,
breaking into the top ranks by volumes on several lanes. Valuations suggest investors expect
this to keep happening. Against that backdrop, a 20% reduction in headcount at Flexport in
Q12023 suggests deeper problems: we are concerned the product is struggling to sustain
demand at adequate levels.
THIS IS A STILL A LOGISTICS
BUSINESS, AND THE
INCUMBENTS ARE FAR
AHEAD
One of the traps for digital forwarders is too much emphasis on "digital" and not enough on
"forwarder." Some companies choose to be software providers to the logistics industry, and do a
great job of it. However, those that go down the path of freight forwarding need to build logistics
capabilities in-house. That means a deep understanding of how to plan, manage, and route
freight. That means a physical presence and boots on the ground at major logistics locations
around the world. That means — if you want to make money — doing the hard work of logistics
yourself rather than outsourcing it. This is something where industry incumbents — even the
BERNSTEIN
EUROPEAN AIRLINES & AIR TECH: OPPORTUNITIES FOR AI IN FREIGHT FORWARDING 115
technology dinosaurs — are far ahead of the new breed (Exhibit 5).
•You hire a forwarder to solve your logistics problems. Freight forwarders with strong
logistics capabilities can be incredible with the right technology, but even the best technology
will not save a forwarder that lacks the ability to get shipments from A to B. The value a
forwarder brings to its customers is in finding better routings for freight, consolidating loads,
booking capacity, consolidating loads, and seeing shipments through to execution. Without
that, there is no point in the business existing.
•Wide global networks are a critical advantage in executing shipments and intervening
in problems. Things go wrong in logistics all the time. Trucks get stuck in traffic, break down,
or don't turn up, goods are offloaded at the wrong port, or not loaded onto the right vessel,
items get misplaced. It happens. But when things go wrong, how can a forwarder ensure
service recovery? Frequently, emails are not enough — there is value in having your rep go to
the person in charge at the terminal, point at the box, and ask them to get it to where it needs to
be. Without an on-the-ground presence at one forwarder, customers will often choose to use
another that can take responsibility for that particular goods flow. DSV has 1,500+ offices and
logistics facilities globally; we would expect other large forwarders to be similar. Flexport has
29 locations. An extensive footprint takes a long time to build out, and digital forwarders, less
than a decade old, are far behind the curve. Scaling a forwarding business is not like scaling
software: it takes the hard yards of renting space, hiring people, and building relationships.
•To make real money, you need to do hard things. There is a huge difference in profitability
between individual shipments, and this is in part driven by shipment complexity. A full
truckload move from A to B might have a gross margin of, say, 9%. A groupage transaction,
in which loads from several customers are consolidated onto a single truck and then
deconsolidated at a later point, might have triple the gross margin. Similarly, in ocean, less-
than-container load shipments generate far higher unit gross profits than full container load
shipments. But who captures the profit? Ask a large, competent forwarder and the answer
is "I do." Many digital forwarders do not have the same logistical capability in-house, and
outsource this to a third party, who then captures the value from providing the service. The
digital forwarder is able to grow its customer book and number of consignments, sure — but
without building out the logistics service is missing out on much of the value of the transaction.
BERNSTEIN
116 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
EXHIBIT 4: Large global forwarders have networks 50x
more extensive than the largest digital disruptor
Total locations, including offices, terminals, and warehouses
1,500
1,294
29
DSV Kuehne+Nagel Flexport
Source: Company reports, Bernstein analysis
EXHIBIT 5: Less-than-container load shipments are more
logistically complex and generate higher gross profit
Gross profit per TEU, DSV, 2021
10x
Full container load Less than container load
Source: DSV presentation, Bernstein analysis
BEST-RUN INCUMBENTS ARE
INVESTING IN TECHNOLOGY
FASTER THAN DIGITAL
FORWARDERS CAN GROW
THEIR LOGISTICS FOOTPRINT
Incumbent freight forwarders are not standing still on the technology front. The three in our
coverage — the world's three largest firms — are investing seriously in technology across both
front-end portals to improve customer service and back-end systems to improve efficiency. This
is a priority for Kuehne+Nagel, DSV, and DHL. In the race of digital forwarders building logistics
capabilities and traditional forwarders improving digital capabilities, the incumbents should be
favorites.
Kuehne+Nagel
•Back end: In-house-developed transport management system gives K+N the lowest
marginal IT cost around…Kuehne+Nagel has made the boldest investment in technology.
The ubiquitous transport management system CargoWise One has 23 of the top 25 freight
forwarders as its customers — but it doesn't have one of the top three. Kuehne+Nagel carved
its own path, building its own transport management systems: AirLOG for Air and SeaLOG for
Sea. This took years of focused effort and was a high-risk strategy, but with this system rolled
out, should have a higher marginal conversion rate on each incremental shipment.
•…while eTouch creates a virtuous circle of productivity and market share gains. A
key longstanding pillar of Kuehne+Nagel's strategy is cost reduction under the umbrella
of eTouch. The immediate impact of eTouch is to automate routine steps in the forwarding
process. However, the upside from reduced manual intervention is not targeted at lowering
headcount. Rather, unit cost savings can be realized by allowing existing employees to
process many more volumes — which leads to falling unit opex. As manual intervention is
removed from the process, margins will rise, and Kuehne+Nagel may choose to use its
stronger cost position to gain more volume too. Expect more volumes, further gains from
automation, and a virtuous circle where scale builds scale.
BERNSTEIN
EUROPEAN AIRLINES & AIR TECH: OPPORTUNITIES FOR AI IN FREIGHT FORWARDING 117
•Front end: myKN. Kuehne+Nagel's customers can use myKN as an online portal for instant
quotation, plus booking and tracking. The tool aims to improve customer service and eliminate
the need for disparate systems and manual intervention such as phone calls to quote, book,
and track every shipment.
DSV
•Back end: Higher data quality, more automation. DSV has a well-architected IT stack,
with a small number of data systems and one transport system per mode. Simplification
is the watchword, as with much else it does. Current initiatives are trying to improve data
quality in the company. For example, 68% of air and sea bookings are still made manually;
the aim is to move as much onto myDSV and APIs as possible. Road is somewhat better, with
76% of bookings coming through Electronic Data Interchange (EDI), but again the aim is to
move this to more advanced tools with higher data quality. Beyond booking, other tools have
been developed to assist in the flow of freight. For example, AI to automatically fill Customs
documents, digital twins of cargo to assist with load consolidation, and drones to do automatic
stock counts (Exhibit 6).
•Front end: myDSV. DSV customers can manage their shipments through the online portal
myDSV. This gathers all of a customer's dealings with DSV in one tool, eliminating the need to
use separate phone calls, emails, and systems. It allows booking, an overview of shipments,
tracking, and report generation.
DHL
•Back end: CargoWise One rollout closes the gap to peers. DHL was running on a legacy
transport management system for a long time, after an attempt to replace the in-house system
with an SAP product was aborted in 2015. Following a change of CEO, CargoWise One was
rolled out to Ocean and Air, and completed by 2021. This is greatly improving the ability of DHL
to manage freight flows and the forwarding business, including automating certain routine
tasks, integrating with vendors and carriers, and having a view on the true profitability of each
shipment. Further functionality in the system will be implemented over time, with additional
productivity gains.
•Front end: myDHLi. On the front end, the company rolled out myDHLi in 2020, a portal for
customers to manage quotation and booking, tracking, documentation, analytics, and reports.
The aim is better customer service — and therefore at the end of the day, happier customers
that stick with DGFF (Exhibit 7).
BERNSTEIN
118 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
EXHIBIT 6: DSV is pursuing more automated ways of booking, and the journey is not even half-done yet
Booking share by tool, DSV, 2021
Source: DSV presentation
EXHIBIT 7: DHL Global Forwarding's myDHLi: a digital, multi-channel portal for customers to manage freight flows
Source: Company website
BERNSTEIN
EUROPEAN AIRLINES & AIR TECH: OPPORTUNITIES FOR AI IN FREIGHT FORWARDING 119
VALUATION METHODOLOGY See Disclosure Appendix for valuation methodology.
EXHIBIT 8: Ratings and target prices
25-May-2023 Target
Ticker Rating Currency Closing Price Price
DPW.GR M EUR 41.29 43.50
DSV.DC O DKK 1,313.00 1,500.00
KNIN.SW O CHF 255.40 300.00
IDS.LN M GBp 196.85 300.00
EDM 1,123.01
Source: Bloomberg, Bernstein analysis and estimates
RISKS See Disclosure Appendix for risks.
Alex Irving alexander.irving@bernstein.com +44 207 170 0539
BERNSTEIN
120 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
INDIA INDUSTRIALS & INFRA:
OPPORTUNITIES FOR AI IN THE POWER
SECTOR
HIGHLIGHTS •Smart grid (the most real consequence). Before AI solutions, we need data and control. The
future grid will have two-way communication and "prosumers." Investments in grid digitization
are expected to be focused on automation control, communication, and distributed energy
management (Exhibit 6). Smart meters are critical for this story as well; India plans for 250
million smart meters by 2025, of which only 5.6 million are done.
•Weather forecasting. The ongoing use case of AI is in weather forecasting, a critical part of
renewable power scheduling. Solutions from IBM and research from DeepMind claim to see
15-30% improvement in weather predictability and up to 20% impact in effective realization
using AI based prediction tools (Exhibit 9).
•Demand-supply-tariff modeling and virtual power plants. Many traders in Europe are
leveraging AI/smart algorithms to predict demand-supply-tariff better and act as virtual
power plants (managing many small renewable power plants, storage, and consumers as one
single power plant). Statkraft (private) claims to have 10GW under management, and some
claim 85% algo trades in intra-day. This is a need in India as well (trading margin cap is a
barrier), and entities such as PTC (not covered) should benefit.
INVESTMENT IMPLICATIONS We maintain Outperform rating on ReNew — it continues to be our top pick in our coverage. Its
investments in AI/digitization will likely give it an edge over its competitors in the long run. L&T
(Outperform) is a big beneficiary of the transition toward a smart grid, which we think has a high
certainty of being executed irrespective of how big a role AI plays. On Indian Energy Exchange
(Underperform), if its digital journey is executed in the right manner, it could make this a key
differentiator against competitors, especially in case "market coupling" becomes a reality. We
maintain our Underperform rating on Adani Green.
WE KNOW UTILITIES/ENERGY
IS NOT WHAT COMES TO MIND
WHEN YOU THINK OF AI
Well, what did you expect? We don't think we needed charts to reach this conclusion, but in
Exhibit 1 and Exhibit 2 we can see utilities/energy ranks on the lower end of sectors when it
comes to the percentage of jobs related to AI and quantum of private investment in AI. We were
actually a bit surprised to see even 1.3% of job openings in utilities being in AI (can only happen
in the US!).
BERNSTEIN
INDIA INDUSTRIALS & INFRA: OPPORTUNITIES FOR AI IN THE POWER SECTOR 121
EXHIBIT 1: Utilities/energy sector ranks on the lower end
compared with other sectors in terms of percentage of
jobs related to AI
0.6%
0.7%
0.9%
1.0%
1.2%
1.3%
1.3%
1.3%
1.4%
1.5%
1.6%
3.3%
3.3%
4.1%
5.3%
0% 2% 4% 6%
Waste Mgmt. and Admistrative Support
Transport and Warehousing
Real Estate and Rental and Leasing
Wholesale Trade
Mining, Quarrying, and Oil and Gas Extraction
Utilities
Retail Trade
Public Administration
Mgmt. of Companies and Enterprises
Educational Services
Agriculture, Forestry, Fishing and Hunting
Manufacturing
Finance and Insurance
Professional, Scientific and Technical Services
Information
AI Job Postings (% of All Job Postings) in US by sector
2022 2021
Source: Artificial Intelligence Index Report 2023, Stanford, Lightcast, Bernstein
analysis
EXHIBIT 2: Private investments in AI by focus area
Source: Artificial Intelligence Index Report 2023, Stanford, NetBase Quid,
Bernstein analysis
Our research below threw up some areas where AI can be of relevance in power/utilities.
Moreover, it also shows how renewable companies can actually differentiate on the basis of these
skills in the future.
ASSESSING POTENTIAL
APPLICATION AREAS OF AI IN
POWER
To assess the possibility of AI applications, we used the framework shown in Exhibit 3, analyzing
what are the key decisions needed to be made at each part of the power value chain. To enable
the use of AI in the power sector, we think the first step is to ensure we have the right equipment
to drive two-way communication, automation, and two-way flow of power — only then does the
role of AI algorithms to optimize decisions becomes relevant. The most immediate impact we see
of AI (and required equipment) in the power sector is in the below categories:
(a) Equipment: Grid digitization;
(b) Equipment: Smart meters;
(c) Software application: Weather forecasting;
(d) Software application: Efficiency improvement and predictive maintenance of thermal power
plants;
(e) Software application: Virtual power plants (including supply-demand-tariff forecasting); and
(f) Software application: Distributed energy management systems.
BERNSTEIN
122 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
EXHIBIT 3: AI in power sector: Our framework based on key decisions to assess where could AI play a role in the
power sector
Area Key Decisions Software Equipment
Renewable Energy
Generation
How much to sell at what time? Weather forecasting
a) How much to sell? a) Efficiency improvement
b) How to operate the plant?
b) Capacity modelling/
Digital Twin
c) What to upgrade? c) Predictive maintenance
a) Sell or Buy or Store? a) Demand-Supply
b) How to sell? b) Tariff estimation
a) How to optimize load in the system? a) Network planning
b) Distributed Energy
Management
a) When to run heavy loads?
b) When to send back to grid?
Conventional
Generation
Power Trader/
Power sales
Customer
a) Smart-meters
b) Sensors
c) Automation and control
d) Communication system
e) Batteries
f) EV Smart-charging
g) Storage other - Pumped
Transmission/
Distribution Network
Source: Bernstein analysis
(A) EQUIPMENT: FUTURE GRID Investments in grid digitization: The future grid (smart grid) will have two-way flow of power
and information, from the one-directional flow of power today and limited communication
(Exhibit 4).
EXHIBIT 4: Future grid: allows two-way communication and power flow
Source: CLP, Bernstein analysis
We think the foundation of AI is data, for which we need systems capable of communicating with
each other, monitoring performance, and implementing automated decisions. Hence, the first
step for AI to become real in the power sector is grid digitization/smart grid implementation.
Between 2022 and 2050, capex expenditure on grid digitization is expected to be US$5.1Tn
BERNSTEIN
INDIA INDUSTRIALS & INFRA: OPPORTUNITIES FOR AI IN THE POWER SECTOR 123
globally and US$0.6Tn in India (Exhibit 5). Grid automation, advanced distribution management
systems (ADMS), and distributed energy resources management systems (DERMS) will likely be
the biggest contributors to this capex. A lot of this capex is also driven by a need for a flexible
grid for renewable integration — to monitor, control, and optimize renewable and supporting
generation equipment. Digitization is also required to prevent grid failure by preempting
equipment failure (Exhibit 7).
EXHIBIT 5: Between 2022 and 2050, capex expenditure on grid digitization is expected to be US$5.1Tn globally and
US$0.6Tn in India
0
2
4
6
2022 2030 2040 2050
$ trillions
Cumulative Capex for power grid digitalization
China US Europe India Middle East and North Africa Rest of World
1.3
0.8
0.7
0.6
0.4
1.4
2022-2030
Total = $ 0.65 Trillion,
India = $ 0.03 Trillion
2031-2040
Total = $ 1.84 Trillion,
India = $ 0.15 Trillion
2041-2050
Total = $ 2.65 Trillion,
India = $ 0.38 Trillion
Source: Bloomberg New Energy Finance (BNEF), Bernstein analysis
EXHIBIT 6: Grid automation, ADMS, and DERMS will
likely be the biggest contributors to this capex
34
123
228
292
0
50
100
150
200
250
300
350
2022 2030 2040 2050
Global annual capex on grid
digitalization
($ Bn)
Analytics
DER flexibility
Communications
Monitoring
Automation &
control
Source: BNEF, Bernstein analysis
EXHIBIT 7: Expected grid digitization spend per km during
2022-50 by different countries
2.7
2.2
2.1
2.1
1.9
1.8
1.6
1.2
0 1 2 3
Japan
Germany
UK
China
France
Australia
US
India
Average (2022-50) annual digital
spend per grid length ($ thousand
per km)
Source: BNEF, Bernstein analysis
(B) EQUIPMENT: SMART
METERS
India has big plans for a smart grid, and especially for smart meters as an enabler to digitize the
grid. While the primary objective of this initiative is to reduce AT&C losses in the system (e.g.,
prepaid smart meters), these meters are critical for two-way communication between the buyer
and the seller. These additional data points from each customer would be critical to develop any
BERNSTEIN
124 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
AI-based solutions for demand-supply forecasting, tariff prediction, network planning, etc.
India has already installed 5.6 million smart meters, with L&T having the biggest mandate,
followed by Genus (Exhibit 8). But the ambition is much bigger — to reach 250 million smart
meters by 2025. We could see another 5.6 million that are sanctioned but to be completed, and
another 94 million in the tendering/proposed/partially awarded stage in the India Smart Grid
Forum database.
EXHIBIT 8: Smart-meter deployment companies
# of Smart Meters Sanctioned Installed Balance
L&T 5,185,776 1,987,790 3,197,986
EDF, Genus 2,350,000 1,299,033 1,050,967
Genus 666,570 597,190 69,380
Anvil Cables 511,766 271,620 240,146
Intellismart 620,100 203,020 417,080
BOSCH 368,204 199,900 168,304
Landis Gyr 195,000 195,000 0
Schneider Electric 151,740 151,015 725
Purbanchal Ent. 134,000 134,000 0
Techno Electric 377,722 124,666 253,056
MP Smart Grid 350,000 124,477 225,523
PSPCL, HPL 96,000 88,107 7,893
Others 268,861 225,540 43,321
Total 11,275,739 5,601,358 5,674,381
Note: From the companies in the list, we only cover L&T.
Source: India Smart Grid Forum, Bernstein analysis
(C) SOFTWARE APPLICATION:
WEATHER FORECASTING
The most real and already ongoing use of AI is in weather forecasting for the power sector.
Especially with the rise in renewable energy as a share of grid power, the need for accurate solar
wind forecasting tools is critical to ensure grid stability. India is setting up Renewable Energy
Management Centers to use AI to forecast weather. PGCIL (not covered) is setting up one for
better load dispatch management (see here).
Environmental Intelligence Suite, IBM's AI- and advanced analytics-based weather prediction
offering, can achieve a 15-20% improvement in wind and solar forecasting accuracy. India case
study: IBM implemented its weather forecasting tool in the states of Uttar Pradesh and Bihar in
2020. While we don't think it made use of the full suite of IBM's AI modeling, the impact of quality
weather forecasting was visible in the results. When the process was manual, error margins were
10-15%. With IBM’s weather data as a key input, they achieved an initial error rate of only 5-6%.
DeepMind is an AI lab acquired by Google for ~£400Mn in 2014. One of the areas they are
working on is improving predictability of wind speeds using AI and ML to optimize power sale
from wind plants. They claim machine learning is able to boost the value of their wind energy
output by ~20%, compared with the baseline scenario. DeepMind claims to predict output 36
hours in advance (Exhibit 10 and Exhibit 11) (source: Article with Google-DeepMind AI claims,
DeepMind article).
BERNSTEIN
INDIA INDUSTRIALS & INFRA: OPPORTUNITIES FOR AI IN THE POWER SECTOR 125
EXHIBIT 9: IBM's weather prediction offering
Source: IBM, Bernstein analysis
EXHIBIT 10: DeepMind claims to predict wind output
36 hours in advance with good accuracy — resulting in
optimal utilization of wind resources
Source: DeepMind, Bernstein analysis
EXHIBIT 11: Potential value add to a wind plant by
accurate forecasting (DeepMind study)
Source: DeepMind, Bernstein analysis
(D) SOFTWARE APPLICATION:
EFFICIENCY IMPROVEMENT
AND PREDICTIVE
MAINTENANCE OF THERMAL
POWER PLANTS
An interesting case study is the work done by McKinsey (private) for a fleet of coal-fired power
plants in the US, where they leveraged 400+ AI models (multi-layer neural network model)
to save US$60Mn/year for a fleet of 67 power plants (full roadmap to ~US$250Mn EBITDA
improvement) (Exhibit 12 and Exhibit 13). The role of the AI algorithm was to essentially process
current and historical data from multiple sensors, ambient conditions (temperature), network
conditions, plant conditions, etc., and optimize the following decisions (and get better at it with
time): (a) Capacity forecasting: optimal capacity available to supply power for the next hour/
day, etc.; (b) Plant reliability: predictive maintenance; (c) Efficiency enhancer: AI models to
suggest optimal operating parameters for the plant to ensure maximum plant efficiency; and
(d) Performance management: tools to optimize day-to-day plant operations, e.g., start-up
optimization.
BERNSTEIN
126 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
EXHIBIT 12: AI analytical models can be leveraged to
improve plant maintenance decisions (predictive or
value-based)
Source: McKinsey AI Power Play report, Bernstein analysis
EXHIBIT 13: Role of AI models becomes even more
apparent for day-to-day decisions to operate the plant —
hard for human beings process so many data points
Source: McKinsey AI Power Play report, Bernstein analysis
(E) SOFTWARE APPLICATION:
VIRTUAL POWER
PLANTS (INCLUDING
SUPPLY-DEMAND-TARIFF
FORECASTING)
Demand-supply forecasting and the evolving role of a trader
Demand-supply-tariff forecasting — Toshiba's example (Exhibit 14). Toshiba's Virtual Power
Plant's demand forecasting uses AI to optimally combine weather forecast and ensemble
learnings from the relation between weather and demand to predict future grid demand values.
For electricity market price forecasting, Toshiba has developed a technology to make accurate
predictions based on the relation between its weather forecasting system and past exchange
actual data, using an analog ensemble. For supply, as discussed above, it is a combination of
weather forecasts and equipment response generation in that weather, which is used to forecast
power supply from renewables. With adequate automation and controls, it can remotely schedule
and control storage batteries based on their predictions of demand-supply-tariff.
Virtual power plants. With the evolution of the network model, access to more data, and
emerging AI algorithms to make more accurate predictions of demand, supply, and potential
power tariffs, a key new role that emerges is of an aggregator or a virtual power plant (VPP). A VPP
is essentially a role a trader could play — aggregating power generators/storage solutions from
multiple sources (wind, solar, bio-energy, hydro, and batteries) spread across geographies by
connecting them virtually making them act like one large-scale generator. One can even connect
demand centers to this. VPPs play the following roles:
•Optimize power purchase-sale across these sources to ensure each asset makes higher
return than it would on a standalone basis (e.g., when to sell on the grid, when to store, and
when to ask the demand center to reduce load to offset shortfall in injection).
•Minimize deviations penalties: Renewable generation is at the weather's mercy. Hence, it is
very difficult to predict the exact supply schedule of a small power plant, which can result in
substantial penalties on renewable generators for deviating from their committed schedule
(in India, it is applied via the deviation settlement mechanism (DSM)). By aggregating supply
across several plants and leveraging AI tools for better weather/demand-supply predictions,
a VPP can minimize deviation penalties for power plants.
BERNSTEIN
INDIA INDUSTRIALS & INFRA: OPPORTUNITIES FOR AI IN THE POWER SECTOR 127
•Tools: Computer algorithms for weather forecasting, power demand-supply forecasting, and
tariff forecasting. An image from Statkraft's website is shown below (Exhibit 15) — it has more
than 10GW of installed capacity from over 1,000 power generators as part of its VPPs.
EXHIBIT 14: Toshiba VPP offering
Source: Toshiba website, Bernstein analysis
EXHIBIT 15: A VPP allows multiple small generators/load
centers to act as one single power plant
Source: Statkraft, Bernstein analysis
(F) SOFTWARE APPLICATION:
DERMS
Other than utility-scale renewable penetration, a key expected trend is higher penetration of
distributed generation (rooftop solar, batteries at home, etc.). The majority of these distributed
energy systems are on grid, and managing them is not an easy task, especially as they become
sizeable contributors to the energy balance of the grid. Hence, most utilities need to invest in
DERMS to remove intermittency from these resources and convert them into a dispatchable grid
resource that supports reliability and efficiency goals (Exhibit 16).
AI-based software can enable optimal utilization of distributed energy resources in a grid. They
can even allow usage of storage solutions across homes/offices/cars/industries to support
grid balancing and in improving their commercial viability/returns. They can be very helpful for
network stability and planning for future capex deployment (Exhibit 17).
BERNSTEIN
128 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
EXHIBIT 16: Illustration from Oracle: dispatch modeling
Source: Oracle, Bernstein analysis
EXHIBIT 17: Illustration from Oracle: dispatch scheduling
and management
Source: Oracle, Bernstein analysis
COMPANIES ReNew: the developer
ReNew has used digital analytics and AI over the past few years to improve the efficiency of its
energy assets and has set targets to improve efficiency by 1.5% to 2% over its FY22 values by
FY25 (Exhibit 18 and Exhibit 19). It has started ReNew Digital (ReD), which has two broad buckets
of using digital and AI to differentiate these:
•Energy management services: ReNew acquired Climate Connect, a digital analytics,
software, AI, and ML company specializing in the Indian power market in June 2020 for its
energy management services. Its custom product TropoSkope, which is used for weather
forecasting, has helped ReNew in efficient forecasting of wind speeds. A TropoSkope module
provides weather parameters at a very high resolution and accuracy right from a 15-minute
baseline up to a year-ahead.
•Remote asset monitoring and maintenance: New analytical models and remote
monitoring tools (e.g., predictive maintenance) have increased the energy efficiency of
ReNew's plants by detecting malfunctioning of assets faster, hence reducing downtime and
unplanned maintenance (effectively reducing the need for manual inspection and hence O&M
costs).
BERNSTEIN
INDIA INDUSTRIALS & INFRA: OPPORTUNITIES FOR AI IN THE POWER SECTOR 129
EXHIBIT 18: ReNew's operations control room
Source: Company website, Bernstein analysis
EXHIBIT 19: Digital and AI initiatives being taken by ReNew will help it differentiate against peers
Source: Company website, Bernstein analysis
Indian Energy Exchange: the enabler
While the company might not be leveraging AI to a large extent at present, the digitization of
its interface and automation of its processes are crucial for its customers (generators, traders,
DISCOMs, large consumers) to build AI-based decision tools to automate tasks such as bidding
in the real-time market. Key initiatives in this direction include:
•Automated bidding through application programming interface (API) for the real-time market
(Exhibit 20);
BERNSTEIN
130 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
•Web-based platform services for digital registration, data insights and analytics and hence
improving the efficiency in various aspects of the bidding process such as registration,
bidding, physical delivery, financial settlement, etc.;
•System integration with clearing banks for faster financial transactions;
•Adopting robotic process automation (RPA) in the near future for eliminating human
dependency in market operation processes; and
•Plans to introduce web-based bidding to provide anytime, anywhere, easy, and secure bidding
experience to its market participants.
We believe these tools can help differentiate Indian Energy Exchange from its peers even if
market coupling is initiated in the power exchanges in the future.
EXHIBIT 20: Digital developments at Indian Energy Exchange can be a key enabler for AI use in the Indian power
sector for better decision-making
Source: Company website, Bernstein analysis
L&T: the implementer
L&T, through its subsidiary L&T Smart World & Communication, offers smart grid solutions
including smart meters, advanced communication networks, digital sensors, predictive analytics,
and substation monitoring for power utilities (Exhibit 21). It has partnered with EESL (private)
to spearhead implementation of an advanced metering infrastructure (AMI) system for more
than five million smart electricity meters with a GPRS-based communication module. It has
also mentioned in its FY26 strategic plan "Lakshya 2026" its intention to invest in battery
manufacturing and green hydrogen — which are key to the functioning of a smart grid.
BERNSTEIN
INDIA INDUSTRIALS & INFRA: OPPORTUNITIES FOR AI IN THE POWER SECTOR 131
EXHIBIT 21: L&T smart energy solutions
Source: Company website, Bernstein analysis
INDIAN DISCOMS — WHAT
COULD BE THE FUTURE?
ENEL (not covered) in Europe is a good example of a future power distribution/retail company
that started off as a traditional power utility. It launched its ENEL X offering to leverage the grid
as a platform (Exhibit 22). Its offerings include smart homes and cities, electric transportation,
intelligent public lighting, integration of power (including renewables) through smart grids, etc
(Exhibit 23). Indian DISCOMs (at least the private ones) could leverage ENEL's learnings and
incorporate them into their license areas. This would be a source of additional income and would
also set up the platform for new use cases of AI in power retail (e.g., smart grid operations to
optimize demand-supply from homes/distributed energy sources).
EXHIBIT 22: The future grid: grid as a platform
Source: Enel website, Bernstein analysis
EXHIBIT 23: ENEL X offerings go far beyond the role of
traditional power distribution companies
Source: Enel X website, Bernstein analysis
BERNSTEIN
132 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
VALUATION METHODOLOGY See Disclosure Appendix for valuation methodology.
EXHIBIT 24: Ratings and target prices
Ticker Rating Currency
25- May-2023
Closing
Target price
RNW O USD 5.47 9.6
ADANIGR.IN U INR 970.65 402.0
IEX.IN U INR 155.50 109.0
LT.IN O INR 2,204.70 2,452.0
ASIAX 1,141.60
Source: Bloomberg, Bernstein analysis and estimates
RISKS See Disclosure Appendix for risks.
Nikhil Nigania nikhil.nigania@bernstein.com +91 226 842 1414
Anusha Madireddy anusha.madireddy@bernstein.com +91 226 842 1 444
BERNSTEIN
INDIA INDUSTRIALS & INFRA: OPPORTUNITIES FOR AI IN THE POWER SECTOR 133
BERNSTEIN
134 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
BERNSTEIN
Financials
FINANCIALS 135
BERNSTEIN
136 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
ASIA INSURANCE: CHATBOTS TO MAKE
"LIFE" BETTER?
HIGHLIGHTS •AI competition goes beyond tech companies. Leading insurers have already joined
the battle. AI is one of the key investment areas of insurance companies, with applications
including lead generation, distribution, underwriting, pricing, claims, and customer services.
In Asia, although some leading insurers choose to go bold from incubation, most insurers are
investing through M&As and strategic partnerships. AI and data processing technologies have
been adopted more widely in auto insurance. AI helps drivers choose a safer route to reduce
the risk of accidents. Data collected based on driving behaviors enables dynamic pricing
decisions. With image-based AI assessment, insurers can make car accident claim decisions
within seconds in China today. The low-hanging fruit for life insurance names lies in claims and
customer services. Insurers in Japan have invested in AI-automated claim systems to replace
staff by as much as 30% today.
•Ping An is clearly leading AI applications for financial companies in Asia. It is the only
financial name that competes directly with the top global tech companies. Ping An ranks #5
globally in AI patents (after Tencent (covered by Bernstein analyst Robin Zhu), Baidu (covered
by Bernstein analyst Boris Van), IBM (covered by Bernstein analyst A.M. (Toni) Sacconaghi, and
Samsung (covered by Bernstein analyst Mark Li) and has obtained patents the fastest in the
past five years.
•That said, there are still more challenges to overcome. Challenge 1 — better data
doesn't mean better pricing. No one wants to be charged 5% more simply because he/
she slams on the brakes. In theory, the overweight should be charged x times higher than the
gym goers; in reality, insurers would lose customers with different (or higher) pricing. It is not
a simple process to leverage data. Not to mention that it takes years from data generation
to business insights and pricing decisions. Challenge 2 — a better robot doesn't mean
better services. A chatbot is useful to deal with simple requests, but it is hard to generate
personalized assistance. Agency face-to-face sales interaction is important to develop "long-
term relationships" and loyalty to brands. The margin of the tied agents is the highest among
all the distribution channels in Asia, thanks to the complexity of products that only agents
can sell. Challenge 3 — better tech doesn't mean better financial results…at least in
the near term. At this stage, insurers are investing in technologies to improve efficiency
and customer services. They are still far from seeing direct financial results after years of
investments. The game changer is for agents to embrace tech — it is not a competition,
but a partnership.
INVESTMENT IMPLICATIONS We rate Ping An, Prudential, and AIA outperform, and China Life Market-Perform.
BERNSTEIN
ASIA INSURANCE: CHATBOTS TO MAKE "LIFE" BETTER? 137
LEADING INSURERS HAVE
ALREADY JOINED THE
BATTLE…THOUGH STILL IN
THE EARLY STAGE
Testing the waters. Asia Insurers are still at the early stage with simple AI applications.
AI is one of the key investment areas of insurance companies, with applications including lead
generation, distribution, underwriting, pricing, claims, and customer services (Exhibit 2). In Asia,
although some leading insurers choose to go bold from incubation, most insurers are investing
through M&As and strategic partnerships. AI cannot go standalone; thus, long-term technology
investments and commitments are needed. Together with the development of big data, cloud
services, the Internet of Things, and blockchain technologies, insurance companies are exploring
various ways to improve efficiency, grow revenues, control risks, and reduce costs (Exhibit 3).
AI and data processing technologies have been adopted more widely in non-life auto
insurance. AI helps drivers choose a safer route to reduce the risk of accidents. Data collected
based on driving behaviors enables dynamic pricing decisions. With image-based AI assessment,
insurers can make car accident claim decisions within seconds in China today.
The low-hanging fruit — for life insurance names at this stage — lies in claims and
customer services. The life insurance market is still in the early stage of exploring its long-term
upside. Claim handling is typically time-consuming with human interaction and data collection.
Fukoku Mutual (not covered), the leading insurer in Japan, has invested in AI-empowered
claim systems to replace 30% of its claim workers since 2017. Ping An makes auto accident
claim assessments within seconds after drivers upload the photos. Northwestern Mutual (not
listed), a 166-year-old financial services company in the US, has also been piloting automated
underwriting and fraudulent claims detection through partnerships/M&A investments.
Know Your Customers (KYC) — it's all about the data! Data analytics allow insurers to access
target customers more effectively. Lifetime events trigger insurance purchases, thus, knowing
when the customer would get married, buy a new house/car, have a baby, visit doctors, and
plan for retirement becomes a secret weapon for insurance companies to unlock the top line
upside. Ping An P&C sells car insurance through AutoHome (not covered) as soon as a new car
is purchased. Leveraging the integrated data platform, Ant (not listed) accumulates over 100
million policyholders of its online mutual insurance product within a year.
Ping An is clearly leading AI applications for financial companies in Asia. Insurers are
still at an investing stage before meaningful financial impacts can be expected. Among all life
and P&C insurers, Ping An has clearly led in AI development. It is the only financial name that
shows up on the list of AI patent applications, competing head-to-head with the top global tech
companies. Ping An ranks #5 globally in AI patents (after Tencent, Baidu, IBM, and Samsung) and
has obtained patents the fastest in the past five years (Exhibit 1).
BERNSTEIN
138 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
EXHIBIT 1: Ping An is the only financial name among the top five leading AI tech companies and has obtained
patents the fastest in the past five years
Number of active AI and machine learning patent families held by the company
Microsoft is covered by Bernstein analyst Mark L. Moerdler; Alphabet is covered by Bernstein analyst Mark Shmulik.
Source: LexisNexis PatentSight, Statista
EXHIBIT 2: Leading insurers have already joined the battle, with applications including lead generation,
distribution, underwriting, pricing, claims, and customer services
Source: Bernstein analysis
BERNSTEIN
ASIA INSURANCE: CHATBOTS TO MAKE "LIFE" BETTER? 139
EXHIBIT 3: AI cannot go standalone...with the development of big data, cloud services, IoT, and blockchain
technologies, insurers are exploring various ways to improve efficiency, grow revenues, control risks, and reduce
costs
Source: Bernstein analysis
CHALLENGES BEFORE "LIFE"
GETS BETTER
Challenge #1 — better data doesn't mean better pricing
No one wants to be charged 5% more simply because he/she slams on the brakes or fails to
walk 10,000 steps a day. The beauty of insurance is that actuaries cannot predict risks at the
individual level. In other words, if actuaries are accurate in calculating every policyholder's risk,
there is no "risk" by definition. In theory, the overweight should be charged x times higher than
the gym goers; in reality, insurers would lose customers with different (or higher) pricing. It is not
a simple process to leverage data. Not to mention that it takes years from data generation to
business insights and pricing decisions.
Challenge #2 — a better robot doesn't mean better services
Generally speaking, higher-margin insurance products are more complicated in design. A
chatbot is useful to deal with simple requests, but it is hard to generate personalized assistance.
Agency face-to-face sales interaction is important to develop "long-term relationships" and
loyalty to brands. The new business margin of the tied agents is the highest among all the
distribution channels in Asia, thanks to the complexity of products that only agents can sell. That
said, technology can empower agents to better understand customers and differentiate their
selling strategies.
Challenge #3 — better tech doesn't mean better financial results…at least in the
near term
At this stage, insurers are investing in technologies to improve efficiency and customer services.
They are still far from seeing direct financial results after years of investments. Insurers can
choose to replace legacy systems either with third-party tech solutions or in-house development;
in any case, it is cash-burning with initial one-off investments and ongoing maintenance fees
each year (Exhibit 4). Ping An has claimed to invest 1% of its revenue, or about 10% of its profit
each year, in technology R&D, taking years to see meaningful financial impacts on P&L (Exhibit 5).
The game changer is for agents to embrace tech — it is not a competition, but
a partnership
We argued in the previous note (China InsurTech: Digital, Distributor, or disruptor? Who is the
beneficiary among all competitors?) that although digital cannot replace life agents in China, tech
can empower agents and differentiate the selling process.
BERNSTEIN
140 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
With chatbots handling routine queries and repetitive tasks, insurance agents can focus more
on complex requests and relationship management to improve efficiency and provide a better
customer experience. Technologies can also help agents select better leads, tell a better story
to address the pain points, facilitate faster fulfillment, and automate the claims process, with a
goal to transform into a "tech-equipped quality agency" model, delivering productivity gains and
financial impacts (Exhibit 6).
EXHIBIT 4: Insurers investing in AI technologies...a long-term commitment
Source: Media
BERNSTEIN
ASIA INSURANCE: CHATBOTS TO MAKE "LIFE" BETTER? 141
EXHIBIT 5: Early stage of direct and indirect financial impacts from Ping An's AI investments
Source: Company Investor Day presentation
EXHIBIT 6: The game-changer is for agents to embrace tech — it is not a competition, but a partnership.
Source: Bernstein analysis
USE CASES OF GLOBAL
INSURANCE COMPANIES
Ping An (2318.HK/601318.CH, covered)
Ping An has accumulated over 227 million retail customers in the past three decades in China.
Over recent years' investments and expansions, there are also:
•6.6 million SME owners borrowing through Lufax;
•35.5 million DAU (car owners and car dealers) on AutoHome's platform;
•460 Chinese banks and over 1,800 small and medium financial services using Ping An's
OneConnect financial platform; and
BERNSTEIN
142 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
•373 million registered users and 1.1k+ enterprises covered by Ping An Good Doctor (Exhibit
7).
Today, the integration model benefits Ping An Group in terms of: (1) cross-selling and sharing
customer resources; (2) product + services bundling; and (3) technology empowerment, with
applications in:
•"AI customer services" for online policy administration, underwriting, and claims settlement
(Exhibit 8);
•"AI doctor" for smart diagnosis and treatment;
•"Face recognition" to enable customers to collect their insurance payments quickly, efficiently,
and securely;
•"Cloud claims" in auto insurance, a personalized service based on analysis of customers' data;
•"Smart investment consultancy service" on Ping An Pocket Bank, customizing investment
portfolio accounting to risk appetites;
•"Advanced anti-fraud monitoring system" in Ping An Bank's credit card center to reduce direct
and indirect economic losses;
•"AI smart stock investment" for Ping An Securities to recommend stock positions and smart
asset allocation solutions;
•"Smart Wealth Management" for high-end customers, with one-on-one asset allocation
service; and
•"KYC system" of Lufax to enhance the precise matching of products and investors.
EXHIBIT 7: Ping An ecosystem: it is about more than selling insurance
Source: Company reports, Bernstein analysis
BERNSTEIN
ASIA INSURANCE: CHATBOTS TO MAKE "LIFE" BETTER? 143
Zhong An (6060.HK, not covered)
ZhongAn, the first pure online insurer in China, has developed a life insurance platform (SpeedUp)
and served 450 million customers to date. The system is able to process insurance claims
and policies at a speed of up to 14,000 units per second and respond within one millisecond
for searching. Besides this, it is able to integrate the obtained customer information, cater to
individual demands, and propose a precise selling strategy (Exhibit 9).
AIA (1299.HK, covered)
AIA has developed the Vitality program with Discovery SA (DSY, not covered) to collect
customers' lifestyle data. DSY tracks customer activity through wearables and offers pricing
discounts on a range of products and services. It has encouraged more than 5.5 million members
to lead healthier lifestyles and redefines traditional insurance pricing.
Progressive (PGR.US, not covered)
Progressive, one of the largest auto insurers in the US, launched Snapshot in 2011. Customers
install Snapshot in their car, which can automatically detect driving behaviors and update
insurance premiums. The algorithm seemingly makes the premium update based on the habit
records and updates the new pricing to customers without human interaction (Exhibit 11).
Northwestern Mutual (not listed)
AI has been adopted for automating the underwriting process, understanding customers'
"frustration and fears," forecasting claims, detecting frauds, and predicting propensity scores.
The company has rolled out an underwriting process in life insurance since 2017 by leveraging an
intuitive application — a client-declared medical history questionnaire — for fast and automated
underwriting without medical tests. The advanced analytics, historical data, and predictive
underwriting models help shorten underwriting time from 28 days to one day.
Aditya Birla Sun Life (JV of Aditya Birla Group and Sun Life Financial, not listed)
Aditya Birla Sun Life has launched a virtual insurance agent built on conversational AI technology
— "DISHA," specializing in automating simple customer interactions on a large scale (Exhibit 10).
To date, DISHA has answered over 2.3 million questions without human interaction and helped
reduce call center volume by 15%.
BERNSTEIN
144 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
EXHIBIT 8: Use case 1 — Ping An: insurers can be smarter
Source: Company presentations
EXHIBIT 9: Use case 2 — Zhong An: insurance can be simpler and faster
Source: Company presentation
BERNSTEIN
ASIA INSURANCE: CHATBOTS TO MAKE "LIFE" BETTER? 145
EXHIBIT 10: Use case 3 — Aditya Birla Sun Life: chatbots can reduce costs and handle simple requests
Source: Bernstein analysis
EXHIBIT 11: Use case 4 — Progressive: insurance can be innovative
Source: Company presentations
BERNSTEIN
146 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
VALUATION METHODOLOGY See Disclosure Appendix for valuation methodology.
EXHIBIT 12: Ratings and price targets
Source: Bloomberg, Bernstein analysis and estimates
RISKS See Disclosure Appendix for risks.
Tianjiao Yu tianjiao.yu@bernstein.com +852 2918 5798
Cally Yang cally.yang@bernstein.com +852 2918 5790
BERNSTEIN
ASIA INSURANCE: CHATBOTS TO MAKE "LIFE" BETTER? 147
BERNSTEIN
148 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
BERNSTEIN
Healthcare
HEALTHCARE 149
BERNSTEIN
150 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
GLOBAL CDMOS: USE CASES OF AI IN
BIOLOGICS DRUG DISCOVERY
HIGHLIGHTS •In recent years, many AI algorithms and tools enabling the discovery and optimization of
biologics have been developed (e.g., AlphaFold), and many companies have been founded
that focus exclusively on biologics drug discovery. A COVID-19 therapeutic (Bamlanivimab)
became the first AI-enabled biologic to score a US FDA EUA in 2020. We identified 14 AI-
native companies for whom AI is central to the biologics/protein discovery program. These
companies have disclosed >580 programs at various stages of development. They represent
~6% of the total biologics under development in 2023.
•AI finds applications across the biologics drug discovery workflow — in structure prediction
and epitope definition, in diversity generation and screening where AI can be used for
optimization of naturally occurring repertoires of biologics and in the design of novel libraries,
in humanization and engineering where AI can help optimize the PK/PD profile, and several
others. The focus seems to be exclusively on the discovery process, but we note some
business models spilling over into preclinical and cell line development, which has so far been
the domain of CDMOs.
•Wuxi Biologics and Biocon (through Syngene) in our coverage offer biologics discovery
services. Wuxi Biologics claims to leverage AI to generate novel diversified binders. Syn.AI
is Syngene's AI offering, though the focus seems to be small molecules. Lonza, Samsung
Biologics, Catalent (not covered) and Fujifilm (not covered) do not offer discovery services
today. To keep abreast of technology and to protect themselves from potential future scope
creep (by AI-native companies), it will be important for CDMOs to keep a close watch on the
space and jump in when the time is right through partnerships/deals.
INVESTMENT IMPLICATIONS We rate Wuxi Biologics Outperform (target price HKD88), Samsung Biologics Ourperform (target
price KRW1,031,932), Lonza Outperform (target price CHF672), and Biocon Outperform (target
price INR314).
AI IN BIOLOGICS DRUG
DISCOVERY
AI is already transforming the way we discover and study biologics. More than 50 AI-enabled
biologics are in different stages of discovery, preclinical, and clinical development. In this chapter,
we briefly discuss the biologics discovery workflow, use cases for AI and emerging players, and
their offerings. We were fascinated by what we learnt and the promise AI/ML holds.
Over the last 5-10 years, we have seen rapid progress in AI applications in small molecules with
several drugs in the clinic. In the recent past, we have seen attention to the field of biologics
discovery as well. Many AI algorithms and tools enabling the discovery and optimization of
biologics have been developed (e.g., AlphaFold), and many companies have been founded that
focus exclusively on biologics drug discovery. A COVID-19 therapeutic (Bamlanivimab) became
the first AI-enabled biologic to score a US FDA EUA in 2020.
BERNSTEIN
GLOBAL CDMOS: USE CASES OF AI IN BIOLOGICS DRUG DISCOVERY 151
The field has made long strides in: (a) creation of large databases such as Genomics
(DNA sequencing), Proteomics (protein sequencing), and Cryo-EM and AlphaFold (protein 3D
structures, folding and visualization), (b) advances in machine learning algorithms that enable
mining of these databases, and (c) exponential growth in computational power that serves as the
backbone. The convergence of these factors promises to reduce timelines and cost, improve the
quality and novelty of molecules, and increase R&D success rate.
•We identified 14 AI-native companies for whom AI is central to the biologics/protein discovery
program. These companies have disclosed >580 programs at various stages of development.
They represent ~6% of the total biologics under development in 2023.
•AI finds application across the biologics drug discovery workflow — in structure prediction
and epitope definition, in diversity generation and screening where AI can be used for
optimization of naturally occurring repertoires of biologics and in the design of novel libraries,
in humanization and engineering where AI can help optimize the PK/PD profile, and several
others. Most companies today are exploring one or the other part of the value chain, and we
expect more value will be unlocked in the future from end-to-end applications.
•The focus seems to be exclusively on the discovery process, but we note some business
models spilling over into preclinical and cell line development, which has so far been the
domain of CDMOs. AbSci (not covered), for example, boasts of a proprietary E.coli cell line
(SoluPro) that is scalable and can support manufacturing from discovery to commercial. Usual
practice is to use a microbial transient cell line in early stages and move to a CHO cell line
when larger quantities of the protein are required.
•Wuxi Biologics and Biocon (through Syngene) in our coverage offer biologics discovery
services. Wuxi Biologics leverages AI to generate novel diversified binders. Syn.AI is Syngene's
AI offering, though the focus seems to be small molecules. Lonza, Samsung Biologics,
Catalent, and Fujifilm do not offer discovery services today. To keep abreast of technology and
to protect themselves from a potential future scope creep (by AI-native companies), it will be
important for CDMOs to keep a close watch on the space and jump in when the time is right
through partnerships/deals.
The biologics drugs market has been growing with a greater share of total drug sales and more in
approvals in recent years. Biologics drugs are estimated by consensus to gain more penetration
in the pharma industry, with their share of total drug sales likely to grow from ~30% in 2012 to
~53% in 2028. (Exhibit 1).
BERNSTEIN
152 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
EXHIBIT 1: Biologics drug market is expected to grow fast and continue to increase its share of total drug sales and
more in approvals
Source: Evaluate, US FDA, Bernstein analysis
Biologics therapies have made breakthroughs in the past few decades with several blockbuster
drugs, such as Rituxan, Herceptin, Remicade, etc., with new targets and indications (Exhibit 2).
BERNSTEIN
GLOBAL CDMOS: USE CASES OF AI IN BIOLOGICS DRUG DISCOVERY 153
EXHIBIT 2: Biologic therapies have made breakthroughs in the past decades with several blockbuster drugs
Source: Adapted from Journal of Biomedical Science, Bernstein analysis
Drug discovery and development could be complicated and costly (Exhibit 3). In total, the time for
a drug from discovery to commercialization could be as long as 13 years with potential cost as
high as US$2.7Bn. AI could jump in and help on some aspects in this whole process with possible
shortened time and some cost savings.
BERNSTEIN
154 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
EXHIBIT 3: Drug discovery and development can be complicated and costly; AI could jump in and help on some
aspects
Source: Tufts CSDD "Cost of Developing a New Drug," US FDA, UK Office of Health Economics, company disclosures, Bernstein analysis
We listed several traditional ways of therapeutic antibodies' discovery, demonstrating the
complexity (Exhibit 4). The screening process can take a long time, and AI tools could help
accelerate.
BERNSTEIN
GLOBAL CDMOS: USE CASES OF AI IN BIOLOGICS DRUG DISCOVERY 155
EXHIBIT 4: Several traditional ways of therapeutic antibodies discovery, demonstrating the complexity
Source: Adapted from Journal of Biomedical Science, Bernstein analysis
The traditional approaches we have listed have their pros and cons. Challenges exist in aspects
such as 3D structure maintenance and technical aspects of selection and screening (Exhibit 5).
BERNSTEIN
156 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
EXHIBIT 5: Traditional approaches have their pros and cons
Source: Adapted from cell.com, Bernstein analysis
Lilly's bamlanivimab is one example of AI-assisted drugs, demonstrating potential in the power
of AI in the drug discovery space. AI tools are powerful in several aspects such as 3D structures
modeling and can improve efficiency when working with a large volume of data (Exhibit 6).
BERNSTEIN
GLOBAL CDMOS: USE CASES OF AI IN BIOLOGICS DRUG DISCOVERY 157
EXHIBIT 6: Lilly's bamlanivimab is one example of AI-assisted drugs, demonstrating the potential power of AI in
the drug discovery space
Source: Company disclosures, Bernstein analysis
There are several areas where AI could help throughout the drug discovery process. For example,
deep learning algorithms can predict protein functions based on trainings in protein sequences
and 3D structures, neural networks could help identify target proteins by screening large scale
databases, and so on (Exhibit 7).
BERNSTEIN
158 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
EXHIBIT 7: There are several areas where AI could help throughout the drug discovery process
Source: Bernstein analysis
There are a few players in the AI drug discovery field, including Absci, AbCellera (not covered),
OmniAb (not covered), Generate Biomedicines (private), Adimab (private), and Wuxi Biologics.
Most focus on the pre-IND part in lead identification and lead optimization. Wuxi Biologics AI
application could assist in the whole drug value chain from discovery to commercialization,
providing novel binder generation and de novo antibody development (Exhibit 8).
BERNSTEIN
GLOBAL CDMOS: USE CASES OF AI IN BIOLOGICS DRUG DISCOVERY 159
EXHIBIT 8: There are a few players in the AI drug discovery field
Source: Company disclosures, Bernstein analysis
VALUATION METHODOLOGY See Disclosure Appendix for valuation methodology.
EXHIBIT 9: Ratings and target prices
25-May-2023 Target
Ticker Rating Currency Closing Price Price
LONN.SW O CHF 565 672
207940.KS O KRW 782,000 1,031,932
2269.HK O HKD 43 88
BIOS.IN O INR 240 314
ASIAX 1,142
MXAPJ 504
EDM 1,123
Source: Bloomberg, Bernstein analysis and estimates
RISKS See Disclosure Appendix for risks.
Nithya Balasubramanian nithya.balasubramanian@bernstein.com +91 226 842 1433
Parth Shah parth.shah8@bernstein.com +91 226 842 1464
BERNSTEIN
160 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
EU MEDTECH: AI IN MEDICAL IMAGING
HIGHLIGHTS •AI in medical imaging (radiology) is a reality today. Analyzing images is one of the
applications particularly well-suited to machine learning/AI. Thus, one of the most obvious
opportunities for AI in healthcare is identification of anomalies in medical scans. In the past
few years, large imaging players, some tech titans, and numerous start-ups have developed
various applications designed to enable more accurate and faster diagnosis of lesions. While
it is unlikely that computers will replace radiologists any time soon (not least of all due to
regulatory complexities), clinicians are gradually acknowledging that incorporating AI into the
radiology suite is a necessary move, as radiologists today are overwhelmed with an ever-
increasing amount of data. According to a 2020 study by the American College of Radiology,
an estimated 30% of US radiologists use AI tools. While today's AI programs are largely aimed
at diagnosis, some of the most commercially popular applications are focused on speeding up
scan times and improving radiologist workflows. In clinical therapy, AI is becoming integrated
into treatment pathways in order to improve outcomes in the cath lab and beyond.
•Who are the competitors in the field today, and who is best-positioned? The leading
imaging OEMs include Siemens Healthineers, Philips, and GE Healthcare, and Elekta in
radiotherapy. Among the tech titans are primarily Google DeepMind and Merative (formerly
IBM Watson Health; private). There are also countless start-ups (e.g., Arterys, HeartFlow, and
MIM). As access to data is the lifeblood of AI, thus far, not only have the equipment players
managed to avoid the fate of the "dumb pipe," in fact they benefit from unrivalled access, as
they have agreements with their customers (hospitals and healthcare systems) that provide
them with rights to use patient scans for research purposes (Siemens Healthineers has a
library of 1.5 billion curated images).
•Imaging equipment manufacturers are well-positioned to identify the commercial
opportunities. As imaging OEMs are closest to the end-customers (both physicians including
radiologists, cardiologists, oncologists, etc., and procurement departments and senior
management at healthcare facilities), they have a deep appreciation for the aspects of today’s
imaging workflows that could be improved with AI, and what applications would actually
be commercially attractive. Meanwhile, they have the deepest understanding of the current
limitations of each of the medical imaging modalities within their own product offerings;
their reps are in the radiology suite day-in and day-out, best positioning them to monetize
opportunities by identifying and understanding the real bottlenecks in imaging diagnostics.
Overall, we expect these commercial opportunities to expand the size of the total imaging
market, as the sale of software (either as one-time packages or via SaaS contracts) becomes
an increasingly larger proportion of sales.
INVESTMENT IMPLICATIONS We rate Siemens Healthineers Outperform with a target price of €60.50. We rate Philips Market-
Perform with a target price of €17.00.
BERNSTEIN
EU MEDTECH: AI IN MEDICAL IMAGING 161
INTRODUCTION For the full version of this chapter (where we also discuss AI in radiotherapy, amongst other
things), see EU Medtech: AI in Medical Imaging & Radiotherapy - Getting Smarter... the Real
Opportunities for Artificial Intelligence.
All the major imaging players are involved in AI development today, and as the Philips CFO likes
to frequently point out, they employ more software engineers in their R&D division than any
other type of professional. While AI is still in its early days, based on our research and analysis
of the industry over the past several years, we have reached the following conclusions and
observations:
•Data is the lifeblood of AI, and OEMs have access to the best. Access to high-quality
data (i.e., professionally annotated and curated training sets) is perhaps the most important
determinant of being able to build a successful AI program, and OEMs have a natural
advantage via their existing relationships with their customers. Over the past several years
they have leveraged these relationships as they have increasingly entered into long-term
partnerships with healthcare systems that help expand their access to patient scans for
research purposes. Also, given their long history of collecting sensitive patient data and using
it to improve imaging technology, they are already a trusted partner (not so for tech companies
in recent years — see here and here). Siemens Healthineers says it now has 1.5 billion curated
medical images within its research library.
•Implementation of AI across the medical imaging market will ultimately be somewhat
fragmented. Today, there are already several small software firms involved in radiology with
an outsized presence in the market (e.g., TeraRecon (private) and Carestream (before its sale
to Philips). We expect there to be some winners within the landscape of today's start-ups
(although it is still a bit early to call which ones), either as they reach commercial scale or get
acquired. It is also possible for the likes of DeepMind to commercialize an application that is a
must-have for a specific use case (e.g., its foray with the NHS into eye exams — see here). But
overall, we expect most of the commercially successful applications to come from the large
equipment manufacturers (and some of the smaller ones such as Canon, Fujifilm, etc.).
•We expect the development of AI in medical imaging to be more evolution than
revolution. AI will help make workflows more efficient, lead to quicker and more accurate
diagnoses, and will ultimately become integrated into clinical decision-making in every
radiology suite around the world. But we think we are unlikely to see a future where AI
programs "replace" human expertise and render radiologists' jobs redundant any time in the
near to mid-term, not least due to regulatory complexities as well as legal liability reasons. But
it is possible that there will be some AI applications that will enable better access to imaging
technologies in resource-constrained environments such as emerging markets, where the
choice is between a computer or nothing.
AI IN HEALTHCARE The list of potential applications for AI in healthcare is almost endless — from enabling population
health, to designing individualized treatment plans, to creating new channels for care delivery, to
discovering better drugs and biomarkers. AI can be applied to various types of healthcare data,
both structured (e.g., x-rays, CT scans, photographs, laboratory results, etc.) and unstructured
(e.g., clinical notes, medical journals, textbooks, etc.). Machine learning methods are useful for
structured data, while natural language processing is required for unstructured data.
BERNSTEIN
162 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
The global pandemic highlighted how woefully underprepared most developed market
healthcare systems were for not only a major exogenous shock (which is somewhat
understandable), but also how ill-equipped they are to handle increased patient volumes, which
is a reality of an aging population that is also getting more unhealthy.
AI (and its less advanced cousins machine learning and data analytics) is expected to become
an integral part of healthcare, from how we learn to how we deliver care. In order to understand
where AI may have the greatest promise in the landscape, it is necessary to identify what
problems AI could help solve.
AI will make doctors more efficient, but also potentially better. We do not expect AI-enabled
platforms to replace doctors just yet, as human judgement is likely to be as relevant in healthcare
in 50 years as it is today. Physicians encounter things every day that don't conform to previously
known patterns, while they also must consider non-medical implications in treatment path
selection, and have to make judgments and decisions quickly and accurately based on a deluge
of information.
We see three areas where AI more broadly has the potential to help improve healthcare systems:
(1) speeding up workflows for clinicians and increasing efficiency, (2) improving accuracy and
reducing error, and (3) eventually being able to do things that are beyond human capabilities.
AI IN DIAGNOSTIC IMAGING
AND CLINICAL THERAPEUTICS
Diagnostic imaging is an obvious area where AI can help clinicians, and a key business for
two of the companies in our coverage — Siemens Healthineers and Philips. Scans produce
structured data (images) which are analyzed by doctors. Medical imaging software already exists
to help practitioners analyze these images, but most are currently still manual, tedious, and only
use a fraction of available data. Meanwhile, conclusions drawn by medical professionals based
on existing analytical software are ultimately subjective. AI can help modernize this process,
moving interpretation toward being automatic, accurate, quantitative, data-driven, consistent,
and fast. Today, existing software can help dermatologists distinguish cancerous lesions, use
facial recognition to diagnose rare genetic diseases (e.g., DiGeorge Syndrome), or automatically
highlight anatomical features in MRIs, x-rays, and other medical images.
In the long term, there is potential in transforming radiology from a discipline of qualitative
interpretation to one of quantitative analysis. This shift is often referred to as radiomics. But in
imaging, just like in healthcare in general, in the near-to-medium-term, AI's main impact will likely
be more around speeding up workflows, increasing the efficiency of radiology suites, reducing
errors, and improving the accuracy of diagnosis.
Revolution or evolution? We asked ChatGPT what it thinks
While we share the excitement about a potential AI revolution in healthcare in the long-term,
we believe the first phase will be more of an evolution of the current system. In the near-term,
AI innovation will focus on improving the routine and freeing time for healthcare professionals,
i.e., taking away the "a monkey could do this" parts of imaging analysis, such as the labeling
and contouring of anatomical features and tumor boundaries, leaving radiologists to focus on
complex image interpretation, thus increasing productivity.
AI will ultimately become integrated into clinical decision-making in every radiology suite around
the world. In fact, according to a 2020 study by the American College of Radiology, an estimated
BERNSTEIN
EU MEDTECH: AI IN MEDICAL IMAGING 163
30% of US radiologists already use AI tools. But we think we are unlikely to see a future where
AI programs "replace" human expertise and render radiologists' jobs redundant any time in the
near-to-mid-term, not least due to regulatory complexities as well as legal liability reasons. It
is possible, however, that there will be some AI applications that will enable better access to
imaging technologies in resource-constrained environments such as emerging markets, where
the choice is between a computer or nothing.
Speeding up workflows and increasing efficiency
Worldwide, there is a big discrepancy between the number of doctors trained in radiology and
the rising demand for diagnostic imaging. This bottleneck creates a great need for efficiency
and productivity solutions. Between 2015 and 2019 in the UK, the average number of people
waiting at least six weeks for a CT or an MRI increased from 2,000 to 9,000 — a 350% increase.
In the same time, the number of imaging professionals increased by 23%, while the number of
exams carried out increased by only 8%. So, if the imaging staff rose more than the number of
tests, why has the wait list increased so much? It implies that imaging throughput of machines is
the limiting factor to getting patients scanned. And based on the number of installed machines
in 2018 and 2019, the maximum number of tests which can be performed in a year is ~33
million, irrespective of the number of staff. Therefore, in the absence of building new hospitals
or radiology departments, the only way to reduce wait times is to increase the number of scans
that can be performed by the same number of machines and employees through faster image
acquisition and faster interpretation, driven by AI.
REDUCING SCANNING TIMES A highly useful application for AI in medical imaging is the ability to speed up scanning times.
Medical imaging today is relatively time-consuming for technicians and patients. Even excluding
preparation time, once a patient is in the scanner, the image acquisition process can take up to an
hour depending on the scan (particularly MRI). As speeding up this process would improve patient
experience, reduce costs, and improve throughput (allowing hospitals to provide more scans),
there is significant interest in reducing scanning times. Speeding up the process by under-
sampling can reduce scan time by a factor of six. However, this produces a low-quality image.
With machine learning, programs have been developed that can reconstruct usable images from
under-sampled data. These reconstructed images can be comparable in quality to traditionally
acquired images. Exhibit 1 shows a low-quality image on the left, a reconstructed usable image
in the center, and a traditionally acquired image on the right.
BERNSTEIN
164 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
EXHIBIT 1: Image reconstruction using AI can reduce scan times
Source: Dr. Ben Glocker at 2018 Bernstein AI Impact on Sectors Conference (with permission), Bernstein analysis
SPEEDING UP
INTERPRETATION TIMES WITH
AUTOCONTOURING
Helping speed the process of interpretation is perhaps the most promising near-term benefit
of AI applications. More than half (58%) of radiology leaders say they do not have enough
diagnostic and interventional radiologists to keep patients safe. According to the Royal College
of Radiologists, the NHS radiologist workforce is now short-staffed by 33% and needs about
2,000 consultants to meet safe staffing levels and pre-COVID-19 volumes of scans.
An average radiologist can interpret an image every three to four seconds, eight hours a day. But
in certain situations, it can take hours to fully annotate a radiology report. One of the seemingly
simple ways in which AI can help speed workflows is via automatic labeling of scans. The software
can automatically identify and label major items of anatomy, saving significant time for the
clinician. This feature is referred to as "autocontouring."
Designing algorithms that could complete these sorts of tasks with more accuracy than human
experts is a focus for many developers, but it is worth noting that there is significant value in just
matching human performance, if you are doing it in less time. Algorithms that can perform these
tasks in a fraction of the time it would take human radiologists have the potential to accelerate
patient care and free up radiologists for other tasks and more important responsibilities. In
a healthcare landscape struggling with increasing cost pressures and growing demand for
diagnostic imaging, along with a relative shortage of radiologists, any efficiency gains are
valuable.
Reducing errors and improving accuracy
Radiologists can scrutinize hundreds of images before identifying an area of concern in a
patient's body. It is like playing "Where's Waldo" on each of those images. Once an abnormality is
spotted, the radiologist then needs to find evidence of disease, and if so, what disease, and if it
needs intervention. In other words, whether it is Waldo, Wilma, Willy, or Wally. A study by Kakinuma
et al. (1999) analyzed 1,443 subjects screened for lung cancer, and found that of 22 recognized
tumors, seven were visible on images from earlier CT examinations. The causes of these missed/
failed diagnoses were size, location, and the presence of confounding factors. This is another
area where AI could be a game changer.
BERNSTEIN
EU MEDTECH: AI IN MEDICAL IMAGING 165
POSITIONING OF PATIENTS Among the "basic" uses of AI-driven software in imaging, patient positioning is one of the most
important. AI efforts here focus on getting patients correctly positioned in the MRI or CT machine
prior to their scan, which ensures better image quality, limited dosing of radiation, and avoids
requiring the patient to come back for a second scan (therefore limiting lost time/money for
hospitals). The major imaging companies now offer software platforms with their machines that
are designed to make sure that patient positioning is correct the first time (and done quickly), e.g.,
Siemens' FAST Integrated Workflow and Philips' iPatient platform.
Healthineers' FAST Integrated Workflow, in conjunction with the SOMATOM CT machine's 3D
infrared camera, is able to capture the patient's physical characteristics and then use algorithms
to determine table height, whether the patient should lie head or feet first, prone or supine, etc.
Philips' iPatient platform for the Ingenia MRI offers tools to optimize patient positioning, as well
as automating many of the processes that operators would otherwise have to do manually.
TREATMENT PLANNING Another way in which AI can be integrated into treatment is through personalized procedure
planning. AI solutions have the computing power to flex a huge number of variables when
performing risk assessment and determining treatment plans for individual patients. As they are
able to draw on the imaging results already garnered at diagnosis, and potentially an enormous
history of past cases as well, the imaging players should be well-placed to develop solutions that
help guide physicians throughout the treatment pathway. One example of a company already
participating in this space is HeartFlow (private), a medical start-up founded in 2007 and focused
on personalizing cardiac care. Specifically, the company seeks to improve the diagnosis and
treatment of coronary artery disease (CAD), where plaque in the artery walls is obstructing blood
flow to the heart (Exhibit 2).
EXHIBIT 2: Heartflow uses deep learning to help doctors choose treatment pathways for specific patients
Source: Dr. Ben Glocker at 2018 Bernstein AI Impact on Sectors Conference (with permission), Bernstein analysis
ENHANCING IMAGING IN
REAL-TIME IN THE CATH LAB
TO IMPROVE OUTCOMES
Starting in cardiology, interventions in the cath lab are particularly interesting areas for
development within the AI field. A major obstacle to producing useful images of the heart is that
the heart is always moving. This typically necessitates the use of lower-resolution images. Here,
AI is not used for diagnosis, but rather for image segmentation on a real-time basis, as well as to
make imaging clearer (retouching an image).
BERNSTEIN
166 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
One particularly challenging procedure is Chronic Total Occlusion Percutaneous Coronary
Intervention (CTO PCI), a minimally invasive treatment for complete blockages of coronary
arteries. The procedure involves inserting a balloon catheter into the artery via blood vessels
near to the blockage. To perform the procedure successfully, surgeons must overcome anatomic
ambiguity between major vessels and side branches, as well as identify calcium build-ups, which
can prevent successful entry into the vessel. Current CT imaging technologies can help surgeons
know where in the vessels there is a build-up of calcium, reducing the failure risk, as well as
highlight and delineate between main vessels and side branches (Exhibit 3). However, they are
less good at providing high-quality real-time imaging of a structure in motion than traditional
fluoroscopy.
Deep learning offers the opportunity to optimize the fusing of images from multiple modalities
(e.g., CT and fluoroscopy), which could allow high-quality real-time imaging of coronary arteries
and vessels, while offering live image segmentation and annotation (Exhibit 4).
EXHIBIT 3: Current technology can color code the main
blood vessels and side branches for vascular surgery
Source: Siemens symposium at TCT 2017, Bernstein analysis
EXHIBIT 4: Deep-learning could improve access to high-
quality real-time imaging of vessels with segmentation
Source: Siemens symposium at TCT 2017, Bernstein analysis
WHAT DOES AI MEAN FOR
DIAGNOSTIC IMAGING
AND RADIOTHERAPY
INCUMBENTS?
A key question about the evolution of AI is: What is the potential value creation from these
platforms? And who will be the winners? Currently, nobody knows the full answer. When we
wrote our primer in 2018, we were unsure if imaging players would be able to monetize their
investments in AI, or whether the technology would simply get integrated into machines and
processes with limited ability to charge for it. Five years on, it is clear that there is a route to
monetization, largely in the form of incremental software add-ons, but also premium pricing for
machines that incorporate AI functionality. Philips also has some software-as-a-service (SaaS)
business lines, but this remains less common. It is also clear today that the medical capital
equipment OEMs are actually well-positioned versus the tech competition.
Competition from the land of tech
At this intersection of technology and healthcare, medical equipment OEMs are competing
against numerous players from the tech world. This includes some tech titans, notably Google
DeepMind and Merative (formerly IBM Watson Health). There are also start-ups in Silicon Valley
and beyond such as Arterys (Tempus), Quantib, MIM Software (all private), Nanox.AI (formerly
Zebra Medical Vision; not covered), and countless others. But all the major imaging players are
also involved in AI development today, and as the Philips CFO likes to frequently point out, they
employ more software engineers in their R&D division than any other type of professional.
BERNSTEIN
EU MEDTECH: AI IN MEDICAL IMAGING 167
The US FDA provides a comprehensive list of FDA-approved AI platforms (see website here),
which shows that 154 different companies have gained approval for one or more AI application in
radiology. While it is difficult to tell who is ahead of whom in the race to build out AI capabilities, our
research indicates that imaging equipment players are well-placed due to their close relationship
with hospitals and health systems, which puts them in a better position to access data.
Collaboration is required between the companies developing AI and the healthcare providers
who have control over patient data, which is the lifeblood required for any AI application. In
our extensive field interviews, AI experts have also commented that imaging incumbents "really
understand" what the customers (clinicians) want, as they are in the hospital interacting with them
every day.
Ultimately, we believe the implementation of AI across the medical imaging market will be
somewhat fragmented. There are already several small software firms involved in radiology with
an outsized presence in the market (e.g., TeraRecon and Carestream (before its sale to Philips)).
We expect there to be some winners within the landscape of today's start-ups (although it is still
early to call which ones), either as they reach commercial scale or get acquired. It is also possible
for the likes of DeepMind to commercialize an application that is a must-have for a specific use
case (e.g., its foray with the NHS into eye exams — see here). But overall, we expect most of
the commercially successful applications to come from the large equipment manufacturers (and
some smaller ones such as Canon, Fujifilm, etc.).
Who controls the data — avoiding the fate of the dumb pipe
A key question as the AI ecosystem develops is who gets to place themselves where in the value
chain. Particularly important for imaging equipment manufacturers, using a template from the
telecom industry, is that they don't just want to be the "dumb pipe," supplying image data to
AI software players who extract most of the value. To avoid the fate of the dumb pipe, Philips,
Siemens Healthineers, and GE Healthcare (not covered) are deploying and developing their own
AI applications. All three are integrating AI into their imaging machines to accelerate workflows
and increase diagnostic accuracy. OEMs are also looking to provide software platforms for other
AI developers and researchers to integrate their offerings in a centralized location. This gives
manufacturers some control of AI development around their machines and allows them to extract
some of the value from third-party software offers.
IMAGING EQUIPMENT
MANUFACTURERS ARE
WELL-POSITIONED TO
IDENTIFY COMMERCIAL
OPPORTUNITIES
As imaging OEMs are closest to the end-customers (radiologists, cardiologists, oncologists, etc.,
as well as procurement departments and senior management at hospitals), they have a deep
appreciation for the aspects of today’s imaging workflows that could be improved with AI, and
what applications would actually be commercially attractive. Meanwhile, they have the deepest
understanding of the current limitations of each of the medical imaging modalities within their
own product offerings. Overall, we expect these commercial opportunities to expand the size
of the total imaging market, as the sale of software (either as one-time packages or via SaaS
contracts) becomes an increasingly larger proportion of sales.
DATA IS THE LIFEBLOOD OF AI,
AND OEMS HAVE THE BEST
ACCESS
Given that access to high-quality data (i.e., professionally annotated and curated training sets) is
perhaps the most important determinant of being able to build a successful AI program, OEMs
have a natural advantage via their existing relationships with their customers. Over the past
several years they have leveraged these relationships as they have increasingly entered into long-
term partnerships with healthcare systems that helps expand their access to patient scans for
BERNSTEIN
168 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
research purposes. Also, given their long history of collecting sensitive patient data and using
it to improve imaging technology, they are already a trusted partner (not so for tech companies
in recent years — see here and here). Siemens Healthineers says it now has 1.5 billion curated
medical images within its research library.
WHAT PHILIPS,
HEALTHINEERS, GE
HEALTHCARE, AND ELEKTA
ARE DOING TODAY
Philips
Philips is utilizing AI in many of its products today, from patient positioning in an MRI machine to
advanced diagnostics tools for ultrasound. It is also focused on being a centralized platform for
use by hospitals and other clinical practices.
•Patient positioning: Philips' iPatient platform for the Ingenia MRI offers tools to optimize
patient positioning, as well as automating many of the processes that operators would
otherwise have to do manually (Exhibit 5). Once the operator has inputted the patient's
characteristics (via an "ExamCard"), the SmartSelect feature automatically decides which
coils and elements should be activated to produce the highest quality image (a low signal-
to-noise (SNR) ratio produces grainy images). In conjunction with Ingenia's coil design, this
prevents operators from running a scan with the wrong coil selected, and allows for a faster,
more comfortable patient setup process. Premium options such as "Premium motion-free"
reduce the likelihood that patient movement during the scan compromises image quality.
Philips claims iPatient can offer a 30% improvement in throughput for MRI systems.
EXHIBIT 5: Philips iPatient Platform
Source: Company website (with permission), Bernstein analysis
•Faster scanning times: Philips has highlighted its Compressed SENSE technology as an
important step forward for the Philips MRI portfolio. Compressed SENSE reduces scan times
across several different MRI modalities (Exhibit 6). Customers can pay more for Compressed
SENSE as part of their system purchase, or the software is also available on a subscription
basis (i.e., SaaS).
BERNSTEIN
EU MEDTECH: AI IN MEDICAL IMAGING 169
EXHIBIT 6: Compressed SENSE can reduce scanning times across a range of target areas
Source: Philips CMD 2018, DI Zoom Session (slide 7), Bernstein analysis
•IntelliSpace Discovery: Philips launched IntelliSpace Discovery in November 2018, an open
platform that enables researchers to develop and deploy AI in radiology. It provides a platform
for researchers to develop AI and manage data; this embeds Philips into the flow of data and
likely gives it some foresight of future innovation and potential disruptive technology.
•HealthSuite: HealthSuite is a cloud-based open architecture platform that allows devices
and data from a variety of settings to be integrated. This includes in-hospital monitors, but
also sensors and devices in the home. In addition to Philips-developed programs, third-party
programs are also available on the HealthSuite platform. Philips partnered with Salesforce to
develop HealthSuite. While it is not really AI, it offers a platform for the integration of AI.
•Ultrasound CADx: AI and data analytics are enabling easier identification of previously hard-
to-diagnose conditions such as nonalcoholic fatty liver disease. Philips has incorporated AI
into its ultrasound portfolio with features that allow doctors to quantitatively measure liver
fat. With classic grayscale ultrasound, doctors could only tell if the liver had a high degree of
fatty infiltration or if it was normal. There was no way of telling if the fatty liver disease was
mild, moderate, or severe. This new product utilizes data analytics to identify the severity of
the disease and allow for the tracking of fatty liver disease over time, improving patient care
(Exhibit 7).
EXHIBIT 7: Ultrasound identifies area of possible concern on the liver and also gives a confidence level of its
assessment
Source: Philips website
BERNSTEIN
170 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
Siemens Healthineers
Siemens Healthineers utilizes AI in its imaging offerings in multiple mays. Some are more simple,
such as programs focused on improving workflows and patient positioning. It also has more
advanced applications including its AI-Rad Companion and HyperSight technology.
•Patient positioning: Siemens Healthineers' FAST Integrated Workflow, in conjunction with
the SOMATOM CT machine's 3D infrared camera, is able to capture the patient's physical
characteristics, and then use algorithms to determine table height, whether the patient should
lie head or feet first, prone or supine, etc. Providers can pay extra for "advanced applications"
such as FAST Planning, which alters the scan range (how much of the anatomy is being
scanned) to match the patient's needs. If you believe you can be quicker than FAST Integrated
Workflow, you can play this game on Healthineers' website to test your abilities to position a
patient in a scanner.
•MRI Companion Suite: Siemens Healthineers has a Companion Suite to help with complex
MRI operations, leveraging AI to turn data into integrated expertise and tailored assistance.
The technology is available on the MAGENTOM range of MRI scanners.
•AI-Rad: AI-Rad is an AI-powered workflow solution for radiology departments. It is expanding
its use cases, but at present its applications aid with chest CT and chest X-ray interpretation
(Chest CT, Chest X-ray solutions), automatic contouring of organs at risk to support radiation
therapy planning (Organs RT), prostate gland segmentation for fusion biopsies (Prostate MR),
and brain morphometry analysis (Brain MR). The AI-Rad Companion Chest CT won two prizes
at the R&D 100 awards in 2020: (1) best analytical device; and (2) special recognition: Market
disruptor. AI-Rad Companion Organs RT won the "software/service" category at the R&D 100
2021 awards, for being a disruptor which will change the industry. Siemens Healthineers
states that 35% of all imaging procedures can use the AI-Rad companion (note: "can," not
"do"), with Healthineers aiming to bring that up to 80%.
•Syngo.via: Syngo.via is an imaging software platform for reading medical scans (Exhibit
8). Siemens Healthineers has embedded AI into the platform in several ways. This includes
intelligent analysis that speeds up workflows, a third-party app store, allowing access to a
suite of additional functions, including AI algorithms, developed by third parties, as well as a
platform for auto contouring and delineating tumors.
EXHIBIT 8: Healthineers' Syngo.via RT Image Suite
Source: Company website (with permission), Bernstein analysis
BERNSTEIN
EU MEDTECH: AI IN MEDICAL IMAGING 171
•Digital ecosystem: The company is building what it calls a digital ecosystem. A cloud-based
platform (based on Microsoft Azure Cloud), it will integrate Siemens Healthineers devices and
software programs, as well as third-party apps. The apps will be offered via an online store.
It is not currently commercially available.
GE Healthcare
Based on NVIDIA hardware, GE's Applied Analytics platform allows healthcare providers to use
AI to analyze the vast streams of patient data they generate every day. The insights gained are
used to improve clinical and operational performance. The platform includes numerous purpose-
built applications, including a number centered on medical imaging. This includes examples such
as workflow optimizing software and programs to segment and assess liver lesions. Exhibit 9
shows a list of GE Healthcare's activities in imaging AI.
EXHIBIT 9: GE Healthcare's activities in Imaging AI
GE Activities in Imaging AI Description
Centricity Universal Viewer Brings together patient data from multiple sources i.e. notes, EHR, Imaging
Hepatic VCAR Automates segmentation of liver and liver lesions
Interventional Liver ASSIST Helps to identify vessels feeding live tumours
MR Intelligent SAR Uses machine learning to identify head and shoulders
PET VCAR Streamlines assessment of metabolic data
Radiology Operations
Effectiveness for RIS/PACS
Helps to identify and prioritise opportunities for process improvements
X-Ray Quality Application Enables targeted training of staff to improve efficiencies
Venue - Shock Toolkit Automated tools to allow quick ultrasound whilst triaging patients
Imaging Insights
Summarises machine utilisation and dose to allow workflow management and to
drive efficiencies
Mammography Insights
Summarises machine utilisation and dose to allow workflow management and to
drive efficiencies
Tube Watch Remotely monitors CT machines and predicts failures
DoseWatch Automatically collects and analyses radiation exposure
VesselIQ Xpress Tools to analyse 3D angiographic data
Source: Company website, Bernstein analysis
VALUATION METHODOLOGY See Disclosure Appendix for valuation methodology.
EXHIBIT 10: Ratings and target prices
25-May-23 Target
Ticker Rating Currency Closing Price Price
SHL GR O EUR 52.78 60.50
PHIA NA M EUR 17.82 17.00
PHG M USD 19.13 18.00
MSCI Europe 1,848.02
S&P 500 4,151.28
Source: Bloomberg, Bernstein analysis and estimates
RISKS See Disclosure Appendix for risks.
Lisa Bedell Clive lisa.clive@bernstein.com +44 207 170 5052
Jonathon Unwin jonathon.unwin@bernstein.com +44 207 959 4571
BERNSTEIN
172 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
US HEALTHCARE SERVICES: WILL
ROBOTS REPLACE DOCTORS?
HIGHLIGHTS •Evolution of healthcare innovation focus: (1) access to insurance, (2) affordability,
and (3) access to care. From 1960 to 2010, the primary focus was on financing healthcare
costs with the creation of Medicare and Medicaid, the disruption of health management
organizations (HMOs) and managed care and, ultimately, reform through the Affordable Care
Act (ACA). With access to health insurance largely accomplished, the core issue of affordability
is the current primary focus for US healthcare. Approaches have evolved from attempting to
align consumer interests through financial exposure to now focusing on aligning healthcare
delivery incentives through adoption of VBC. We see the coming disruption of AI and digital
care as addressing the emerging issue of access to care, influencing both quality and cost. We
think AI and digital care can challenge the assumption that doctor supply is limited, leading to
lower unit costs and increased access.
•What are digital care and AI-based care? We think AI and digital care represent three
categories: (1) digital care — to improve care delivery efficiency and effectiveness to
better utilize existing doctor and nurse resources, and to greatly increase data inputs, which
could allow for more precise and personalized care; (2) digital-first health plans — which
leverage these efficiencies to lower costs to enrollees; and (2) care automation — we think
digital care and AI can ultimately replace doctors and nurses in certain circumstance, which
could greatly expand availability of clinical resources.
•The most controversial and disruptive of these concepts is care automation. We think
digital care and AI will ultimately lead to full substitution of certain clinicians and doctors,
greatly increasing clinician supply and reducing unit costs. We expect care automation to
include: (1) aids to allow human clinicians to practice at the top of their license; (2) tools which
perform initial elements of care delivery, which are then assigned to a human clinician; (3) tools
which independently perform a limited, lower-risk clinical function to completion; (4) a tool/
robot performing a clinical procedure with the outcome reviewed by a human clinician; and
(5) ultimately fully independent care delivery by the AI enabled robot, a full replacement for
a doctor.
INVESTMENT IMPLICATIONS In general, this disruption should lead to more units at lower unit costs, with reduced labor costs.
Longer term, we see the digital care and AI disruption benefiting those heavily dependent upon
clinical labor (e.g., hospitals), as well as those that benefit from frequency of touch (e.g., VBC),
while negatively impacting those benefiting from clinical labor scarcity (fee for service clinical
practices) and smaller entities unable to leverage digital/AI capabilities. We see this as modestly
positive for MCOs in that premium inflation would be reduced, which lowers pricing and policy
risk, although slowing long-term premium growth rates. This disruption would likely be most
positive for HCA Healthcare, UnitedHealth Group, and CVS Health Corp.
BERNSTEIN
US HEALTHCARE SERVICES: WILL ROBOTS REPLACE DOCTORS? 173
EVOLUTION OF US
HEALTHCARE SYSTEM
PRIORITIES
Evolution of healthcare innovation focus: (1) access to insurance, (2) affordability, and (3)
access to care. The challenge to the evolution of the US healthcare model has been affordability.
During 1980-2010, the primary focus was on financing healthcare costs through access to
health insurance. This period saw the development of HMOs and managed care, introduction
of managed care for the Medicare and Medicaid government programs, and reform to health
insurance access through the ACA. With health insurance largely accomplished, the core issue
of affordability is now the primary focus. The health insurance sector has evolved its approach
to addressing this issue: during 2000-20, the major focus was on attempting to align consumer
interests through financial exposure (e.g., high deductible health plans), and recognizing the
limitations of this approach, the adoption of VBC.
We see AI and digital care as enabling a coming disruption as access to care is addressed, rather
than just access to insurance. The current supply limit of doctors is assumed to be a given, and
then most approaches such as high deductible healthcare and VBC work around this issue, at
times exacerbating the issue. We think AI and digital care can challenge the assumption that
doctor supply is limited. AI and digital care can greatly expand data inputs, which can lead to
greater personalization of care, and care automation can potentially expand access.
Access to insurance era (1980-2010). As medical costs rose during 1980-2010, the primary
focus was on financing healthcare costs through access to health insurance. Medical costs had
been escalating at a higher rate than inflation. In response to these increasing costs, the focus
was on financing these higher costs through insurance and insurance-related innovations. This
phase kicked off post World War II with the growth in employer health insurance and, ultimately,
the ACA. This led the US to move from virtually no health insurance to today's 90%+ of citizens
with health insurance (Exhibit 1).
To enable cost-efficient growth in health insurance coverage, HMOs and, subsequently, MCOs
developed to achieve more cost-efficient health insurance. HMOs and managed care growth
accelerated starting in the 1980s in the employer health insurance market, and then in the 2000s
saw tremendous growth in the government health insurance market. The ACA further enhanced
access to health insurance through consumer protections, including elimination of the ability to
refuse coverage due to preexisting conditions.
BERNSTEIN
174 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
EXHIBIT 1: Growth in the proportion of US population covered by health insurance has been approximately flat
since 1975
9%
23%
51%
61%
68%
73%
85% 89% 89% 87% 85% 86% 85% 86% 84%
91% 90%
0%
10%
20%
30%
40%
50%
60%
70%
80%
90%
100%
1940 1945 1950 1955 1960 1965 1970 1975 1980 1985 1990 1995 2000 2005 2010 2015 2020
% of U.S. Population Covered by Health Insurance
Source: CMS, company reports, Bernstein analysis and estimates
Focus on affordability. While health insurance coverage has broadly expanded to cover 90%
+ of the US population, the underlying issue which caused the need for health insurance —
affordability — remains. From the pre-health insurance period (1940), healthcare costs grew
from 3.9% of GDP to 19.7% of GDP, reflecting this underlying growth in affordability (Exhibit
3). We have seen a series of efforts to address these fast growing costs — managed care, high
deductible plans, and VBC.
•Introduction of managed care (1980s-2000s): Managed care was introduced in the 1980s
to address the unmanaged costs of healthcare and health insurance. The main efforts by
manged care were trading price for volume by creating approved networks of doctors
and hospitals in exchange for greater discounts off list prices. In addition, managed care
attempted to steer volume and modify consumer behaviors by charging differentiated co-
pays for different services (e.g., US$5 co-pay for a PCP doctor visit and US$200 for
an emergency room visit). This approach worked better than traditional indemnity health
insurance, frequently providing 15% savings to the unmanaged approach. But even with this
model, cost inflation remained well above CPI levels (Exhibit 2).
•High deductible health plans (2000s-10s): In the 2000s, managed care plans attempted to
slow the rate of costs by shifting increased financial responsibility to consumers. The intent
was to make consumers aware of costs for procedures and cause them to shop for services.
This ultimately morphed into high deductible health plans, which have a high cost to the
consumer for non-catastrophic healthcare costs, coupled with a healthcare bank account
(e.g., HRA) so the consumer sees costs coming out of their account.
•VBC (2020s-30s+): Managed care placed responsibility on the insurance company (MCO)
to negotiate lower unit prices, and high deductible health plans placed the responsibility on
the consumer to shop and compare prices. VBC places the responsibility on the primary care
doctor to steer procedures to lower cost settings and to proactively address gaps in care to
stabilize chronics, with the intent of avoiding hospital admissions. This approach is still in early
stages, with only ~15% of the market reimbursed in a fully capitated model today. We expect
BERNSTEIN
US HEALTHCARE SERVICES: WILL ROBOTS REPLACE DOCTORS? 175
this model to continue to evolve from risk-bearing primary care delivered at practices to risk-
bearing multispecialty group care delivered at convenient locations, home and virtually.
EXHIBIT 2: Medical costs have inflated more than 2x the
CPI since 1980
Indexed to 100
Source: CMS, Bernstein analysis
EXHIBIT 3: National health expenditure has grown
significantly in terms of percentage share of US GDP
Source: CMS, Bernstein analysis
Access to care. We see AI and digital care as enabling a coming disruption as access to
care is addressed. The current supply limit of doctors is assumed to be a given, and then
most approaches such as high deductible healthcare and VBC work around this issue, at times
exacerbating the issue. These techniques attempt to shop for a lower-cost setting for a unit
of care (e.g., urgent care versus emergency room) and to reward the reduction of the number
of units, both appropriately by proactively caring for patients before a hospital admission is
necessary, and inappropriately by having consumers not have the financial ability to fund more
expensive care (self-rationing).
We expect AI and digital care can challenge the assumption that doctor supply is limited. We will
outline this much further in the chapter, but basically expect digital care and AI can: (1) improve
care delivery efficiency to better utilize existing doctor and nurse resources; (2) improve care
delivery effectiveness by greatly increasing data inputs, which could allow for more precise and
personalized care; and (3) ultimately substitute for doctors and nurses, in certain circumstance,
which could greatly expand availability of clinical resources (Exhibit 4).
BERNSTEIN
176 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
EXHIBIT 4: Number of practicing physicians per 10k citizens in the US is growing slowly
Source: Statista, Bernstein analysis
The number of visits to physician offices in the US has stalled during the 21st century despite the
US population having grown more than 20% since 2000; in fact, the number of physician office
visits per year was roughly the same in 2000 versus 2023 trends so far.
EXHIBIT 5: Physician visits per capita have been declining over the past decade
Source: CMS, Bernstein analysis and estimates
The decline in physician office visits per capita may be attributed to a supply/demand imbalance
for physicians. US population growth and healthcare expenditures have significantly outpaced
medical school matriculants since the 1980s (Exhibit 6).
BERNSTEIN
US HEALTHCARE SERVICES: WILL ROBOTS REPLACE DOCTORS? 177
EXHIBIT 6: While physician growth has waned, healthcare expenditure has continued to rise
Source: Niskanen Center
WHAT WOULD DIGITAL CARE
AND AI LOOK LIKE?
We think AI and digital care represent three categories: (1) digital care to improve care delivery
efficiency and effectiveness; (2) digital-first health plans; and (3) care automation.
Improved care delivery efficiency and effectiveness
Digital healthcare and AI are improving care delivery efficiency and effectiveness to
better utilize existing doctor and nurse resources, and to greatly increase data inputs, which
could allow for more precise and personalized care. This overlaps with the early stages of the
care automation framework we present later in this chapter.
What would this digital health improvement in efficiency of care delivery look like? An
app gives preliminary diagnosis to a patient and connects them to a virtual PCP; the virtual
PCP, a nurse practitioner, uses AI to ensure correct use of evidence-based medicine and refers
the patient to a specialist; the specialist uses imaging, genetic testing, and AI to diagnose and
recommend the preferred treatment (personalized medicine); inpatient surgery is performed and
the hospital achieves more frequent care during the stay as AI allows licensed nurse practitioners
to perform more functions, alleviating some of the nurse supply constraints; the patient is
discharged to the home with remote monitoring (some digital and some human) to assist in
frequent clinical touches.
BERNSTEIN
178 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
Examples of digital-first roles. We initially identify five categories of digital-first care delivery:
(1) acute/urgent virtual care delivery (telephonic or with video); (2) primary care, including
VBC practices, which are digital-first; (3) behavioral healthcare delivered through phone and/
or virtual; (4) integrated virtual consultation and dispensing for a specific condition; and (5)
asynchronous care delivery leveraging text with or without video/audio.
Value creation. Increased efficiency of access to care, which can reduce the unit cost of care
and can increase adherence to care. This should be valuable for those with chronic conditions,
where costs could be impacted through keeping individuals stabilized. This could also increase
satisfaction for more convenience-oriented individuals, such as younger consumers and those
with younger families.
Digital-first health plans
The more common digital-first health plans provide either a digital app to direct the consumer or
a digital-first practice as an initial access point for care. This digital assistant is intended to steer
patients and assist in the effective (meaning lower cost and higher outcome) management of
care. These products will likely evolve to increase the focus on management of chronic conditions
and would logically incorporate remote monitoring and other digital management tools.
Value adds from this approach include: (1) steerage benefits from concentrating volume for lower
unit costs and directing patients to lower cost settings (e.g., virtual versus emergency room or
outpatient MRI versus the one at a hospital); (2) leveraging evidence-based medicine to avoid
unnecessary care; and (3) better management of chronic conditions, including faster responses
to issues to avoid inpatient admissions.
One possible example of this future state: A consumer could buy a digital-first health plan with
its own digital-first primary care practice. Assuming this consumer had a chronic condition, she
could use a digital remote monitoring device and app to keep herself and her digital physician
up to date on her condition. An integrated care management app would provide guidance and
motivation for a chronic condition program, along with peer support through a social network. If
her diagnostics indicate it, she could have a virtual visit with the specialist. The patient could go
to a digital enablement location to conduct a more thorough exam in a virtual consultation room
with the nurse, and have blood drawn. If drugs were prescribed or dosage modified, the patient
would get advice and guidance from their digital concierge app, which would find the lowest-
cost drugs and deliver them to the patient.
Such a digital-first health plan would have cost of care advantages in line with a narrow network,
with the potential for improved utilization and outcomes depending on engagement through
digital means and/or leveraging a digital-first VBC approach. In addition, operating expenses
may have a potential cost advantage once fully scaled.
How does digital-first fit into a disrupted ecosystem, and what populations will it target?
We anticipate digital-first will be more applicable for employer and individual populations. We see
convenience and efficiency, coupled with cost savings from direction as being major elements
of digital-first. Over time, elements of digital-first could spread to the more expensive, higher
utilization segments of senior citizens with chronic conditions.
This is consistent with our care delivery segmentation, which focuses on three major new models:
(1) digital-first care, which provides easy access and greater efficiency for providers; (2) VBC
BERNSTEIN
US HEALTHCARE SERVICES: WILL ROBOTS REPLACE DOCTORS? 179
utilizing virtual, home, and convenient locations to improve downstream costs and better manage
chronic conditions; and (2) home-based care, coupled with remote monitoring and virtual, to
better manage high-cost frail elderly patients.
Care automation
The most controversial and disruptive of these concepts is care automation. We define care
automation as the extension and substitution of human labor by machines. We expect that
digital care and AI will ultimately lead to full substitution of certain clinicians and doctors, greatly
increasing clinician supply and reducing unit costs. While guardrails and ethical considerations
are important steps in this disruption, within this chapter, we focus on what this may look like in
the future, and the implications for business models.
In the near term, we could see digital apps driven by AI performing initial diagnoses and routing
people to the appropriate virtual or in-person human physician. We expect advances in remote
monitoring and wired environments will allow hospitals to use fewer nurses per patient and
lower-licensed nurses as well. Longer term, we can envision digital care, including robots
powered by AI, replacing doctors and nurses for a variety of functions. The likely parameters
will be variability of diagnosis, agreement on treatment approaches, and consequences
of the clinical situation. Wait times for a new appointment with a physician is 26 days, per a
Merritt Hawkins 2022 survey. If one thinks of this as the number of clinical encounters (demand
units) demanded by consumers (patients) and the supply of clinician time slots (supply units), we
likely are currently at a supply demand imbalance.
Bernstein's care automation framework
The advancement of AI and digital health is likely to initially focus on capabilities that improve
care delivery efficiency and effectiveness. We expect the evolution of these types of digital care
and AI will grow from aids to support humans to clinicians operating under human supervision,
to ultimately fully autonomous clinicians.
1. Human clinician aids. Aids that support human clinicians and allow clinicians to practice at
the top of their license. This will likely involve automation of administrative functions such as
patient questionnaires to be provided to the doctor and for record purposes. These systems
can direct clinicians to more accurate use of evidence-based medicine. An example of this
is a doctor today practicing in a wired environment who is able to compare diagnoses and
treatment recommendations with those suggested by AI tools. Aspects of this evolution are
in place today, but we would expect this to extend as it allows lower-licensed clinicians to
perform more functions, such as remote monitoring, which helps leverage nurse resources
better, and virtual care, which allows for greater efficiency for the patient against incidents
where a patient has to spend time to travel to the doctor's office/hospital. Machine learning
for developing evidence-based medicine, particularly more personalized treatments, would
fit into this category. Likewise, AI for drug development would also fit into this category.
2. Digital care to extend human clinicians. Tools that perform initial elements of care delivery
such as preliminary diagnosis, which is then served up to a human clinician. Examples of this
include digital diagnoses apps powered by AI that perform initial diagnoses and then direct
patients to a human clinician, at times through virtual means. This step of the evolution typically
involves the digital app or machine interacting with the patient first, and then transferring to
BERNSTEIN
180 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
the human clinician. In contrast, the first step (human clinician aid) is intended to be in the
background, enabling greater efficiency and human leverage.
3. Specific function digital tools. Tools that independently perform a limited, lower-risk clinical
function to completion (perhaps involving prescription) within the scope of options selected by
the human clinician. More immediate examples of this would be vision tests, where a machine
could perform the test and prescribe a corrective device, which would then be approved by
a human clinician. In this example, it could be an optometrist, or this might allow opticians
to write prescriptions for glasses. Depending on the type of clinical task, a human would
approve the result or provide a narrow range of options within which the machine would
work. It would seem that dispensing of traditional oral medications (non-narcotic) would also
fit within this category. We have looked at smaller companies that have created automated
remote dispensing machines, kind of like vending machines. These are stocked with a limited
range of medications, and video capability is incorporated to allow both visual inspection of
the pills and a video consultation with the patient.
4. "Junior robot doctor" with clinician oversight. The tool/robot performs clinical procedures
consistently within evidence-based medicine, with outcome reviewed by a human clinician
(this likely involves some human-like emotions and connection to the patient). In this category,
the physician is the robot, with oversight by a physician. This would be sort of like a resident
or physician in training. Categories most likely to be automated in this manner might, for
example, include annual physicals and certain functions of radiology. In both these cases, the
automation would likely perform the function in routine situations and escalate to the senior
human physician for complexities and more consequential situations.
5. "Full robot doctor." Fully independent care delivery by an automated tool, a full replacement for
a doctor. This is the most futuristic category and is more likely for a specific clinical category
or set of roles within a clinical category, as opposed to a full replacement of all physicians.
However, it would seem reasonable with machine learning, AI, and robotic advances to
perform physical functions and perform sensory functions (e.g., touch) that this could be
accomplished. This will warrant more research as we develop our 2023 Disruptor list.
One interesting aspect of research into this field is the cautious way in which researchers have
broached these subjects. Almost all research papers in this subject state that AI and robots will
not replace (fill in the relevant clinician type). This may create a bit of an industry blind spot to the
technological innovation that will allow this replacement to occur. At the same time, surveys of
physicians indicate that there is room for automation in their roles. A recent Stanford Medicine
survey found physicians expect that one-third of their tasks will be automated in 20 years.
Also, the pace of AI has been accelerating, with machine learning-enabled FDA-approved
medical devices a sharp example of this acceleration. Over the past five years, 448 algorithms
have been approved, compared with 58 the prior five years and 11 the five years before that.
Examples of AI/ML-enabled medical devices include: (1) smart stethoscopes, which can detect
faint heart murmurs with greater accuracy, (2) automatic insulin delivery pumps, which administer
insulin boluses or shut off insulin delivery automatically based on blood glucose readings, and (3)
smart implants, such as ZBH's Persona IQ smart knee, an AI-enabled knee implant with a sensor
that captures data during surgical procedures and rehabilitation to customize implant placement
and improve surgical techniques.
BERNSTEIN
US HEALTHCARE SERVICES: WILL ROBOTS REPLACE DOCTORS? 181
Examples of emerging digital health and AI companies
In our 2022 Disruptor list of the most disruptive emerging private companies in healthcare
services, we saw the major disruptive themes were VBC and reinventing care delivery, along with
an acceleration in digital-first health plans and AI. Below are brief descriptions of some of these
disruptive AI and digital care companies.
K Health (#7 on Bernstein's 2022 Disruptor 25 list) is a digital-first healthcare provider that uses
AI and digital care delivery to intake and initially diagnose patients. The current offerings then can
connect to virtual care delivery for primary care.
viz.ai (#8 on Bernstein's 2022 Disruptor 25 list) is an AI-based disease detection and care
coordination platform. Its technology, when integrated into clinical workflow, enables automation
in disease detection, increases diagnostics rates, and enhances care coordination workflow,
leading to better outcomes.
Maven Clinic (#11 on Bernstein's 2022 Disruptor 25 list) is the largest virtual clinic for women's
and family health. Its platform combines more than 30 provider types with individual care
navigation to support all parents and all paths to parenthood, from fertility through pregnancy,
parenting, and pediatrics.
Hinge Health (#12 on Bernstein's 2022 Disruptor 25 list) is a digital musculoskeletal (MSK)
clinic. Hinge Health reduces MSK pain, surgeries, and opioid use by pairing advanced wearable
sensors and computer vision technology with a comprehensive clinical care team of physical
therapists, physicians, and health coaches.
SWORD Health (#15 on Bernstein's 2022 Disruptor 25 list), a digital musculoskeletal (MSK)
therapy provider, offers digitally guided physical therapy, on-demand text-based support from a
physical health specialist, clinical education for members, and predictive AI to identify patients
at risk of requiring surgery.
98point6 (#17 on Bernstein's 2022 Disruptor list), a virtual primary care provider, provides on-
demand, text-based diagnosis, treatment, and consultation with primary care physicians, and
sends prescriptions to the pharmacy.
IMPLICATIONS OF DIGITAL
CARE AND AI DISRUPTION
The major disruptions we described could have the following impacts:
Digital care efficiency and effectiveness would reduce labor costs and allow for higher
volumes. An example would be a hospital floor using remote monitoring to reduce nurses/
patient, and digital aids that allow lower-licensed or less-experienced clinicians to perform more
complex tasks. Digital care delivery comparably improves patient access and time inefficiencies,
which would reduce labor cost per patient.
Care automation would have the impact of driving down unit costs of any clinical services it could
replicate, while also likely reducing ineffective care. The impact would be to reduce labor costs
for targeted clinical services. This reduction in cost, coupled with increase in access, would allow
for a greatly increased number of services.
Our basic interpretation is that we would see a significant increase in access, and that unit costs
drop, while the number of units rises from these types of innovations. We see digital care and AI
BERNSTEIN
182 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
disruption benefiting in long term sectors that: (1) are heavily dependent upon clinical labor, and
(2) benefit from increased volumes.
This would be a positive for digital care and AI disruptive companies that create, own, and
operate digital care delivery, digital care enablement, and care automation businesses.
Today, the landscape is piecemeal with early stage efforts to address this opportunity. We have
identified some of these emerging companies in our 2022 Bernstein Disruptor 25 list including
viz.ai, Scipher Medicine, ZephyrAI, KHealth, 98point6, Hinge Health, Sword, and Maven Health
(all private companies).
For hospitals, this could imply an ability to further stretch labor — very important as
labor-related costs represent around half of a company such as HCA's total expenses.
Additionally, these advances would not change the need for settings to perform procedures
(e.g., surgeries). This would be modestly offset by reductions in emergency room utilization, likely
increasing the importance of owning these digital care capabilities and/or practices as entry
points for the hospital.
We would expect this to be moderately positive for VBC companies, as they would benefit
from an increased "catch radius" — the improved ability to maintain connection with their
patients as they employ digital care and care automation to supplement the less value-added
services their physicians provide. Importantly, VBC companies are not fee-for-service practices,
which will be affected from the competition and reduction in unit cost for certain clinical services.
Rather, they are taking risk on total costs, and so will benefit from the improved total cost
economics.
Likewise, MCOs should modestly benefit from this disruption as they see reductions in medical
cost inflation from reduced unit costs. This will be somewhat offset by reduced willingness to be
exposed to traditional managed care techniques such as prior authorizations in an environment
where unit costs are much lower.
Among our coverage companies, we see HCA Healthcare, UnitedHealth Group, and CVS
Health as the leading beneficiaries of the rise of digital care and AI. HCA Healthcare is the
most exposed to labor costs and would directly benefit in its current model. UnitedHealth Group
would benefit moderately in its OptumHealth and MCO businesses, and may expand into these
emerging digital care/AI businesses directly. CVS Health likewise would benefit in its OSH/care
delivery businesses and MCO business, along with participating in a world of lower labor costs
and increased unit volume (given its exposure on the pharmacy/healthcare retail side). Elevance,
Humana, and Cigna would also be able to participate through their MCO and VBC businesses.
Will expanded access to care result in lower spending?
We have posited that healthcare is an infinite good, so we do not see long-term reductions
in percentage of spend on healthcare. Rather, we expect that consumption would greatly
increase. We see the number of units increasing, while cost per unit would greatly decrease.
The importance of cost containment-only functions would decrease, while the importance of
integrated personalized care delivery capabilities would increase.
Is 20% of GDP peak healthcare spending? We are occasionally asked by portfolio managers
if healthcare is an unattractive investment sector, given healthcare as a percentage of GDP is
so high at nearly 20% and therefore may not be able to grow further. This line of questioning
BERNSTEIN
US HEALTHCARE SERVICES: WILL ROBOTS REPLACE DOCTORS? 183
normally focuses on natural constraints on resource allocation and assumes that US spending
must hit a limit, as other major countries only spend 9-12% of GDP on healthcare.
Some of this is likely impacted by the US's lower social spending. While US healthcare
expenditure is the highest worldwide, its social spending lags many developed nations. With
social spending and healthcare spending combined, US spending is "middle of the pack" relative
to OECD nations.
The concept of a finite good and an infinite good. We believe individuals have an infinite
appetite for health and life, as opposed to a finite appetite for food, housing, or other necessities
(we won't get into the more difficult debate between life and an iPhone). The conceptual
framework for this is that appetites for food, water, shelter, and clothing must be met, but are
finite. We recognize that more expensive food or housing will be purchased as a society's wealth
rises, but we have long since passed the point where an additional dollar of income is spent on
food or one of these other finite goods. In fact, we have seen food as a percentage of GDP decline
from 23% in 1935 to 8% today.
We perceive healthcare to be an infinite good, with infinite appetite for life and quality of life. For
healthcare, we have seen a consistent growth in spending over the past 50+ years, along with
increased utility in terms of life span and quality of life. From 1950 to present, we have seen an
average increase in life expectancy of 1.75 years every decade.
We expect a downward shift to the upward sloping line of healthcare spending. We
believe there is waste and inefficiency in the US system, which could allow for reductions of
15-20%+ of current spending. The major drivers of this cost savings could be combinations
of improved alignment of interests (VBC), shifts to a greater proportion of spending being
reimbursed at government reimbursement levels, and policy efforts on drugs and other high-
growth subsectors. This underpins our view that the major disruptions coming to healthcare
will be an increased shift to government managed care, a shift to VBC, and a likely erosion of
employer coverage shifting to government programs.
While healthcare spending as a percentage of GDP is high, we see it rising over time.
What would be the impact of these cost savings on healthcare spend? A 20% reduction
in spending would lower US healthcare spending to ~14.5%, and this would likely impact
healthcare spending over the course of more than a decade, not as a one-time reset to costs.
Offsetting this cost saving is our history of a consistent increase of ~2.5% points of increase in
spending (portion of GDP) per decade. Therefore, we would expect healthcare spending in the
US to be well above 25% in 50 years. This is the long-term bull case for healthcare spending.
BERNSTEIN
184 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
VALUATION METHODOLOGY See Disclosure Appendix for valuation methodology.
EXHIBIT 7: Ratings and target prices
25-May-2023 Target
Ticker Rating Currency Closing Price Price
CNC O USD 63.43 92.00
CI M USD 247.07 293.00
CVS M USD 67.67 93.00
ELV O USD 450.08 620.00
HCA O USD 264.09 345.00
HUM M USD 500.07 568.00
UNH M USD 477.70 609.00
SPX 4,151.28
Source: Bloomberg, Bernstein analysis and estimates
RISKS See Disclosure Appendix for risks.
Lance Wilkes lance.wilkes@bernstein.com +1 212 407 5826
Amir Farahani amir.farahani@bernstein.com +1 212 756 4136
William Robbins william.robbins@bernstein.com +1 212 969 2487
BERNSTEIN
US HEALTHCARE SERVICES: WILL ROBOTS REPLACE DOCTORS? 185
BERNSTEIN
186 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
BERNSTEIN
Technology
TECHNOLOGY 187
BERNSTEIN
188 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
US E-COMMERCE: OPPORTUNITIES AND
RISKS FROM AI
HIGHLIGHTS •The companies we cover deal in tremendous amounts of data. Machine learning
algorithms are already being utilized to harness that data into better recommendations and
more efficient networks. As these platforms grow, so do their underlying data sets, allowing
for better results. Better search on Etsy using words (and now images) to help you sort through
the long tail, one-day delivery through Amazon Prime, and now models such as upfront fares
at Uber that make driver pay more dynamic. Data reinforces the moat.
•The new wave of generative AI captures the imagination. It is not a stretch to draw up an
outlook where generative AI chatbots transform search across e-commerce marketplaces,
offering personalized and well-curated recommendations. Shopify's (not covered) OpenAI API
is an early taste of what the future of shopping online could look like, as are integrations from
Instacart (not listed) and Klarna (private). Having a personal assistant chatbot that can help you
sift through the limitless online shelf is certainly appealing — personalization at scale. When
search parameters are complex and undefined, and more discovery is required, the value of
these AI models can shine. On the flip side, there is also the disintermediation risk to consider
if these chatbots become consumer apps in their own right. For marketplaces, the value of
offering differentiated merchandise and having strong direct traffic could be amplified.
•The productivity uplift seems imminent. Regardless of how consumers interact with AI
chatbots or change their search behavior, the ability of generative AI to automate content
creation is perhaps the most tangible near-term benefit. Merchants should benefit from
automating processes; for example, writing product descriptions and interpreting sales data.
Knowledge workers should also see a productivity uplift, which will free up time to focus
on higher ROI efforts or allow these companies to run leaner. We see trickle-down benefits
from AI investments by big tech as: (1) cloud vendors; (2) productivity suite providers (e.g.,
Office365 or Workspace); and (3) even the primary digital ad channels for online marketplaces
(Performance Max at Google and Advantage+ at Meta). In other words, R&D, G&A, and
marketing dollars should go further in the new world.
INVESTMENT IMPLICATIONS The new wave of generative AI certainly captures the imagination and raises questions of how e-
commerce may be further transformed. We approach it from the perspective of: (1) consumers;
(2) merchants; and (3) the platforms themselves. As always with disruptive technology, we run
the risk of thinking through the lens of today's constructs. But it is likely that use cases and
applications emerge that are hard for us to envision today, that may end up feeling intuitive in
hindsight. We rate Etsy Outperform, eBay Market-Perform, and Wayfair Underperform.
BERNSTEIN
US E-COMMERCE: OPPORTUNITIES AND RISKS FROM AI 189
HOW AI SHAPES THE
CUSTOMER EXPERIENCE
TODAY
AI is already being used in a number of ways to enhance the shopping experience across e-
commerce marketplaces:
•Improving search: The beauty of e-commerce marketplaces is that they present consumers
with a tremendous amount of choice. But when platforms have billions of product listings, it
becomes incredibly important that search results are accurate, effective, and personalized,
and that the algorithms are able to understand more natural/simple search prompts. These
models continue to improve. In 2021, Etsy rolled out its new model called "XWalk," which uses
2.7 billion data points or 11x more than the 240 million data points used by its prior model to
help improve the relevancy of its search results (Exhibit 1 and Exhibit 2). Another example is
the use of image search, where customers can take photos of products in their surroundings
to find similar listings on the platform (Exhibit 3). eBay recently talked about training its image
search algorithms on over 10 billion images on its platform across 1.7 billion listings.
•Improving product listings: Online shopping involves trust. And trust is hard to build when
product listings look unprofessional, which can often be the case when dealing with SMBs that
have limited resources or casual sellers. To combat this, marketplaces have to invest in listing
technology (and education of best practices). For example, eBay rolled out 3D visualization
tools that sellers can use to produce more dynamic listing images.
•Imagining products in your space: One of the biggest challenges of shopping online
sometimes is the inability to imagine how a particular product might look on you or in your
personal space. Marketplaces such as Wayfair and Amazon now offer consumers augmented
reality tools. This enables customers to visualize products, such as a desk or a chair, in their
personal living spaces, thereby reducing the risk of (expensive) returns (Exhibit 4).
•Recommendation engines: Amazon popularized the "frequently bought together" feature
to drive incremental add-ons to the average basket — in a data-driven, predictive fashion
(Exhibit 5). Stitch Fix is another example of a business built around AI-driven predictive
recommendations. The business curates a personalized box or a "Fix" for clients, which
contains five recommended items that it anticipates the customer would like. To do so, it
utilizes data from a style questionnaire that it asks customers, studying keep versus return
behavior on orders as well as online activity, textual feedback, and other visual cues it is fed
(e.g., a customer's Pinterest board).
•Warehouse/logistics management: There is perhaps no better example here than Amazon,
which uses an incredible amount of data to predict demand trends and forward-position
inventory to optimize delivery routes such that it can consistently meet two-day shipping
windows (and now even one-day) in a cost-effective manner. Not to mention the automation
and robotics that keep its warehouses running. Wayfair, which runs its own fulfillment and
delivery network as well, has had to make similar investments over the years (Exhibit 6).
•Content moderation and fraud detection: Algorithms are used to ensure product listings
and customer reviews are genuine and do not violate any marketplace rules, and that
fraudulent activity is prevented. In 2022, Etsy invested more than US$50Mn in its Trust and
Safety efforts in order to detect and remove non-compliant listings in real time. ~95% of the
violations flagged on the marketplace were generated by its internal automated systems. The
net result is a higher-quality and safer experience for the customer.
BERNSTEIN
190 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
EXHIBIT 1: Etsy's improved search algorithm uses 11x the data than its previous model to drive more relevant
results
Source: Company reports
EXHIBIT 2: Etsy's discussion of search and the opportunities for generative AI
Etsy CEO, Josh Silverman, on 4Q22 Earnings Call:
So how do you tell the search engine what you want in really new and novel ways? Xwalk helps you to process massive amounts of information and neural
models help you to make sense of what people mean not what they say. And in fact, the neural models we're using use the same kind of underlying
algorithms that ChatGPT uses. So we're already using some of the underlying technologies that you're seeing from great innovators like OpenAI and it's having
great impact...
And I would say that similarly figuring out more novel ways to talk to our search engine, to tell it what you want. So multimodal search through maybe pictures
and words is going to be very important. Show me things of the following style that would match well with this couch. And so those kinds of things I think open
up a huge opportunity for Etsy...
I'm particularly interested in is how can we get better at quality? Everyone has their own view of what great looks like and what a really good item on it Etsy
looks like. So how can we start to better anticipate? What's right for you? How can we get more personalized so that you see only the best of Etsy as according to
you? We have a great roadmap and I'm really excited about the work that the team is doing...
One other thing I just want to say is every time that search team develops a new model, a new algorithm, we look at how much extra processing power has it
used? And therefore, how much extra cost has it added? And how much does it improve conversion rate in the lifetime value of the buyers? And is that profitable
to ramp up? And we only ramp up things where the benefit it adds is greater than the cost...
I think there's a lot of opportunities, one, for generative AI. One is in search no doubt. You might see it in the seller experience and making it easier for sellers
to make listings. You might see it in the member services experience and having a better opportunity to get customer support. Our developers may be able to
use it to make themselves more productive.
Etsy CEO Joshn Silverman, on 1Q22 Earnings Call:
What we talked about in the prepared remarks about Xwalk is really important and to simplify it, instead of having one objective function like leather wallet,
you can now search-- we can now optimize for multiple objectives at the same time. Leather wallet that's near me will arrive within this time and it's brown ,
and the ability for the search engine understand plain English and convert it into things that people want is just getting better and better.
Note: Bolded emphasis is ours
Source: Bloomberg, Bernstein analysis
BERNSTEIN
US E-COMMERCE: OPPORTUNITIES AND RISKS FROM AI 191
EXHIBIT 3: Visual search is being used to help shoppers
find products through use of their camera...
Source: Company reports
EXHIBIT 4: ...and augmented reality tools help consumers
visualize online merchandise in their personal space
Source: Company reports
EXHIBIT 5: Data-driven recommendations, at scale
Source: Company reports
BERNSTEIN
192 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
EXHIBIT 6: For e-commerce companies built around logistics, machine learning is critical to improving the
productivity and effectiveness of the network
Wayfair CEO, Niraj Shah, in 4Q22 Earnings Call:
[Another] initiative is leveraging our enormous database of orders to understand the relative rate of damage in other incident risks for items based on delivering
location and the factor that into the amount of exposure that items receive on our platform, lowering cost and also improving the customer experience.
Wayfair CTO, Fiona Tan in 1Q22 Earnings Call:
One last area to highlight is how we are applying machine learning into our logistics network. We are dynamically curating customer search to show products
that not only fit the parameters of the search, but are also located in fulfillment centers closest to them. Doing so can significantly cut the distance products
travel, lowering costs and prices, reducing damage, and driving higher conversion. We also have strong teams thinking about what it will need to shop for the
home years from now.
Source: Bloomberg, company reports, Bernstein analysis
HOW GENERATIVE AI COULD
CHANGE THE WAY WE SHOP —
BRING ON THE AI CHATBOTS
The new generative AI wave certainly captures the imagination. It is not a stretch to draw
up an outlook where generative AI chatbots transform search across marketplaces, offering
personalized and well-curated recommendations. Personalization at scale is a powerful notion
that should drive better conversion rates and could accelerate e-commerce adoption.
Shopify's OpenAI API is an early taste of what the future of shopping online could look like, and
it is not the only one integrating with ChatGPT. Klarna and Instacart are two other examples
(Exhibit 7 and Exhibit 8). What's the dream? A 24/7 personal assistant chatbot that can help you
sift through the limitless online shelf and serve results that are increasingly personalized to your
needs and preferences. And, of course, these chatbots could also help with customer service if
something goes wrong with the order.
Now, watching some of these demos, we can't also help but wonder the extent to which the
current search experience is truly being improved. In instances where you know what you need or
want, you're probably just better off typing that straight into the Amazon search bar. It seems both
quicker and easier. Let's not forget that search has been getting optimized for decades now and
is quite effective. There is a reason that 63% of US consumers1 start their product searches on
Amazon and that a very low percentage (estimated 8%2) of online shoppers regularly use more
cutting-edge technology such as visual search.
Nonetheless, we can see the utility of these tools when search parameters are more complex
and undefined, so a longer search and discovery process would be required. Take the Instacart
example. Say you're like me, and you are clueless in the kitchen. You could aimlessly Google for
some lunch recipes, find your way to a cooking blog and get inundated with terrible pop-up ads.
Or you could ask the ChatGPT plug-in within the Instacart app for some healthy and affordable
lunch options under US$10, generate a recipe, build a shopping cart and order the basket within
a few clicks. Sign me up!
The extent to which consumers engage with generative AI assistants/chatbots when shopping
online remains to be seen. We know just how hard it can be to change consumer behavior, and
how much time it can take. 85% of retail still happens in-store, and we've been at e-commerce for
over 25 years now. Alexa never lived up to its promise of being that easy-to-use, voice-controlled
assistant. However, this feels like a true paradigm shift in technological power with the potential
1 Jungle Scout Consumer Trends Report: Q1 2023.
2 The Insider Intelligence Ecommerce Survey conducted in February 2023 by Bizrate Insights.
BERNSTEIN
US E-COMMERCE: OPPORTUNITIES AND RISKS FROM AI 193
for these chatbots to live up to the hype of being that dynamic and effective digital assistant. If
rolled out effectively, it should help drive better conversion rates and accelerate the adoption of
e-commerce, even in categories that are hard to crack today.
IS THERE
DISINTERMEDIATION RISK? A
"WHAT IF" SCENARIO
Of course, we are looking at the technology from the lens of consumer experiences and platforms
we live with today. We have assumed so far that consumers still seek out marketplaces and use
generative AI APIs to help them discover products. However, we could see entirely new platforms
and applications built entirely around AI in ways that are hard to conceive at the moment, but
could fundamentally shift how we engage with the online world. There is a disintermediation risk
for e-commerce marketplaces if consumers gravitate toward these tools as the jumping off point
for their product search journey.
In other words, if generative AI chatbots were to become consumer apps in their own right or
foundational to the operating systems of the future, to what extent do consumers need to seek
out online marketplaces to search for products? Presumably that search process would start
one-step removed. Marketplaces could still facilitate the underlying sale, but in losing direct
traffic they would risk losing: (1) customer loyalty; (2) access to data; and (3) the ability to sell-
through additional services such as advertising.
We believe this scenario would be a bigger challenge for marketplaces that: (1) struggle with
direct traffic; and/or (2) have commoditized or overlapping merchandise because it could
increase the risk of price and feature comparisons across marketplaces. Scanning a sample of
several large online shopping destinations in the US, the breakdown of web traffic data (mobile
and desktop) suggests that only 42% of e-commerce traffic to the top sites is direct, 33% comes
from organic search, and 13% from paid search on average in our sample (Exhibit 9).
A similar threat already exists today with Google Shopping, which has not really materialized into
a structural headwind for the sector. However, the core Google Search product naturally shares
in the profit pool of e-commerce, thanks to its leading position in the product search journey
(commanding 46% of web traffic between organic and paid search). Any potential outcome that
could disintermediate traffic and the economics of a marketplace bears watching. Perhaps we
will even see tech companies resist integration with chatbot APIs to maintain the status quo.
BERNSTEIN
194 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
EXHIBIT 7: Snapshot of the Klarna and OpenAI integration
Source: Company reports
EXHIBIT 8: Instacart's OpenAI integration
Source: Company reports
BERNSTEIN
US E-COMMERCE: OPPORTUNITIES AND RISKS FROM AI 195
EXHIBIT 9: Disintermediation threat? Scanning web traffic to some of the top online shopping destinations shows
that only 42% of traffic is direct on average (this excludes mobile app engagement)
54% 52% 51% 49% 48% 47%
39% 39% 38%
34%
30% 29%
26%
35%
21%
31%
41%
25%
43%
36%
42%
26%
30%
40%
8%
5%
10% 11%
0%
19%
10%
13% 13%
25%
27%
21%
0%
10%
20%
30%
40%
50%
60%
Amazon eBay thredUP Best Buy Costco Chewy Target Etsy Walmart Rent the
Runway
Wayfair Overstock
US Desktop and Mobile Search Traffic by Direct and Search Channels (% of Total; won't add to 100% as some
traffic channels not shown)
Direct Traffic Organic Search Paid Search
Note: Amazon is covered by Mark Shmulik; Costco, Walmart, and Target by Dean Rosenblum. Best Buy, threUP, Rent the Runway, Overstock, and Chewy are not
covered by Bernstein.
Source: Similarweb, Bernstein analysis
AN ENHANCED MERCHANT
EXPERIENCE — FEWER
BARRIERS TO BUILDING A
DIGITAL BUSINESS
Marketplaces are two-sided, and it is not just about the end consumer. Platforms have to nurture
and grow the supply side of the equation to keep the marketplace healthy and improve the quality
of the service for consumers. This often requires digitally capable small businesses. The barriers
to building an online business continue to erode, and we expect that trend to continue with the
help of generative AI.
Today, platforms such as Shopify have democratized access to the tools you need to get a
digital business up and running. We think generative AI can take it a step further by helping with
content creation for the average business. This includes landing pages, blog posts, social media
posts, product descriptions and designs, and even ad creatives. As a quick experiment, we asked
ChatGPT to help us write a listing description for a pair of sneakers that we could post to eBay
(Exhibit 10). A little robotic, but not bad at all!
Beyond creating content, generative AI tools can also be integrated in marketplaces to help
merchants better analyze data around product sales and returns. Watching the Copilot demo
from Microsoft, it is incredible how these tools can be used to generate easy-to-read tables,
charts, and summaries from broad-based data sets (e.g., customer reviews) — all with a few
clicks.
BERNSTEIN
196 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
EXHIBIT 10: We asked ChatGPT to help us write a listing description; the results were pretty good
Source: ChatGPT, Bernstein analysis
PLATFORM EFFICIENCIES —
THE PRODUCTIVITY UPLIFT
SEEMS IMMINENT
Regardless of how consumers interact with AI chatbots or change their behavior around search,
the ability of generative AI to automate content creation is perhaps the most tangible benefit.
Knowledge workers across these companies should see a productivity uplift, which will either
free up time to focus on higher ROI efforts or allow these companies to run leaner.
We can see how generative AI could be used to assist workers across an organization. It goes
beyond engineers and writing code to also include non-technical roles in marketing, finance,
legal, HR, etc. By now, you have probably seen the research paper from academics at the
University of Pennsylvania that discussed how 80% of the US workforce could see 10% of their
work tasks affected by the introduction of large language models ("LLMs"). And, with access to
LLMs, about 15% of all worker tasks could be completed "significantly faster" with the same level
of quality — a percentage that massively improves to between 47% and 56% of all tasks when
coupled with software and tooling built on top of LLMs.
We also see trickle-down benefits from big tech investing in AI. The reality is big tech, namely,
Google, Amazon, Microsoft, and Meta, supports much of the underlying infrastructure for
internet companies — public cloud (AWS, GCP, and Azure), productivity tools (Workspace and
Office365), and customer acquisition (Google Search and Meta direct response advertising).
BERNSTEIN
US E-COMMERCE: OPPORTUNITIES AND RISKS FROM AI 197
Take investments from Google and Meta into Performance Max and Advantage+ as examples.
These are AI-led solutions to improving return on ad spend in a privacy-first world, and our
coverage companies should benefit from it.
R&D, G&A, and marketing dollars should go further in the new world. And for a lot of SMID-Cap
tech struggling with FCF generation, improved opex management is an opportunity to run leaner.
It is not yet clear to us how much these models could save. But we compiled a list of e-commerce
businesses and found that last year the median company in the group spent 39% of revenue on
marketing, G&A, and R&D (Exhibit 11). Illustratively, if we assume these companies can cut these
expenses by 5% to 10% on the back of AI developments, it would generate 200 basis points to
400 basis points of margin expansion, all else being equal (Exhibit 12 and Exhibit 13).
An offset to this math would be if the platforms have to take on an incremental cost burden for
utilizing AI in their day-to-day operations. Also, if it is easier for companies to do more with less,
then presumably the barriers to increased competition will also be lower under the new paradigm
— another risk to keep an eye on for incumbents.
EXHIBIT 11: E-commerce companies spend ~40% of their revenue on sales & marketing, R&D, and G&A (median)
11% 8% 6% 12% 12% 22% 28% 22% 13% 22%
6% 14% 10%
14%
16% 27%
22%
54%
4% 2%
21%
6% 21%
10%
12% 13% 51%
21%
22% 25% 27% 29% 34%
45%
56% 61%
87%
98%
0%
20%
40%
60%
80%
100%
120%
Overstock Amazon Chewy MercadoLibre Wayfair eBay Etsy Shopify Rent the
Runway
(2021)
thredUP
Select Operating Expenses as % of Revenue (FY22; R&D, G&A, S&M)
S&M R&D G&A Total
Note: Excludes cost of sales. Overstock, Chewy (R&D not disclosed separately), MercadoLibre (excludes provision for doubtful accounts), Shopify (excludes
transaction & loan losses), Rent the Runway (based on year-ended Jan 2022, excludes rental depreciation and fulfillment), and thredUP not covered. Amazon
covered by Mark Shmulik and excludes fulfillment costs (R&D used is disclosed technology and content costs). eBay excludes provision for transaction losses and
amortization, Wayfair does not disclose R&D separately (advertising cost used for S&M and consolidated SOTG&A for the rest).
Source: Bloomberg, Bernstein analysis
BERNSTEIN
198 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
EXHIBIT 12: Illustratively, if we assume generative AI tools enable these companies to cut 5% of their operating
expenses, it would drive 200 basis points of margin improvement for the median company...
22% 25% 27% 29% 34% 39% 45%
56% 61%
87%
98%
21% 24% 26% 28% 32%
37% 43%
53% 58%
82%
93%
0%
20%
40%
60%
80%
100%
120%
Overstock Amazon Chewy MercadoLibre Wayfair Median eBay Etsy Shopify Rent the
Runway
(2021)
thredUP
Select Operating Expenses % of Revenue (Illustrative scenario with FY22 expenses 5% lower)
FY22 Opex % of Revenue OpEx 5% Lower Scenario
+200bps margin
leverage
Source: Bloomberg, Bernstein analysis
EXHIBIT 13: ...and 10% operating expense savings would drive 400 basis points of margin expansion at the median
e-commerce platform in our simplistic scenario
22% 25% 27% 29% 34%
39%
45%
56% 61%
87%
98%
19% 22% 25% 26% 30% 35% 41%
50% 55%
78%
88%
0%
20%
40%
60%
80%
100%
120%
Overstock Amazon Chewy MercadoLibre Wayfair Median eBay Etsy Shopify Rent the
Runway
(2021)
thredUP
Select Operating Expenses % of Revenue (Illustrative scenario with FY22 expenses 10% Lower)
FY22 Opex % of Revenue OpEx 10% Lower Scenario
+400bps margin
leverage
Source: Bloomberg, Bernstein analysis
BERNSTEIN
US E-COMMERCE: OPPORTUNITIES AND RISKS FROM AI 199
VALUATION METHODOLOGY See Disclosure Appendix for valuation methodology.
EXHIBIT 14: Ratings and target prices
5/25/2023
Ticker Rating Currency Closing Price Target Price
EBAY M USD 43.62 50.00
ETSY O USD 86.23 120.00
W U USD 34.87 30.00
SPX USD 4,151.28
Source: Bloomberg, Bernstein analysis and estimates
RISKS See Disclosure Appendix for risks.
Nikhil Devnani nikhil.devnani@bernstein.com +1 212 969 6331
Eva Zhang eva.zhang@bernstein.com +1 212 969 1485
BERNSTEIN
200 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
SOUTH & SOUTHEAST ASIA TECH:
OPPORTUNITIES FOR AI IN CONSUMER
TECH
HIGHLIGHTS •AI tools commonly being deployed by all platforms in the ASEAN region include: (1)
Chatbots: connecting consumers to drivers/merchants through AI-powered bots and chat
technologies; (2) Personalization and recommendation engines: machine learning-powered
product recommendations and personalized search results for consumers to attain higher
buyer engagement and conversion, thereby achieving more sales; (3) Merchant analytics:
provide data-driven insights and pricing recommendations for merchants; (4) Targeted
advertising: allows merchants to deploy machine learning-powered advertising to target
consumers; and (5) Credit scoring: to assess the credit profile of consumers by analyzing
usage behaviors, spending patterns, transaction frequency, and other types of alternative
data.
•Grab has been particularly active in many aspects of business that require AI in some form:
GrabMaps, personalized recommendations, Grab Ads, and merchant tools, in addition to
broader experiments on the deployment of autonomous vehicles (AVs), etc. The heart of
GrabMaps solutions is the new data it generates from millions of orders and rides daily, with
real-time feedback from partners on road closures, address changes, etc. Grab's vast data
capabilities generate valuable insights into consumption patterns and consumer behaviors.
Deep AI and machine learning systems use these vast datasets to deliver intelligent and
personalized experiences such as predictive ride recommendations so that a ride can be
booked with one tap. It also uses machine learning and GPS data from driver-partners to
detect potentially unmapped roads and generate food merchant recommendations to ride
users. At a smaller scale, efficiency drivers such as order batching and reducing wait times at
merchants are being deployed.
•Sea Ltd, apart from personalized recommendations, has been experimenting with features
such as beauty cams and livestreaming with mixed success until now. As part of its strategy
for 2023, Shopee plans to use AI to recommend more personalized live stream content and
deals based on shopping behavior and user interests. In addition, it has built an ads tool for
merchants to capture ad revenues.
INVESTMENT IMPLICATIONS AI can help enhance the consumer experience, win share against old economy networks,
increase operational efficiency, reduce operations cost, and improve revenue streams through
appropriate use of the vast trove of data consumer tech companies collect. AI needs investments,
where Grab and Sea are leading in their respective areas. Start-ups have an additional challenge
of lack of resources and data, so the investments by leading players serve as entry barriers and
value enhancers. Accordingly, we see Sea and Grab dominating their respective verticals.
BERNSTEIN
SOUTH & SOUTHEAST ASIA TECH: OPPORTUNITIES FOR AI IN CONSUMER TECH 201
AI IN CONSUMER TECH When we talk about the implementation of AI in consumer tech, we largely include elements of
machine learning and predictive learning. The consumer tech companies' vast data and analytics
capabilities can empower various AI applications, which can eventually bring efficiency gains. We
present the various ways SEA tech companies are deploying AI to gain a share and increase the
efficiency of operations.
GRAB AT THE FOREFRONT OF
AI ADOPTION
Some key areas where Grab is implementing/piloting the use of AI/machine learning are:
Driver-side initiatives
•GrabMaps: Grab started investing in its own mapping solutions in 2017 and implemented
its mapping solution by the mid-CY22 end. Grab has built its maps using its proprietary map-
making camera (Kartacam), which feeds data into a mobile app the drivers maintain (Exhibit
1 and Exhibit 2. While the need to make its own maps arose from incomplete data and
the suboptimal routing performance of third-party mapping providers, it is now increasing
system-wide efficiency and saving costs. GrabMaps has led to a cost reduction of more
than US$10Mn for Grab and also brought additional revenue streams as it has roped
in clients such as Microsoft/Amazon for their Azure/Bing and AWS mapping needs.
•At a broader level, GrabMaps can be understood as an all-inclusive map that covers all the
undiscovered streets and lesser-known shortcuts that can help drivers navigate even the
tiniest roads that can only fit a motorbike. A pictorial representation is shown in Exhibit 3
and Exhibit 4. GrabMaps has more than 50 million addresses and points of interest (POIs)
(versus seven million in 2017), and powers more than 800 billion API calls per month,
helping drivers and passengers travel in the most time-/cost-effective way. Not just routes
and POIs, it provides rich data such as turn restrictions, speed limits, toll locations and
charges, and a 360-degree street view.
•The heart of GrabMaps solutions is the new data it generates from millions of orders and
rides daily, with real-time feedback from partners on road closures, address changes,
etc. Driver-partners contribute to maps in collecting POIs and other rich data such as
street images, street names, traffic signs, etc., for additional income. This gives GrabMaps
an edge in accuracy and coverage while being highly cost-effective. From driver-partner
allocations, ETA calculations, and route planning to cost optimization, GrabMaps is critical
to saving cost and bringing efficiency to Grab. Based on an internal study of GrabMaps
versus a leading third-party mapping provider, GrabMaps has a 4x lower error rate and 10x
lower latency. Further, the ease of finding the right POI for transport bookings improved
by 3%, while ETA accuracy improved by 1% regionally, with some countries seeing
improvements of up to 7.8% post-implementation. Grab is planning to launch application
programming interfaces (APIs) and mobile software development kits (SDKs) in 2023,
which will allow developers to enhance/build their own applications and geolocation
capabilities leveraging GrabMaps technology, such as Grab’s routing, search, traffic and
navigation features.
BERNSTEIN
202 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
EXHIBIT 1: Grab driver using Kartacam for GrabMaps
Source: Grab investor day presentation, Bernstein analysis
EXHIBIT 2: Building map data through the KartaView app,
as well as Grab driver and user superapp
Source: Grab investor day presentation, Bernstein analysis
EXHIBIT 3: Snapshot of Jakarta map with lower POIs
Source: Grab investor day presentation, Bernstein analysis
EXHIBIT 4: Growth in POIs with GrabMaps in one year
Source: Grab investor day presentation, Bernstein analysis
Grab's vast data capabilities generate valuable insights into consumption patterns and consumer
behaviors. Deep AI and machine learning systems use these datasets to deliver intelligent
and personalized experiences such as predictive ride recommendations so that a ride can be
booked with one tap. It also uses machine learning and GPS data from driver-partners to detect
potentially unmapped roads and generate food merchant recommendations to ride users (Exhibit
5).
BERNSTEIN
SOUTH & SOUTHEAST ASIA TECH: OPPORTUNITIES FOR AI IN CONSUMER TECH 203
EXHIBIT 5: AI helps in ride recommendations, precise ETA calculations, unmapped roads detection, and cross-
vertical merchant recommendations (from left to right)
Source: Grab investor day presentation, Bernstein analysis
•Just-in-time arrival: On the driver supply side, GrabFood is piloting a system that targets to
reduce their wait time (just-in-time arrival of drivers in merchant premises). To implement this,
it is deploying Bluetooth devices (currently with large merchants), which track the food prep
time, arrival time for drivers, and the exact time when they leave. This data helps improve the
accuracy of food preparation time estimation models, which is then used to ensure drivers
reach close to or slightly after the food is ready for collection. This is combined with high-
quality mapping to help drivers navigate outdoor roads and indoor paths to merchants. Grab
highlighted that it helped save 12 million minutes of driver capacity in July 2022, which
eventually helped driver-partners earn more money and reduce the dependency on
driver incentives (Exhibit 6).
BERNSTEIN
204 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
EXHIBIT 6: Grab making online and offline investments for "shorter stops" of Drivers
Source: Grab investor day presentation, Bernstein analysis
•Orders Batching: Grab is piloting batching orders by combining the orders generated from
a concentrated area such as a shopping mall and targeting less time-sensitive consumers by
allowing delivery fees to be saved by waiting longer. Many of these pilots are either new
or not even launched, but Grab is seeing encouraging results with a 22% reduction
in driver wait time in merchant premises, a 19% increase in batch rate, and an 11%
increase in the number of trips per logged hour. All these tech system advancements boil
down to reducing reliance on subsidies while delivering stable earnings/affordable services
to drivers/consumers (Exhibit 7).
EXHIBIT 7: Grab piloting large batches for orders
Source: Grab investor day presentation, Bernstein Analysis
BERNSTEIN
SOUTH & SOUTHEAST ASIA TECH: OPPORTUNITIES FOR AI IN CONSUMER TECH 205
•Driver partner superapp: Driver partner superapp has a feature called day planner for
drivers, which recommends different services for different periods of the day based on
demand trends, as the timing of demand surge in food delivery is different from ride-hailing,
grocery, and parcel delivery. The superapp also recommends shifting zones to drivers when
order density in a particular region surges; this helps drivers get more orders and earn more
(Exhibit 8 to Exhibit 10).
EXHIBIT 8: Day planner feature in
driver partner superapp
Source: Grab investor day presentation, Bernstein
analysis
EXHIBIT 9: Shift zone feature in driver
partner superapp
Source: Grab investor day presentation, Bernstein
analysis
EXHIBIT 10: Drivers can toggle
between different services within
the superapp
Source: Grab investor day presentation, Bernstein
analysis
Customer-side initiatives
•Targeted promotion: Grab has leveraged AI and machine learning to optimize different
merchants to be shown at the top for different customers, and target the promotional spend to
maximize the ROI on every promotion dollar spent. This has reduced consumer incentives
by ~7% across five core markets while sustaining GMV and user growth.
•Customized experience: Grab also uses its vast user data to provide users with a
customized/unique app version. The experiences are tailored to users' likes, dislikes, habits,
and predicted preferences to create better user experiences. This has led to a 23% increase
in conversion in a pilot study for Grab (Exhibit 11).
BERNSTEIN
206 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
EXHIBIT 11: Grab targets users based on their likes and dislikes
Source: Grab investor day presentation, Bernstein analysis
Others
•Autonomous vehicles/robots: Grab has also been the first company in the region to pilot
new technologies such as AVs for taxis and robots/AVs for food delivery. Grab was the first
ride-hailing company globally to test an AV for ride-hailing in Singapore in 2016, in partnership
with nuTonomy, which is now a part of Motional (JV between Hyundai (auto OEM company;
not covered by Bernstein) and Aptiv(auto tech)). Grab also piloted the first indoor robot runner
service in Singapore in 2021, which helped consolidate orders across several restaurants
within a mall before handing them off to delivery partners at a central collection point. Grab
also piloted an autonomous vehicle robot for food delivery service in 2022, designed to serve
guests in Singapore (Exhibit 12 to Exhibit 14).
EXHIBIT 12: Grab piloted AVs for food
delivery in Singapore in 2022, in
parternship with NCS
Source: Grab presentation, Bernstein analysis
EXHIBIT 13: nuTonomy and Grab
launched partnership to expand
public trial of autonomous car
service in Singapore
Source: Grab presentation, Bernstein analysis
EXHIBIT 14: Robot to consolidate
restaurant orders from a mall
Source: Grab presentation, Bernstein analysis
BERNSTEIN
SOUTH & SOUTHEAST ASIA TECH: OPPORTUNITIES FOR AI IN CONSUMER TECH 207
•Fraud prevention tools: Grab has reduced fraud rates related to driver identity from 20-30%
to less than 1% in the ASEAN region by implementing an AI-powered facial authentication
system for driver partners. A pictorial representation of the same is included in Exhibit 15,
where a driver is trying to bypass facial verification using face masks. Grab has also built
its credit scoring models based on a machine learning system, which empowers its lending
business (Exhibit 16). Apart from this, Grab has partnered with Microsoft to improve its natural
language processing system so that its platform supports the localized language of the
ASEAN region, which is key to its localization strategy.
EXHIBIT 15: Anti-fraud system built by Grab based on AI
Source: Grab investor day presentation, Bernstein analysis
EXHIBIT 16: Machine learning-driven credit scoring model
built by Grab
Source: Grab investor day presentation, Bernstein analysis
•Advertising: Grab has built a unified marketing platform for merchant partners to start with
self-serve ads. This is integrated with point-of-sale and helps merchants update prices and
items on the menu. The main target is to digitize the sales and order management of all
merchants, including smaller ones that have never operated online.
Specifically, on advertising, Grab is also creating more inventory and slots on its app online, and
using data for offline attribution signals such as "Did you go to a shopping mall?" Grab is targeting
ads business both from on-app and off-app modes. The vast data capability allows it to target
customers through ads, leading to higher conversion (Exhibit 17 and Exhibit 18).
BERNSTEIN
208 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
EXHIBIT 17: Self-serve ads platform launched by Grab
Source: Grab investor day presentation
EXHIBIT 18: How Grab advertises on its platform
Source: Bernstein photos, Bernstein analysis
SEA LTD IS NOT FAR BEHIND With the rapid change in technologies, new approaches to providing immersive shopping
experiences to customers have emerged in the e-commerce space. We broadly categorize
these approaches into three key buckets: (1) 3D experience through AR; (2) livestreaming/feeds
shopping experience; and (3) an efficient ads tool.
3D experience through VR/AR
Several consumer tech companies globally are launching VR experiences through AI, which gives
users a near-to-realistic experience of the product remotely. Shopee has led this innovation in the
ASEAN region by launching these features in 2020. It launched the Shopee BeautyCam feature
in 2020, which allows users to try virtual makeup products through an AI-powered face detection
system. Shopee highlighted that it has led to an average 3x increase in conversion for
beauty brands and will partner with more beauty brands in 2023 to implement this feature for
a wider variety of products (Exhibit 19 to Exhibit 21).
BERNSTEIN
SOUTH & SOUTHEAST ASIA TECH: OPPORTUNITIES FOR AI IN CONSUMER TECH 209
EXHIBIT 19: AR tool allowing users to
try lipstick shades virtually
Source: Bernstein photos, Bernstein analysis
EXHIBIT 20: AR tool allowing users to
try foundation shades virtually
Source: Bernstein photos, Bernstein analysis
EXHIBIT 21: AR tool allowing users to
try eyelash shades virtually
Source: Bernstein photos, Bernstein analysis
Livestreaming/video/feeds
Livestreaming is another form of shopping where AI plays a key role. While livestreaming has
been tried by many participants, only a few players such as Tiktok have made a dent, mainly
due to its strong algorithm system, which curates videos of products based on user data. TikTok
targets customers in the ASEAN region through the #ForYou feed, which curates videos for
users based on prior shopping experience and other user data. Shopee has launched Shopee
Live in the region, but has not been able to scale up that business in a meaningful way. As part
of its strategy for 2023, Shopee plans to use AI to recommend more personalized live
stream content and deals based on shopping behavior and user interests.
Continuously evolving advertising platform
Shopee is taking initiatives to make its advertisement offering more seamless and valuable for
brands. It first collaborated with Google in July 2020 to allow brands (registered in Shopee)
to create Google shopping ads through the Shopee Ad offering system, which can then be
published in multiple online channels supported by Google. Furthermore, Shopee has also
rolled out flexible tools to measure and optimize brand campaigns. Below is a snapshot of ad
optimization tools Shopee provides:
•To increase exposure. Shopee provides a click-through rate for each ad, which measures
the effectiveness of ads in driving shoppers to click on ads. It also tracks the number of
impressions and clicks for each ad. A brand can optimize its ads based on these parameters
by switching from exact match to broad match and increasing the number of keywords for a
broad match for products that have not met the desired impressions level.
•To increase sales. Shopee provides: (1) GMV generated per ad, and (2) ROI for each ad, which
is revenue earned per dollar of advertisement spent. For keywords with low GMV growth and
ROI, an advertiser can stop spending on keywords that bring less revenue per advertising
dollar spent and vice versa. An advertiser can also reallocate the advertising budget for higher-
priced products from low-priced products to increase ROI.
•To increase profit. Shopee provides a cost-to-income ratio (CIR) to understand the
percentage of the cost from advertising for each keyword. A brand can track each ad's
BERNSTEIN
210 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
profitability, measured at product profit margin ex ads reduced by the CIR for that product.
For products with higher CIR, a brand can lower bid prices carefully to reduce cost-per-click
while ensuring impressions and clicks remain stable.
Shopee also announced a partnership with five established ad media agencies (Dentsu Aegis
Network, Omnicom Media, Publicis Groupe, Havas Group, and Mediabrands) in September 2020,
aimed at increasing brands' presence and scale. In 2022, Shopee expanded its suite of marketing
solutions and tools to optimize marketing impact and returns for businesses.
•Facebook Ads on seller center. Shopee is deepening collaboration with Meta and will be
one of the first e-commerce platforms to integrate Facebook ads in the ASEAN region. It will
enable millions of sellers to create and manage campaigns easily, and earn higher ROI on ad
spend. As per Shopee's experience, Facebook ads have shown to deliver 14 times the ROI in
the electronics category in 2021.
•New Shopee display ads. Shopee now allows brands on Shopee Mall to purchase new
homepage banner display ads to run their campaigns. This will give them the highest exposure
in-app and help drive on-site traffic to their stores and products. Sales generated by Shopee
ads increased by 200% YoY in 2021.
•New Customer Intelligence Dashboard. Shopee's new dashboard gives brands more
insights on shopper demographics and segmentation to customize their e-commerce and
marketing strategies. With a better understanding of customer preferences, brands can
improve customer engagement and deliver a more personalized experience in-app (Exhibit
22).
BERNSTEIN
SOUTH & SOUTHEAST ASIA TECH: OPPORTUNITIES FOR AI IN CONSUMER TECH 211
EXHIBIT 22: Shopee ads platform
Source: Company website, Bernstein analysis
VALUATION METHODOLOGY See Disclosure Appendix for valuation methodology.
EXHIBIT 23: Ratings and target prices
Ticker Company Rating Currency
Closing
Price
Price
Target
SE SEA O USD 60.48 100
GRAB Grab O USD 3.0 4.3
GOTO.IJ GoTo M IDR 111 115
SPX S&P 500 Index USD 4151.3
ASIAX Bloomberg Asia ex Japan USD 1141.6
Note: Closing prices as of May 25, 2023
Source: Bloomberg, Bernstein analysis and estimates
Risks See Disclosure Appendix for risks.
Venugopal Garre venugopal.garre@bernstein.com +65 6230 4651
Ankit Agrawal ankit.agrawal@bernstein.com +91 226 842 1441
BERNSTEIN
212 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
GLOBAL AUTOMATION: LATEST
DEVELOPMENTS IN THE
MANUFACTURING SECTOR
HIGHLIGHTS •We are witnessing an exciting inflection of AI adoption in manufacturing.
•In machine vision, leading players have been actively launching new products/solutions for
scaled adoption.
•The fusion of robotics with AI and vision is a key enabler for the Robot Renaissance.
INVESTMENT IMPLICATIONS We reiterate Outperform for Keyence, Cognex, Estun, and Hikvision, and Market-Perform for
FANUC.
THE PATH TO SCALE The entire field of AI in manufacturing is at the dawn of mass adoption, with some players
moving faster, entering the initial stage of product and customer multiplication (Exhibit 1). Going
forward, efforts will continuously be made on three fronts: (1) developing platform algorithm and
application software; (2) expanding the spectrum of "first uses" through customized projects;
and (3) standardizing "projects" to become "products." Leading players in this field make all three
efforts, but with their own focus and style.
•Platform algorithm and application software. In AI-based machine vision, Aqrose (private),
Cognex, and Hikvision offer standardized AI software suites, integrating important core
functions in inspection to target long-tail integrators/user-developers. Cognex and Hikvision
also integrate AI functions with their more comprehensive rule-based vision platforms. Since
our last in-depth analysis of the field (see here), Aqrose has introduced or greatly improved
new AI functions, including an AI-based defect generator (Exhibit 2), unsupervised learning for
classification/segmentation (Exhibit 3), incremental training function, and AI review system.
These new functions aim to lower the barrier in adoption by improving the efficiency in project
development and deployment (Exhibit 4). It has helped the company's exponential growth
over the past few years (Exhibit 5). Similarly, Cognex's AI-based vision functions have also
expanded, although fewer details are available to us (Exhibit 6).
•Expanding spectrum of "first uses" through customized projects. AInnovation (not
covered) has demonstrated the most success on this front. Among the latest flagship projects
are AI-based power grid scheduling and peak-shaving, AI solutions for sunroof gluing, interior
and exterior trim inspection, rain test, intelligent transmission system, vision-guided robot
welding projects, etc. (Exhibit 7). Following the initial projects, AInnovation's path toward
scalability is through two types of "repurchases" — the repurchase of the same solutions
by many more customers (1XN) and the repurchase of many more solutions by the same
key account (1+N), an approach similar to Hikvision's expansion in the enterprise digitization
space (see here). Its growth trajectory and quickly improving opex efficiency are evidence of
BERNSTEIN
GLOBAL AUTOMATION: LATEST DEVELOPMENTS IN THE MANUFACTURING SECTOR 213
its progress (Exhibit 8 and Exhibit 9).
•Product standardization. In the new AI era, Keyence stays true to its decade-long
strategy of product standardization — focusing on applications addressable by plug-and-play
products (Exhibit 10) to serve a wide range of customers at once, and waiting with discipline
for customized needs to become "productizable" before embracing the opportunity. For
example, its AI products currently mainly focus on plug-and-play inspection of manufacturing
objects with limited types of defects/features, but are used by thousands of customers,
including those completely unexperienced with AI. Keyence sees AI as one of the many tools in
its toolbox to solve challenging problems, and its hardware and software innovations reinforce
each other. A good example is its approach to incorporate AI in machine vision not only for
image analysis, but also to improve image acquisition by auto-optimizing image acquisition
parameters.
EXHIBIT 1: AI in manufacturing is at the dawn of mass adoption, with some players moving faster and just entering
the "multiplication" phase
Note: In this exhibit, the assessment on Google and Amazon (both covered by other Bernstein teams) is solely on their efforts in machine vision used in the
manufacturing sector. Their broader AI capability is a different topic and is not discussed here. AInnovation and OPT are not covered by Bernstein. Aqrose and
Smartmore are private companies.
Source: Bernstein analysis
BERNSTEIN
214 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
EXHIBIT 2: Aqrose has started using AI-generated dataset for training (launched April 2022)
Source: Aqrose, Bernstein analysis
EXHIBIT 3: Aqrose: unsupervised learning for classification (launched November 2022)
Source: Aqrose, Bernstein analysis
BERNSTEIN
GLOBAL AUTOMATION: LATEST DEVELOPMENTS IN THE MANUFACTURING SECTOR 215
EXHIBIT 4: Aqrose: shortening project development time to support accelerating adoption of AI in manufacturing
Source: Aqrose, Bernstein analysis
EXHIBIT 5: Aqrose annual delivery of AI products and cumulative number of factory customers
Source: Aqrose, Bernstein analysis
BERNSTEIN
216 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
EXHIBIT 6: Cognex AI-based software core functions
Source: Cognex, Bernstein analysis
EXHIBIT 7: AInnovation's coverage of industry verticals
Source: AInnovation, Bernstein analysis
BERNSTEIN
GLOBAL AUTOMATION: LATEST DEVELOPMENTS IN THE MANUFACTURING SECTOR 217
EXHIBIT 8: AInnovation delivered strong growth
Source: AInnovation, Bernstein analysis
EXHIBIT 9: AInnovation opex breakdown
Source: AInnovation, Bernstein analysis
EXHIBIT 10: Keyence has actively developed AI products since 2019 and targets to produce smart, simple, and stable
AI products
Source: Keyence filings and websites, Bernstein analysis
ROBOTICS + AI In addition to vision, robotics is another area where AI technology is quickly adopted and proves
the most impactful. The fusion of vision, AI, and robotics is a key driver of the ongoing Robot
Renaissance.
AI and vision allow robots to perform autonomous functions. These functions, unlike automotive
spot welding, for instance, do not have fixed, preprogrammable robot paths. With AI helping
recognize and locate objects, and plan the optimal trajectory of the robot arm in real time, new
functions such as bin picking (Exhibit 11), adaptive welding, machine tending, and assisted
BERNSTEIN
218 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
assembly move from labs to factories.
Thanks to these emerging robotic functions, an inflection is taking place — in Japan, the most
mature robotic market (link), robot density increased 30% in just four years after staying flat for
almost two decades (Exhibit 13). Much of the new adoption takes place in traditional and labor-
intensive industries, such as machinery, electronics, and metal products, where the required
robotic technologies are far from traditional (Exhibit 14). The same trend is manifesting outside
Japan as well, driving a global robotic adoption super cycle.
In addition to robot makers, many players from other fields are jointly making it happen. Analyzing
bin picking as an example, we identify three key technology building blocks — robotics, AI
path planning, and vision (Exhibit 12). Leading robot makers are expanding their capability to
drive technology development, e.g., FANUC develops all three in-house, UR integrates Energid;
Keyence expands from one block (vision) to two (path planning), hence silently making its own
TAM bigger; and independent third-party start-up Mech Mind (private) in our view is world-
leading in this area and renowned tech leaders in adjacent fields (e.g., Nvidia, covered by
Bernstein US Semiconductors team) are entering.
WHAT ABOUT CHATGPT? We came across a Microsoft study of potentially extending ChatGPT to robotics (link). The article
contemplates the prospect of human-robotic interaction via natural language instructions, and
argues that this can reduce workload in system integration and make it easier for non-technical
users to adopt robots.
We read with great interest, but remain unconvinced about the feasibility and necessity from this
first study which, according to the disclosure at the end of the article, "was largely written with
the assistance of ChatGPT, with prompts provided by the authors." That is, indeed, some good
"human-robot" collaboration in a very different form.
BERNSTEIN
GLOBAL AUTOMATION: LATEST DEVELOPMENTS IN THE MANUFACTURING SECTOR 219
EXHIBIT 11: Mech Mind 3D vision-based bin picking
Source: Mech Mind, Bernstein analysis
EXHIBIT 12: Path planning is a new "market" — players enter the space from many directions; collectively, they
enable the important robotic bin picking application
Source: Company websites, Bernstein analysis
BERNSTEIN
220 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
EXHIBIT 13: Robot density in Japan seemed saturated for 1.5 decades, but hiked in the past four years
Source: International Federation of Robotics, Haver, Japan Ministry of Internal Affairs and Communications Labor Force Survey, Bernstein analysis
EXHIBIT 14: Robot Renaissance primarily driven by a "downward" proliferation, i.e., the multi-fold increase of robot
intensity in traditional industries like machinery and metal products
Source: Statistics Bureau of Japan, Japan Robot Association, Japan Automobile Manufacturers Association, Bernstein analysis and estimates
BERNSTEIN
GLOBAL AUTOMATION: LATEST DEVELOPMENTS IN THE MANUFACTURING SECTOR 221
VALUATION METHODOLOGY See Disclosure Appendix for valuation methodology.
EXHIBIT 15: Ratings and target prices
Note: Share prices are as of May 25, 2023. FY2022 means FY3/23 for Keyence, FANUC. Closing prices for reference stock indices as of May 25, 2023: ASIAX —
1,141.60; JPL — 1,379.29; SPX — 4,151.28.
Source: Bloomberg, company reports, Bernstein analysis and estimates
RISKS See Disclosure Appendix for risks.
Jay Huang, Greater China Director of Research jay.huang@bernstein.com +852 2918 5746
Dien Wang dien.wang@bernstein.com +852 2918 5743
Weibin Liang weibin.liang@bernstein.com +852 2918 5242
BERNSTEIN
222 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
INDIA TECHNOLOGY, MEDIA & INTERNET:
OPPORTUNITIES FOR AI IN TECHNOLOGY
SERVICES
HIGHLIGHTS •We believe AI will accelerate the shift from commodity IT services to high-value IT services
such as consulting and custom app design.
•Deep technical expertise, consulting strength, and deployment of AI in internal processes
would be key providers of competitive strength.
•Among the publicly listed IT players, we see Infosys, TCS, and Accenture as emerging leaders
in AI, with Accenture as the likely winner, driven by its strength in Digital.
INVESTMENT IMPLICATIONS We are Outperform on TCS, Infosys, and Tech Mahindra; Market-Perform on HCL Technologies;
and Underperform on Wipro.
AI IS FINALLY GOING
MAINSTREAM
AI is becoming big
The urgent need for visibility and insight at scale became undeniably clear during the COVID-19
pandemic. From factories that shut down, to supply chain disruptions, to consumer goods
companies unable to pivot fast enough — many organizations struggled because they were not
prepared. Based on an Accenture survey, 39% of the technology enterprises currently utilizing
AI are using AI in scaling up processes across the production system, while 16% (Exhibit 1) are
still experimenting with AI technologies.
EXHIBIT 1: Technology organizations scaling up versus experimenting in AI
19%
27%
39%
16%
45%
12%
18%
24%
22% 19%
36%
15%
20%
20%
9% 12%
Scaling Experimenting
5G Artificial intelligence Cloud Digital Twins Edge Computing IoT RPA Robots/Autonomous robots
Note: Scaling: applying technology to change a process (or multiple processes) in a full production system; Experimenting: applying technology to testing within a
process (e.g., A/B testing), but not yet changed the process in a full production system.
Source: Accenture Technology Vision 2021, Bernstein analysis
BERNSTEIN
INDIA TECHNOLOGY, MEDIA & INTERNET: OPPORTUNITIES FOR AI IN TECHNOLOGY SERVICES 223
ACCENTURE AI ACQUISITIONS Accenture (not covered) has an increased focus on infusing data and AI across its offerings that
enables clients to generate higher value from their digital transformation initiatives. Accenture
has continued to invest in AI companies to drive growth and acquire new capabilities (Exhibit
2). With Flutura, Accenture continues to build its data and AI capabilities for clients. In 2022, it
acquired data science company ALBERT in Japan. Other recent AI acquisitions include Analytics8
in Australia; Sentelis in France; Bridgei2i and Byte Prophecy in India; Pragsis Bidoop in Spain;
Mudona in the UK; and Clarity Insights, End-to-End Analytics, and Core Compete in the US.
EXHIBIT 2: Accenture: AI acquisitions
AI Acquisitions Description
Month, Year of
Acquisition
Flutura
Industrial AI company focused on unlocking high-value
operational outcomes for the Energy, Chemicals, Process
Manufacturing & Heavy Engineering industries.
Mar-23
ALBERT
ALBERT offers AI and big data analytics services, AI-based
algorithm development, AI implementation consulting, and
data science training support
Nov-22
Bridgei2i
BRIDGEi2i provide contextual AI-powered analytics solutions to
solve complex business problems and deliver digital
transformation outcomes
Oct-21
Core Compete
A cloud analytics services, which enables digital
transformations with cloud-native solutions that deliver AI and
machine learning-infused business outcomes
Apr-21
End-to-End
Analytics
An applied analytics and data science consultancy, with deep
expertise in applying optimization, machine learning and AI
technologies across a broad range of industries
Dec-20
Sentelis
Data consulting and engineering company that specializes in
designing and scaling data and artificial intelligence (AI)
capabilities.
Jul-20
Byte Prophecy
An automated insights and big data analytics company to meet
the growing demand for enterprise-scale AI and digital analytics
solutions
May-20
Mudano
A data and machine learning company focused on financial
services.
Feb-20
Clarity Insights
A Chicago-based company that provides AI and machine
learning services to companies in the healthcare, financial
services and insurance industries.
Dec-19
Pragsis Bidoop
A company with strong expertise in big data, AI and advanced
analytics, with a diversified client base
Sep-19
Analytics8
Analytics8 is a data and analytics consulting firm that specializes
in data strategy and all aspects of implementation.
Aug-19
Source: Company reports, Bernstein analysis
WHAT IS AI? WHY NOW? AI — a definition
We find the most straightforward definition of AI to be the group of technologies that
takes on "human-like" skills, such as processing language, perceiving, representing
knowledge, reasoning, and planning and acting. Underlying these technologies are
enablers such as learning algorithms and supporting infrastructure.
The technologies considered to be AI today can be grouped into five categories resembling
human skills, including systems that: (1) process language; (2) perceive their environments;
BERNSTEIN
224 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
(3) represent knowledge; (4) reason; and (5) plan and conduct an action.
•Language processing: Technologies focused on speech recognition. Prominent examples
include OpenAI's ChatGPT, Amazon Alexa, Apple Siri, Google Home, and Google Translate.
•Perception: Technologies that perform visual and sensory classification tasks. Prominent
examples include computer vision applications such as Kinect.
•Expertise: Technologies that develop knowledge "libraries" for specialized areas. Prominent
examples include expert systems such as IBM Watson (IBM is covered by A.M (Toni)
Sacconaghi).
•Reasoning: Technologies that solve specific tasks or puzzles. Classic examples are game-
playing machines such as Deep Blue and Alpha Go.
•Planning and acting: Technologies that plan, navigate, and act in the physical world.
Examples are self-driving vehicles, such as Tesla and Alphabet's Waymo.
Underlying these technologies are enablers, including learning algorithms, and
supporting infrastructure such as sensors, unstructured data management, and hyperscale
cloud computing. Common types of learning algorithms include machine learning, deep learning,
and reinforcement learning.
Another way to think of AI is through its use cases/applications. Exhibit 3 summarizes some
major use cases across different sectors of the economy. What stands out is the remarkable
diversity of AI applications, both in terms of the sophistication of the underlying AI technologies
and the contexts in which they are applied.
EXHIBIT 3: Major AI use cases
AlI use cases Description/Examples Industries
Customer experiences and
back-office processes
Humanoid robots in branches, machine vision and Natural processsing language to
scan and process documents, Real-time transaction risk monitoring, detecting fraud
patterns and cyber attacks
Banking
Autonomous transportation
Self-driving vehicles (e.g., Alphabet's Waymo, Tesla's Autopilot), smart cities (e.g.,
congestation control, smart highway pricing), navigation services (e.g., Google
Maps, Mapquest), etc.
Automobiles,
Transportation
Recommender systems
Filtering systems that suggest recommendations for individual customers (e.g.,
Amazon, Hulu, Netflix, LinkedIn)
Retail, Entertainment,
Social Media, etc.
Robotics
Boston Dynamics' Atlas, Honda's ASIMO, RobotCub Consortium's iCub, Aldebaran
Robotics' Romeo, etc.
Manufacturing, Military,
Home Services,
Healthcare, etc.
Chat bots Various examples at Uber, Sephora, Bank of America, Domino, Pizza Hut, etc. Retail, Food, Finance, etc.
Cybersecurity risk
assessment
Largely machine learning technologies used in dynamic risk analysis, anomaly
detection, antimalware, etc.
Cybersecurity
Healthcare analytics and
prediction
Automated imaging diagnostics (e.g., IBM Watson, Imagia), virtual care (e.g.,
Sense.ly), behavioral tracking (e.g., Ginger.io)
Healthcare
Visual and hearing aid Examples include Oticon's Opn, OrCam's MyMe, Alphabet's Lens, etc. Healthcare, Entertainment
Intelligent grading and
tutoring systems
Teaching robots (e.g., PLEO rb, Ozobot), online learning systems (e.g., Duolingo,
SHERLOCK), testing systems (grading tools used by ETS and Pearson)
Education
Source: Company reports, Bernstein analysis
BERNSTEIN
INDIA TECHNOLOGY, MEDIA & INTERNET: OPPORTUNITIES FOR AI IN TECHNOLOGY SERVICES 225
AI — WHY NOW? AI has been around for more than 60 years (Exhibit 4), yet is only just now gaining widespread
commercial momentum — at least 20 years after the Internet did. So what took so long? There
are, in our view, two reasons for why AI development is accelerating now:
•One, a revival of AI research driven by a focus on a different set of learning approaches.
One can divide the development of AI into three distinct periods. The first is the pre-1970s
era, in which much of the basis and inspiration for AI technologies was first formulated and
received significant enthusiasm and funding. However, from the 1970s to the early-1990s,
the lack of significant practical achievements and a series of research debacles led to a
prolonged period of public disillusionment and funding cuts in AI — a period called the "AI
Winter." It was only during the 1990s that researchers such as Rodney A. Brooks successfully
advocated for a rethinking of AI as a practical, task-solving endeavor and opened the doors
for renewed interest in the field.
•Two, advancements in other technologies have enabled an acceleration in AI research.
The growth of cloud technologies and plummeting computing, storage, and networking costs,
combined with advancements in technologies such as sensor networks, unstructured data
management, and hyperscale cloud computing have, in recent years, made the "brute force"
approach to AI more viable.
Because of the proliferation of data and the maturity of other innovations in cloud processing
and computing power, AI adoption is growing faster than ever. Companies now have access to an
unprecedented amount of data, including dark data they didn’t even realize they had until now.
These treasure troves are a boon to the growth of AI.
EXHIBIT 4: Brief history of AI
- 1940s 1950s 1960s 1970s 1980s 1990s 2000s -
1936
1943
1956
1959
1957
1960s
1970s
1972
1990s
1997
2009
2011
2005
2016
Turing machine
formulated
A formal model of
"artificial neurons"
was published
The General Problem
Solver was created
NLP systems
(SHRDLU, ELIZA)
were developed
Stanford researchers
designed MYCIN,
an expert system
Deep Blue
defeated
Garry Kasparov
"AI Winter"
Google began its
self-driving project
Honda introduced
ASIMO
Researchers pushed
for a more practical,
task-based approach
in AI
Shakey the robot
was invented
"Pandemonium"
paper laid the basis
for computer vision
Workshop on AI
at Dartmouth
Watson defeated 2
Jeopardy champions
AlphaGo
beat Lee Sedol
History of Artificial Intelligence
2016
Sophia debuts
as first robot
citizen
2018
AI model
outscores
humans
2020
OpenAI GPT-3
released
Source: Stanford's The One Hundred Year Study on Artificial Intelligence, Gartner, Bernstein analysis
BERNSTEIN
226 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
We believe AI will impact IT services in two major areas:
•Impact area #1 — efficiency, when IT services firms apply AI to facilitate the running and
management of clients' IT environments
•Impact area #2 — growth, when IT services firms help clients deploy AI technologies into
specific industry applications
Impact area #1 — AI applied to improve efficiency of IT management
We believe that AI deployed for efficiency will likely be the next "must have" capability to compete
in IT management, similar to offshore labor in the last decade. While we don't see it displacing
the need for service providers, we are skeptical it will enable IT services firms to break the labor-
revenue "linearity problem," as we expect efficiency gains from AI will be competed away.
Robotic process automation (RPA): The clearest and most popular context in which AI
technologies are applied to run IT environments is RPA (also known as "IT automation").
In applying AI into RPA, the common approach of IT services companies has been to layer AI
technologies and AI-related building blocks such as machine learning and natural language
processing on top of traditional RPA platforms (Exhibit 5). Some examples include:
•Infosys NIA: AI platform built by Infosys that helps businesses automate and streamline their
processes, and improve operational efficiency. NIA is applied in several business processes
such as customer service, supply chain management, and finance.
•TCS ignio: IT automation platform that can "ingest" both structured and unstructured
information to build a knowledge base of information on servers, networks, databases, etc.
•Cognizant's Intelligent Automation Platform: Combines NLP and machine learning to
organize unstructured information and automate repetitive tasks and compliance tracking.
•Wipro Holmes: IT system that uses pattern recognition, visual learning, and NLP to identify
incidents on the application stack, automate repetitive tasks, and conduct monitoring and
self-healing.
Productivity benefits from AI: These kinds of use cases led by AI pose a shift for labor-
based IT services, as smart machines replace the need for human labor, potentially disrupting
the traditional IT services business model. And this is no small problem: IT services is a nearly
US$900Bn industry, and the top 10 listed IT services players employ more than three
million people. However, the industry has seen stagnant revenue per employee: The revenue
per employee has maintained a narrow band of US$45,000-US$55,000 per annum (Exhibit 6).
EBIT per capita has reduced to US$11,000-US$13,000 per annum.
BERNSTEIN
INDIA TECHNOLOGY, MEDIA & INTERNET: OPPORTUNITIES FOR AI IN TECHNOLOGY SERVICES 227
EXHIBIT 5: Digital has a new stack — AI, IoT, 5G, blockchain becoming more mainstream
Source: Company reports, Bernstein analysis
EXHIBIT 6: Revenue/head — IT services (US$K)
48
44
49 50
48
52 52 51 52
54 55 54 55
57
47
42
46 47
45
47 50 49
47
49
51 50
48 47
45
43 45 46
44
45
47
44 43
47
49
47
43
46
30
35
40
45
50
55
60
FY09 FY10 FY11 FY12 FY13 FY14 FY15 FY16 FY17 FY18 FY19 FY20 FY21 FY22
Infosys TCS Wipro
Source: Company reports, Bernstein analysis
In our view, AI will likely be the next "must have" capability to compete in IT management,
similar to offshore labor in the last decade. Although adoption of AI in IT management
remains in its early stages, we believe it will rapidly become a "must-have" capability among IT
services firms.
AI can be an important source of competitive differentiation in a commoditizing industry,
as it may significantly drive down the cost of running IT. In this sense, the wave of AI can be viewed
as the next step after companies have taken advantage of outsourcing, offshoring, and cloud
delivery models to drive efficiency in IT management. We expect AI-augmented automation to
become the next driver of efficiency in IT management and, similar to offshore labor, may create
the opportunity for IT services players to differentiate themselves for a period.
BERNSTEIN
228 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
Impact area #2 — Growth will open new revenue opportunities for IT services
companies
We believe AI deployed for growth — helping clients deploy AI solutions — will likely open
major new revenue opportunities for IT services companies, as a core technology within
"Digital." According to recent Gartner CIO surveys, 63% of CIOs are actively deploying
AI, and a whopping 93% are planning to use third-party help.
IT services firms play the role of "facilitators" in the AI wave in two ways: (1) helping clients
navigate the AI environment, and (2) directly deploying AI technologies and building blocks into
clients' products, services, and processes.
Shift to high-value services: high-value services (e.g., design, consulting, etc.) will see
limited impact. A key defining characteristic of the new leading model in IT services is the ability
to deliver high-value services, as needs shift from commodity services toward more specialized,
customized requirements (e.g., custom analytics, cloud, etc.).
•Industry focus and specialization: Industry focus and specialization is a natural implication
of characteristics of high-value services. We believe it is of sufficient importance to be worth
noting separately. IT services companies have built vertical specific products/platforms to
differentiate in the market.
•Digital/core modernization (spend rotation from ERP): We believe CIOs are shifting
spend from ERP into Analytics, Mobile, and Digital — this reflects the normal lifecycle of an
application type.
•Cloud capability: Cloud has been the driver of shift in the IT services landscape. We believe
the ability to modernize core and work with hyperscalers to shift to the cloud and manage a
complex cloud environment will be a key differentiator.
Digital — a key growth area: AI presents a compelling new growth area for IT services
firms, as a core technology within "Digital." Digital can be defined by blending three different
layers of enterprise architecture — the experience layer, the intelligence layer, and the digital core
later. AI would be most active in the intelligence layer.
In our view, AI presents a compelling new growth area for IT services firms. There is a long
runway ahead for AI adoption. Enterprises are adopting AI technologies in specific, limited use
cases before rolling out AI across multiple processes and functions.
•IDC forecasts global spending on AI-centric systems to reach US$154Bn in 2023, with
banking accounting for a significant chunk at 13.4% (Exhibit 7).
•IDC estimates that AI IT services spending will be ~US$41Bn in 2026, growing at a CAGR of
22% (Exhibit 8). We note that IT services players stand to benefit from much of this spend.
To put these into perspective, Gartner estimates that total IT services spending in 2022 was
~US$1.3Tn.
In the near term, we expect consulting services to benefit from much of the AI-driven
tailwind as clients begin to navigate the wide array of AI technologies and AI vendors. As a result,
consulting offerings with a specific focus on AI (e.g., Accenture Liquid Studios, Cognizant's AI
database, etc.) will enable IT services players to tackle the early wave of AI adoption. As time
BERNSTEIN
INDIA TECHNOLOGY, MEDIA & INTERNET: OPPORTUNITIES FOR AI IN TECHNOLOGY SERVICES 229
goes on, expertise in custom app development and implementation will be critical to
success in AI. In other words, the adoption of AI in IT services will follow the same path as Digital
adoption.
EXHIBIT 7: IDC estimates worldwide spending on AI-
centric systems to reach US$154Bn in 2023
Source: IDC, Bernstein analysis
EXHIBIT 8: IDC forecasts companies to increase AI IT
services spending at a CAGR of 22%
18.4
~41
2021 2025
Worldwide AI services forecast ($ Bn)
CAGR: 22%
Source: IDC, Bernstein analysis
EXHIBIT 9: Majority of respondents in Gartner's survey
believe that AI technologies can impact their businesses
77%
65% 63%
43%
Machine
Learning
Natural
Language
Processing
Smart
Robots
Autonomous
Vehicles
Gartner Survey: Could this technology
have an impact on your organization's
business? (100% = 228 respondents)
Source: Gartner, Bernstein analysis
EXHIBIT 10: 93% of the participants in Gartner's survey
plan to use a service provider in AI adoption
Using/
planning to
use a
service
provider,
93%
Not using/ planning to
use a service provider,
7%
Gartner Survey: Plans for using a Service
Provider in AI adoption (100% = 265
respondents)
Source: Gartner, Bernstein analysis
BERNSTEIN
230 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
Why are IT services relevant in the broader AI development landscape?
There is a considerable enthusiasm to adopt AI technologies among clients. Further, executives
at enterprises that have been adopting AI technologies report a belief in the impact of AI on
their businesses (Exhibit 9). At the same time, 93% of executives who have plans to adopt AI
also think they need help from an IT services provider (Exhibit 10). It appears that many clients
currently face difficulty in self-applying AI effectively in their business settings, hence the need
for IT services.
AI — WHAT DETERMINES A
WINNER IN IT SERVICES?
Given the needs of clients, as mentioned earlier, we believe there are three characteristics of
an IT services company to gain a competitive advantage in AI:
•Deep technological expertise: It goes without saying, but IT services providers that can
forge a strong understanding of the latest developments in AI will stand to benefit. We have
seen that there are three major routes for IT services firms to enhance their AI capabilities.
•Partnerships: IT services firms can partner with AI innovators/platform providers such
as Microsoft (covered by Mark Moerdler), Amazon (covered by Mark Shmulik), and IPsoft
(unlisted) or with research institutions such as MIT and Stanford. This is akin to the SAP/
Oracle (covered by Mark Moerdler) partnerships during the early development phase of the
ERP era (Exhibit 11 and Exhibit 12).
•Organic investment: To give investors a flavor of the direction of in-house AI research
done in the IT services industry, Exhibit 13 shows a list of some recent AI patents filed
by major IT services firms. Unsurprisingly, most patents focus on IT automation and
user interaction, with a few patents targeting industry-specific use cases beyond IT (e.g.,
predictive platform for electrical grids).
•M&A: Finally, some IT services firms have targeted small, tuck-in acquisitions to strengthen
their AI platforms and solutions. For example, Infosys recently acquired "Oddity," a digital
marketing, experience, and commerce agency.
EXHIBIT 11: Cognizant and Accenture have partnered with various companies and institutions to develop AI
solutions
• AutomationAnywhere • Cicero • AutomationAnywhere Research Institutions:
• Arago • Cloudera • Blue Prism • MIT (U.S.)
• Artificial Solutions • Hitachi Data Systems • Fusion • IIT (India)
• Automic • Narrative Science • IBM Watson • DFKI (Germany)
• Ayehu • Nuance • Ipsoft
• Blue Prism • OpenSpan • Mighty AI
• Captricity • UiPath • Nervana
• Celaton • Wise.io • Next IT
• Cenx • WorkFusion • Saffron
Cognizant's SmartSystems Alliance
Accenture's AI Partners
Note: Except IBM, none of the companies are under our coverage.
Source: Company reports, Bernstein analysis
BERNSTEIN
INDIA TECHNOLOGY, MEDIA & INTERNET: OPPORTUNITIES FOR AI IN TECHNOLOGY SERVICES 231
EXHIBIT 12: Infosys and TCS AI partnerships
• 3LOQ • Neenar Analytics • Eightfold AI • Opsview
• ACTIVE.AI • Polygon • Digital.ai • Palo Alto
• ailleron • Value3 • Five9 • MATRIXX
• clinc • Xlrt • Blue Yonder • Twilio
• ENDOR • Glia • Kore.ai • Splunk
• FEATURE SPACE • Xlrt
Infosys Alliance Partners
TCS Alliances and Partnerhsips
Note: Except Infosys and TCS, none of the companies are under our coverage.
Source: Company reports, Bernstein analysis
•Consulting strength: The next major capability required for success in AI is consulting
capability — the ability to identify relevant use cases for AI technologies and to educate
clients on leveraging AI to serve specific business functions. Consulting capability is critical
in the early years of AI development, as consulting will drive a major portion of AI-related IT
spend, and IT services firms with strong consulting capabilities (e.g., Accenture, Capgemini
(not covered) and, to a certain extent, Cognizant) can gain significant benefit.
•Aggressive deployment of AI in in-house operations: To capture the efficiency benefit of
AI, IT services firms need to find ways to apply AI into their own service delivery models. In fact,
we have seen many IT services firms moving in this direction, based on the wide variety of IT
automation platforms that incorporate AI elements, such as Accenture's myWizard, Infosys'
Nia, TCS' ignio, Cognizant's Intelligent Automation, etc. We will be looking for additional signs
of acceleration in penetration of these automation platforms (Exhibit 14 and Exhibit 15).
EXHIBIT 13: Recent AI patents filed by IT services firms with the US Patent and Trademark Office
Date Companies Patent Titles
Relevant
technologies/Use cases
2023 Accenture
Utilizing machine learning and image filtering techniques to
detect and analyze handwritten text
Natural language
processing
2023 Accenture
Utilizing artificial intelligence to predict risk and compliance
actionable insights, predict remediation incidents, and
accelerate a remediation process
Predictive model
2023 Wipro
Method and system for optimizing memory requirement for
training an artificial neural network model
Neural networks
2023 Wipro
System and method of validating multi-vendor Internet-of-
Things (IoT) devices using reinforcement learning
Machine learning
2023 Cognizant
Data mining technique with distributed novelty search
Data mining, Natural
language processing
2022 Infosys
System and method for analysing customer experience from
unstructured social media data
Big data, Data center
automation
2020 Infosys
Automated system for development and deployment of
heterogeneous predictive models
Predictive model
Source: US Patent and Trademark Office, Bernstein analysis
BERNSTEIN
232 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
Among the large public IT services firms, Accenture is a likely winner, building on its
success in Digital, as it: (a) has developed a strong foundation of AI skills and offerings, through
organic and inorganic investment; (b) has strong core capabilities in consulting, solution design,
and custom application development; and (c) has deployed AI-driven IT automation platforms
within its service delivery and operational models. We are particularly impressed with Accenture
myWizard and Liquid Studios.
EXHIBIT 14: Summary of current IT automation platforms/solutions offered by IT services companies (not
exhaustive)
Company Platform Description
Accenture myWizard
myWizard is a collection of 6 virtual agents: 1) architect (recommends IT structure); 2) scrum master (monitors different
requirements and metrics of the IT environment); 3) testing advisor (provides testing suggestions and sample test
cases); 4) data scientist (identifies data patterns); 5) project manager (assesses risks in a delivery process); and 6)
modernization analyst (provides suggestions for modernization initiatives). The platform combines machine learning,
natural language processing, and data analytics with Accenture's database
Cognizant
Intelligent
Automation
Cognizant's Intelligent Automation consists of: 1) a tracking/compliance tool; 2) Automatika (uses NLP and machine
learning to extract and organize unstructured information); 3) HPA (automates repetitive tasks such as financial
transaction clearing); and 4) ADPART (automates test design and implements model-based testing).
Dell Services
Automation
Platform
A fairly standard automation platform that includes automated monitoring services, automated application testing and
repeatble, rules-based task performance
HCL
Technologies
DRYiCE
DRYiCE platform consists of 40+ IT automation modules, which include monitoring capabilities (through receptors), data
collection solutions, testing automation, virtual service desk, visualization interface, and service orchestration. HCL
claims that these services are embedded with machine learning, natural language processing, and data analytics to
augment IT workers.
Infosys Nia
Infosys Nia, launched in April 2017, builds upon the company's previous automation platform, Mana. While use cases
are limited so far, the company claims that it combines machine learning, analytics, character recognition, natural
language processing with robotic process automation capabilities. Some of the example applications include
constructing knowledge models of labor contracts, designing virtual agents to interact with customers, and providing
risk assessment. According to INFY, Nia can be deployed as an IT solution as well as a business operations platform.
Syntel SynBit
An IT automation platform that automates monitoring and incident managment, provides continuous testing/delivery,
and automatically manages repetitive tasks while integrating with the backend. It appears that the platform only has
limited AI elements.
TCS ignio
A platform that can ingest unstructured and structured data to construct its knowledge base; provides monitoring of
performance and incidents; uses a learning subsystem to recognize and complete tasks; and identifies trends and
patterns. The use cases highlighted for the platform include software/hardware installations, "self-healing", cloud
provisioning automation and diagnostics analysis
Tech Mahindra GAiA 2.0
GAiA is an industrialized version of the open-source Acumos platform.G AiA 2.0 will enable comprehensive AI and ML
driven platform capabilities and services to be deployed across mainstream, optimizing enterprise operations in real
time across industry verticals. It offers an enriched marketplace of models and numerous features to empower
enterprises across industry verticals to build, manage, share and rapidly deploy AI and ML driven services and
applications addressing critical business problems.
Wipro Holmes
Wipro Holmes has an user interaction interface/digital agents (that uses facial recognition and NLP through a
partnership with IPsoft's Amelia), combined with predictive and pattern recognition systems, and computer vision
capabilities to conduct monitoring, incident resolution, and perform other repetitive tasks. The platform is supported by
a knowledge virtualization database that accumulates knowledge from human workers and experts.
Source: Gartner, Company reports, Bernstein analysis
BERNSTEIN
INDIA TECHNOLOGY, MEDIA & INTERNET: OPPORTUNITIES FOR AI IN TECHNOLOGY SERVICES 233
EXHIBIT 15: Examples of AI-related solutions deployed by IT services firms
Industry/Area Use Cases
Consulting
AI prototype and proof-of-concept supports (e.g., Accenture Liquid Studios, PwC's Analytics Innovation Accelerator
and Emerging Technology Lab); AI vendor library (e.g., Cognizant's database of AI products); contract resolution
(e.g., Infosys Nia-based solutions)
Transportation
Smart city monitoring (e.g., IBM's Metro Pulse, which gathers local data on weather, sales, local events, social
media), streetlight control (e.g., TCS' IUX software, which optimizes streelight operations), etc.
Manufacturing
Faulty equipment monitoring (e.g., Infosys' AI solution for a mining company); virtual agent (e.g., IPsoft Amelia's
virtual agents can support maintenance engineers in remote locations); analytics platform (e.g., Cisco's Fanuc
Intelligent Edge Link and Drive)
HR Services
Automated query response (e.g., Accenture Amelia)
Finance
Mortgage credit counselor solution (developed by Accenture); tax-related expert system (e.g., Infosys Nia based
solutions); loan decision support (e.g., Wipro Holmes-based solutions
Healthcare
Customer's risk assessment (e.g., Cognizant Life Engage, IBM Watson Health Patient Engagement); health
organization optimization (e.g., IBM Watson Health); diagnostics support (e.g., IBM Watson Health); treatment
recommendation (e.g., IBM Watson for Oncology)
Risk/Compliance
Risk assessment solutions (e.g., IBM Watson Risk and Compliance)
Telecom
Order-to-activation support (e.g., Infosys Nia-based solutions)
Agriculture
Field survey solutions (e.g., Accenture's solution, which combines drones, computer vision, and geospatial analytics
to survey palm fields in Indonesia)
Government
Constituency query automation (e.g., Accenture's query automation solution developed for an Italian government
agency)
Source: Gartner, IDC, company reports, Bernstein analysis
VALUATION METHODOLOGY See Disclosure Appendix for valuation methodology.
EXHIBIT 16: Ratings and price targets
25-May-2023 Target
Ticker Rating Currency Closing Price Price
INFO.IN O INR 1,304.35 1,550.00
INFY O USD 15.66 19.00
TCS.IN O INR 3,293.50 3,560.00
TECHM.IN O INR 1,098.10 1,250.00
HCLT.IN M INR 1,116.00 1,070.00
WPRO.IN U INR 394.10 310.00
WIT U USD 4.73 3.90
ASIAX 1,141.60
Source: Bloomberg, Bernstein analysis and estimates
RISKS See Disclosure Appendix for risks.
Rahul Malhotra rahul.malhotra@bernstein.com +65 6230 2344
Sanjit Shinde Sanjit.Shinde@bernstein.com +91 226 842 1469
BERNSTEIN
234 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
GLOBAL SOFTWARE: AI IS CORE TO THE
FUTURE — HIGHLIGHTS FROM ADOBE
SUMMIT CONFERENCE
HIGHLIGHTS •Adobe has been making investments in AI for quite a while and is currently very well-
positioned in the space — similar to Microsoft.
•The conference saw numerous announcements from Adobe, especially the launch of Adobe
Firefly, which uses multiple large language modules (LLMs) generated by GPT to solve difficult
content creation (images, videos, etc.) and manipulation requirements of creatives.
•Adobe noted that these additions will not negatively impact margins. Going forward, any cost
of revenue impact from GPT and other AI capabilities will be absorbed by the company's
overall margin efficiency and organic revenue growth.
INVESTMENT IMPLICATIONS At the Adobe Summit, Adobe unveiled numerous new solutions and capabilities with a focus
on the use of AI and generative AI. The new capabilities increase the differentiation of Adobe's
offerings and the moat around the business. They should drive sustained revenue growth,
decreased churn, and stable margins. We see this as expanding the TAM for Adobe and in line
with our view that Adobe is a unique company that should sustain/increase growth while driving
strong margins.
Adobe has shown that it is taking an approach very similar to Microsoft — AI is a fundamental
technology that will be infused into everything the company creates. That approach, we
believe, will lead to tightly integrated solutions and much more value for customers.
SUMMIT RECAP Today, every company is an AI company. With the buzz surrounding Microsoft's Bing Chat (GPT)
announcement (see Microsoft Quick Take: The first green shoots of the OpenAI partnership —
our initial thoughts), it does not matter what a company does, it is now being positioned as an
AI company, an AI-powered company, or some other variant. That is where we are in the hype
cycle for AI. We have seen this many times before (e.g., document management, cloud and, most
recently, the Metaverse), and there is often a mix of marketing and reality.
Adobe has always been an AI-driven company. It has been using AI technologies and techniques
for years, but most people have not realized this. Most of Adobe's AI was computational rather
than machine learning (or generative), but what Adobe has been doing is very much AI. As
investors try to pick the innovative and differentiated use of AI- and machine learning-driven
products, and vendors that recently jumped on the bandwagon, one should remember that
Adobe has the internal expertise and has shown the results of these investments, especially in
the Creative Cloud. In fact, Adobe was very early in exposing its AI capabilities as building blocks
that others can use to enhance their solutions via Adobe Sensei (Exhibit 1).
BERNSTEIN
GLOBAL SOFTWARE: AI IS CORE TO THE FUTURE — HIGHLIGHTS FROM ADOBE SUMMIT CONFERENCE 235
At the Adobe Summit, the company's Experience (or digital marketing) conference, it was not
surprising that AI was front and center, and not just for the Adobe Experience Cloud. The
company's numerous announcements at the conference include:
•Launch of Adobe Firefly, the company's AI powered creative tool;
•Expansion of Adobe Experience Cloud analytics to incorporate further AI capabilities;
•Launch of Adobe Express enterprise edition;
•Blending of content creation further into the Experience Cloud to create a more personalized
and targeted marketing experience; and
•An expanded vision for the future of the company where AI and especially generative AI
capabilities weave an easier to use, more powerful, and more personalized experience for the
users of Adobe tools and an increased ROI for Adobe customers.
Adobe believes that AI and, more specifically, generative AI is a platform technology that
will infuse not just Adobe's offerings, but also many other companies, and not just software
companies. Adobe believes it is uniquely positioned within its target markets. It has been
investing in AI for more than 10 years and has extensive expertise, leading experts in the space, a
very large set of curated content (Adobe Stock), data, and technology. The company has already
built multiple LLMs and believes none of its smaller competitors have similar capacity. While
there are likely to be open source and other competition, they are not going to meet the rigorous
requirements of corporations.
In fact, we would argue there are only a couple of companies with the depth and breadth of data,
expertise, and training and inference expertise, and Adobe is one of those companies. Adobe has
also built deep partnerships with leading experts in the space, including Microsoft and NVIDIA,
covered by our US Semiconductors team (which Adobe announced a new partnership with to
drive optimization of tools for generative learning and problem-solving in the content space).
BERNSTEIN
236 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
EXHIBIT 1: Generative AI is becoming part of the fabric of Adobe
Source: Company presentation
WHAT COULD IT MEAN FOR
REVENUE AND MARGINS?
Adobe has not yet defined how all these announcements are going to be monetized, but we
believe, based on our conversations with management, that the company is taking different
approaches toward Digital Media and Experience.
Experience Cloud: Adobe will, we believe, charge for the new Experience Cloud solutions
(Adobe Experience Manager re-imagined and Adobe Product Analytics) as it has historically
monetized new Experience Cloud solutions via usage-based pricing. It is also creating
new solution bundles, including the Adobe Content Supply Chain (which includes Adobe
Creative Cloud for Enterprise, Adobe Workfront, Adobe Experience Manager (Sites, Assets and
segmentation and content profile services from Adobe Experience Platform), Adobe Express for
Enterprise, and Frame.io) via some form of bundle/site license.
Adobe Digital Media has two new offerings: Adobe Firefly and Adobe Express Enterprise
Edition. Adobe Firefly will be available in at least three variants:
1. A free entry-level solution to drive top of the funnel action. Adobe expects to convert many of
those users, over time, to paid users.
2. Adobe Firefly as part of the Creative Cloud will likely be offered as an add-on module or
included in future suite offerings. These offerings will generate what we believe could be
meaningful revenue, given how differentiated Firefly is (as discussed in the next section). It will
likely also decrease churn.
3. As part of Experience Cloud and likely Document Cloud suites/bundles. We started to hear
the value proposition of the integrated solution, and it could not only drive revenue, but also
add meaningful competitive differentiation and, therefore, could be the driver of companies
switching from digital marketing point solutions or even other suites (e.g., Salesforce).
BERNSTEIN
GLOBAL SOFTWARE: AI IS CORE TO THE FUTURE — HIGHLIGHTS FROM ADOBE SUMMIT CONFERENCE 237
Adobe also announced Adobe Express Enterprise, which includes an enhanced version of
Adobe Express as well as Firefly. While pricing for the Express Enterprise edition has not been
disclosed, it is not outside the realm of possibilities that it could be priced similar to Creative
Cloud Express with a list price of US$36/month. Management instead talked about how it hopes
to sell site licenses of Express Enterprise, which is focused on non-creative professionals within
enterprises.
While Adobe did not add any updates to prior comments on margins, it did discuss how it
does not expect the addition of LLMs and GPT capabilities to impact operating margins. This
is very different from many other companies which are going to see margin pressure from the
compute-/GPU-intensive requirements of inference and learning.
We note that while many companies in software/cloud are focusing (often for the first time) on
margins and discussing or guiding for slowing growth, Adobe is delivering strong margins and
showing signs of improving growth going forward. While this chapter is not focused on Adobe's
margins but rather how AI will be a driver of future growth, the company has been investing in AI
and building GPT models while driving strong margins.
As other companies try to add more than basic front-end conversational AI capabilities that do not
add significant functional expansion, Adobe has quietly developed its LLMs and AI capabilities
without negatively impacting margins. Going forward, we understand, that any cost of revenue
impact from GPT and other AI capabilities will be absorbed by the company's overall margin
efficiency and organic revenue growth.
We believe this will give Adobe a competitive advantage across its business, particularly in the
Creative Cloud (where investors have been concerned about small emerging competitors) and in
Experience (Digital Marketing), especially against Salesforce, which is now trying to drive margin
improvement and likely under-investing.
ADOBE FIREFLY — ADOBE'S
NEXT-GENERATION AI-
DRIVEN CONTENT CREATION
SOLUTION
Adobe launched Adobe Firefly at the Adobe Summit and made available a beta version of the
initial offering (Exhibit 2). The reason we say initial offering is that Adobe, we understand, has
numerous additional creative capabilities that will be added to Firefly in the coming weeks and
months.
While most articles about Adobe's Firefly announcement focused on using Firefly to create
images, Firefly is much more than that. It is a next-generation content creation tool powered by
and infused with AI. It is also helping to weave more tightly together the creative process with
experience management/digital marketing. It is designed to be used by anyone wanting to create
and curate digital content assets (photo, video, 3D image, etc.), whether they are consumers/
prosumers, creative professionals, or marketing professionals. With the tight integration of Firefly
into the Experience Cloud, the whole process of creating, managing, and using content changes
and for the better, especially from the marketing perspective.
As management said during the Financial Analyst Day, rather than starting with a blank slate when
creating a content asset, Firefly can start by creating an image guided by a conversation with the
user. But Firefly already goes beyond that, and this is only the beginning.
Firefly beta, which we are playing with, includes:
BERNSTEIN
238 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
•Text Effects, which allow the user to apply styles and textures to images;
•Recolor image components (what Adobe calls vectors);
•Inpainting, which uses a brush to add, remove, or replace content from a digital asset. Note
this could include image assets or text;
•Take an image and extend the image beyond its current borders;
•Create images from 3D objects; and
•Convert sketches to full-blown images.
Adobe's vision is extremely broad, and if the company is able to deliver new capabilities at the
rate management describes, then Adobe is reinventing the creative process.
EXHIBIT 2: Example sketch using Adobe Firefly
Source: Company presentation
BERNSTEIN
GLOBAL SOFTWARE: AI IS CORE TO THE FUTURE — HIGHLIGHTS FROM ADOBE SUMMIT CONFERENCE 239
Protecting artists and creatives
One area of big concerns/complaints we have been hearing has been about OpenAI's generative
AI offerings, DALL-E, as well as many other generative AI image creation solutions), and how it
will impact the ability of artists to monetize their work. DALL-E allows one to ask the system to
create images and art. The art can be based on a specific piece of art or genre and created with a
specific medium. From the buzz we hear in the art world, DALL-E does a really good job of creating
something similar, in much the same way you can ask Bing Chat to create a poem similar to one
written by Dr. Seuss (see here).
The art created is the property, we understand, of the user (the one asking the question), and
the art the system has been trained on is basically everything available on the internet. While
Microsoft's Bing Image Creator (built on DALL-E) adds additional guardrails (e.g., stopping the
creation of destructive or problematic art), there are still concerns. The artists we have spoken
to continue to be concerned about the ramifications for the monetization of an artist's creativity
and work product.
Adobe is all about creators and creative professionals, which is their customer base, and their
concerns are Adobe's concerns. It is therefore not surprising that Adobe's Firefly is built on the
following premises:
•Using only legal-to-use and high quality content.
•Artists can opt in or out from Firefly, which again shows Adobe is focused on creative
professionals.
•Aims to get creators paid and/or their art protected. In other words, Adobe has created a way
for artists to get paid for art they did not have to create but that looks like their art if they so
desire.
•Creators/artists can also be assured that at least Adobe Firefly is not going to "create
knock-offs," allowing them to protect their brand and their art.
•Removing unwanted biases. This one, we would argue, creates a plethora of questions, but is
very interesting. In theory, Firefly's models generate stereotype-free art (e.g., no more doctors
are men and nurses are women).
ADOBE EXPERIENCE CLOUD
SUPERCHARGED WITH AI
Adobe is merging its existing Adobe AEM asset manager with AI to drive faster/better
management of digital content assets. The technology is designed to deliver more effective and
targeted advertising assets to marketing executives. While this may take advantage of some GPT
image technology from content modification (e.g. color, object, composition, and writing style), it
also uses Adobe's Sensei technology to create more effective and targeted marketing.
With many companies hyper-focused on margins, slowing marketing is often an obvious area.
Adobe sees this as a huge opportunity for the following reasons:
•Adobe is adding AI-driven analytic capabilities to its already powerful Experience Cloud
analytic solutions. Marketing professionals will be able to not only gain further insights into
what marketing is effective, but also to target the market much more exactly. Adobe's solution
will now allow users to link results to the specifics of advertising. In other words, they can
BERNSTEIN
240 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
see which colors, content, and writing styles work with different sets of their audience. The
marketers can then tune their advertising based on recommendations to drive increased
conversion of ads to monetization.
•The marketing process is inefficient for most companies. With the addition of AI-driven
analytics, and AI-driven content asset creation/tuning, ROI for clients increases and the
staffing required to drive the process decreases. Management discussed how ROI is going to
be a key differentiator and driver of revenue going forward.
VALUATION METHODOLOGY See Disclosure Appendix for valuation methodology.
EXHIBIT 3: Ratings and price targets
25-May-23 TTM
Closing Price Rel.
Ticker Rating Price Target Perf. 2022A 2023E 2024E 2022A 2023E 2024E
ADBE O USD 392.06 431 -4.10% USD 13.72 15.91 18.52 28.6 24.6 21.2
MSFT O USD 325.92 342 20.70% USD 9.21 9.74 11.16 35.4 33.4 29.2
MDB O USD 284.91 257 12.20% USD 0.85 1.01 1.8 337.2 282.2 158.4
ORCL O USD 104.29 109 41.90% USD 4.91 5.05 5.86 21.2 20.6 17.8
CRM U USD 209.91 145 29.40% USD 5.24 7.21 9.14 40.1 29.1 23
SAP.GR O EUR 120.8 154 26.30% EUR 4.07 5.04 7.21 29.7 24 16.8
SAP O USD 129.44 169 28.80% USD 4.48 5.54 7.93 28.9 23.4 16.3
SNOW M USD 147.91 148 14.30% USD 0.34 0.71 1.12 439 207.1 132.2
SPLK M USD 95.7 98 -9.40% USD -1.24 1.77 2.49 -77.2 54.2 38.4
VMW O USD 127.86 143 1.90% USD 7.25 6.59 7.93 17.6 19.4 16.1
WDAY O USD 196.41 249 34.90% USD 3.63 5.68 7.23 54.1 34.6 27.2
SPX 4,151.28 206.45 221.87 245.55 20.1 18.7 16.9
EDM 1,123.01 74.15 88.65 94.06 15.1 12.7 11.9
Adjusted P/E (x)
Adjusted EPS
Source: Bloomberg, Bernstein analysis and estimates
RISKS See Disclosure Appendix for risks.
Mark L. Moerdler mark.moerdler@bernstein.com +1 212 756 1857
Firoz Valliji firoz.valliji@bernstein.com +1 212 969 1226
Sahr Singh sahr.singh@bernstein.com +1 212 969 2521
BERNSTEIN
GLOBAL SOFTWARE: AI IS CORE TO THE FUTURE — HIGHLIGHTS FROM ADOBE SUMMIT CONFERENCE 241
BERNSTEIN
242 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
BOTTOM-UP APPROACH TO SIZING THE
LLM SILICON OPPORTUNITY
(PART 1 — INFERENCE)
HIGHLIGHTS •Unless you are living in a cave, you must be aware of ChatGPT, OpenAI's machine learning
tool that answers users' disparate questions with human-like responses. ChatGPT can answer
queries. It can write poems. It can tell stories. It may have killed the basic take-home essay
homework dead. And it has captured the fancy of investors who are all trying to figure out what
it means, how it works, how big it will be, and who will benefit, and we have been inundated
recently with questions on how to size the potential opportunity, particularly for NVIDIA whose
GPUs are used to train the neural networks responsible and handle the inferencing of the
queries themselves.
•Numerous attempts have been made by many individuals and organizations to size
the ChatGPT opportunity. However, most attempts have been, to this point, top-down,
comparing to other analogs (e.g., search), or looking at things such as current costs for GPU
time at cloud vendors. In contrast, we have gone bottom up, working through all the steps
in the transformer algorithm underpinning ChatGPT itself, examining the number of compute
operations needed to actually accomplish a ChatGPT query. Key steps include tokenization,
mapping, embeddings, positional encoding, attention, feed-forward networks, and decoding,
with attention and feed-forward steps accounting for the majority of calculations.
•We estimate almost 400 quadrillion operations are needed to accomplish a typical-
sized ChatGPT query response (say ~500 tokens, or ~2000 words). Given this, our math
suggests a GPU TAM in the multiple tens of billions of dollars annually is potentially plausible
once ChatGPT and other large language models are at scale (say, a billion queries per day,
~10% of Google's typical search volume).
INVESTMENT IMPLICATIONS We rate NVIDIA Outperform with a target price of US$475. Opportunities around datacenter, SW,
and auto remain early, and large.
WHAT IS CHATGPT? Unless you are living in a cave, you must be aware of ChatGPT, OpenAI's new machine learning
tool that answers users' disparate questions with human-like responses. ChatGPT can answer
queries. It can write poems. It can tell stories. It may have killed the basic take-home essay
homework dead. And it has captured the fancy of investors who are all trying to figure out what
it means, how it works, how big it will be, and who will benefit, and we have been inundated
recently with questions on how to size the potential opportunity, particularly for Nvidia whose
GPUs are used to train the neural networks responsible and handle the inferencing of the queries
themselves.
BERNSTEIN
BOTTOM-UP APPROACH TO SIZING THE LLM SILICON OPPORTUNITY (PART 1 — INFERENCE) 243
TRANSFORMER
ARCHITECTURES AND GPT
Let us step back and understand this domain of human-language generation, known as Natural
Language Processing (NLP). At first glance, it seems reasonably straightforward to try to do
language processing deterministically, applying various laws and axioms like in physics. However,
human beings do not abide by a specific set of rules in their own language generation, with
numerous exceptions and randomness the norm. Hence, the solution to NLP was found more in
statistical modeling and machine learning, generating language not by applying hard laws but
rather by utilizing probability distributions of characters and words in the scraped text. Deep
neural networks (DNN), a subset of neural networks, have been the preferred way to build these
language models.
Before 2017, recurrent neural networks (RNN) and even convolutional neural networks (CNN)
were the preferred DNNs to build NLP models. These had limitations. RNN works on sequential
feedback loops, which preclude parallelizations, making computations extremely slow. CNN,
though great for image recognition, does not scale well with the length of queries, with the
number of operations increasing as the distance between word positions in queries increases.
Meanwhile, researchers at Google Brain, Google Research, and the University of Toronto
developed a model architecture known as the Transformer, used originally for machine
translation for language (say German to English) or modality (say text to image). Like RNNs,
transformers process sequential input data. However, while RNNs work sequentially on input
data (feedback loop), transformers are able to process the input all at once, allowing for better
parallelization.
Transformer architecture uses what is known as an encoder-decoder architecture (Exhibit 1).
The encoder stack takes an input query and outputs a numerical representation that describes
the context of input, encoding information about which parts are related or relevant to other
parts. The decoder stack takes the encoder's understanding of context and features as well as an
"attention mechanism" (which helps capture relationships between parts of the input sequence)
to iteratively predict the output. Different permutations and combinations of these stacks can
be independently used to produce encoder-only models (e.g., BERT), decoder-only models (e.g.,
GPT), or sequence-to-sequence (both stacks) models (e.g., T5).
Decoder-only architecture (also called autoregressive) was further modified by OpenAI in its
original generative pre-trained transformer (GPT) (Exhibit 2), and subsequently in GPT-2 and
GPT-3 (the latter of which powers ChatGPT).
HOW DOES CHATGPT WORK? ChatGPT, or Chat Generative Pre-Trained Transformer, was launched in November 2022 by
OpenAI and quickly took the world by storm with its ability to provide human-like responses to a
variety of queries. ChatGPT can answer questions or write poetry or make up and tell stories. It
contains ~175 billion parameters.
Simplistically, in response to an input query, ChatGPT's function is to simply put together a
response, writing it by using its transformer architecture to predict the next word in the response
string given the prior words in the string (based on the probabilities and relationships present in
its training data) until the full response is reached.
First, the user inputs an input string, or query. The query is then converted to tokens, with one
token corresponding to ~4 characters in English (OpenAI has a Tokenizer tool to understand the
tokenization process). The input query, now tokenized, is then mapped to the model vocabulary,
BERNSTEIN
244 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
and further encoded to embed information about the relative positions and relationships of and
between all the words/tokens in the query. The results are fed forward through all the hidden
layers of the model (96 layers in this case), before being fed to the decoder, which generates
a probability distribution of potential next tokens for the response, one of which is chosen
(which through the introduction of some randomness is not necessarily the highest probability
selection). All these steps result in a selection for the "next" token in the sequence, and are then
fed back and repeated to generate the next one and the next one until the response is completed.
We note that the randomness built in to the process adds color and richness to the range of
responses (and can be seen as different answers can be given in response to the same question).
However, this mechanism also can lead to factually wrong answers as the tool is purely predictive,
and focused on language generation rather than strictly content.
EXHIBIT 1: Initial Transformer model architecture
Source: Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N.,
Kaiser, L., and Polosukhin, I. Attention is all you need. In Advances in Neural
Information Processing Systems, pp. 5998-6008, 2017, Bernstein analysis
EXHIBIT 2: Transformer model architecture in GPT
Source: Source: Radford, A., Narsimhan, K., Salimans, T., and Sutskever, I.
Improving language understanding by generative pre-training. 2018., Bernstein
analysis
~400 QUADRILLION
OPERATIONS PER QUERY?
Numerous attempts have been made by many individuals and organizations to size the
opportunity presented by ChatGPT (particularly for NVIDIA, whose GPUs are used for these and
other machine learning applications). Many sizing attempts have been to this point top-down,
comparing to other analogs (like search), or looking at things such as current costs for GPU time
at cloud vendors.
In contrast, we thought it might be more useful and insightful to go bottom up, examining the
magnitude of compute operations necessary to actually accomplish a ChatGPT query by walking
through all the steps involved in implementing the ChatGPT algorithm. The various steps in
BERNSTEIN
BOTTOM-UP APPROACH TO SIZING THE LLM SILICON OPPORTUNITY (PART 1 — INFERENCE) 245
the architecture entail tokenization, mapping, embeddings, positional encoding, attention, feed-
forward networks, and decoder, all of which are described below.
•Tokenization: Before model deployments, the input query is converted to tokens (~4
characters in English). OpenAI has a Tokenizer tool to understand the tokenization process.
•Mapping: The input query of tokens is mapped to the model vocabulary, resulting in a
vocabulary encoded matrix (M1).
•Embedding: The vocabulary encoded matrix (M1) is multiplied with embeddings weights to
transform it into the dimension of the model, resulting in embedded matrix (M2).
•Positional encoding: Positional encoding entails converting the input query positions into a
matrix with dimension of the model (M3). This step provides the model with information about
relative positions of words/tokens in the query. The embedded matrix and positional encoded
matrix are added to feed into the heart of the model, which is known as attention.
•Attention: Transformer makes use of a "self-attention" mechanism, which captures the
relationship between different tokens in the input sequence. This mechanism is a step up
from layers in prior models (RNNs and CNNs) in terms of reduced computational complexity,
parallelization of computation, and learning long-range dependencies (i.e., understanding
the relation between two words far apart in the query). GPT uses multi-head attention,
wherein the many attention functions with lower dimension are run parallel and the results are
concatenated (into a matrix M3).
•Feed-Forward Network: The output from multi-head attention (M3) is fed to a feed-forward
network wherein two weights matrices and two biases are applied.
•Repeat across hidden layers: The multi-head attention and feed-forward network, along
with the normalizations, are repeated across all hidden layers of the model (96 layers in this
case, hence these highly computational steps are repeated 96 times and result in the majority
of operations used to implement the algorithm), and then the resulting matrix (M4) is fed to
the decoder.
•Decoder: The decoder maps the output of the multi-head attention and feed-forward
network (M4) to the embeddings and vocabulary, and gives the probability of different output
tokens/words based on what came before. Interestingly, the model does not necessarily
select the highest probability token, but rather chooses one of the output tokens depending
on a statistical parameter, known as temperature. This randomness results in different
responses to the exact same question asked by the user.
Congratulations, we have just predicted the first token in the response! The process is now
repeated again and again for the length of the response. A typical response might stretch ~512
tokens (~2048 characters).
Using these steps, we have built a basic framework around matrix (vector) multiplications to
quantify the operations needed to address a typical ChatGPT query.
GPT-3 uses an input sequence of 2,048 tokens, a model dimension (i.e., number of columns
in the matrix) of 12,288, 96 attention heads, 96 hidden layers, and a feed-forward network
dimension of 4xmodel dimension. Using this, we can add up all the required operations to
BERNSTEIN
246 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
execute a ChatGPT query and response. Barring normalizations or linearizations, we estimate
the required matrix operations on the order of 7x10^14 (seven hundred trillion) operations per
token in the response, or 4x10^17 (four hundred quadrillion(!)) operations for a 2048-character
(~512 tokens) response, with the vast majority implemented in the attention and feed-forward
steps (Exhibit 3).
Two simplifications made in our estimates, which would reduce these operations, are: (1) no
sparsity efficiency (OpenAI uses Sparse Transformer that operates at O(N^1.5) instead of the
traditional O(N^2), making it more efficient), and (2) we assume that computations in calculating
one token cannot be reused for subsequent tokens in the response. Overall though, we believe
our work represents a good order-of-magnitude estimate for the computational complexity
involved in implementing ChatGPT.
EXHIBIT 3: GPT-3 performs ~7x10^14 operations for a token in your query (assuming no inference efficiency
methods such as sparsing); a typical response (~500 tokens) requires ~4x10^17 (~400 quadrillion) operations
Input Output Operations per inference
Operations per response
(512 tokens/ ~2,000
characters)
Input Matrix (2,048X1)
Vocabulary Encoded Matrix
(2,048X50,257)
2.1E+08 1.1E+11
Vocabulary Encoded Matrix
(2,048X50,257)
Sequence Embedded Matrix
(2,048X12,288)
2.5E+12 1.3E+15
Input Matrix (2,048X1)
Sequence Positional
Encoding Matrix
(2,048X12,288)
5.0E+07 2.6E+10
Sequence Embedded Matrix
(2,048X12,288)
Sequence Positional
Encoding Matrix
Positional Embedded
Encoded Matrix
(2,048X12,288)
2.5E+07 1.3E+10
Multi Head Attention
Positional Embedded
Encoded Matrix
(2,048X12,288)
Concatenated Attention
Matrix (2,048X12,288)
2.6E+14 1.3E+17
Feed Forward
Concatenated Attention
Matrix (2,048X12,288)
Feed Forward Output Matrix
(2,048X12,288)
4.8E+14 2.4E+17
Feed Forward Output Matrix
(2,048X12,288)
Output Matrix (2,048X1) 2.5E+12 1.3E+15
Input Matrix (2,048X1) Output Matrix (2,048X1) 7.4E+14 3.8E+17
Steps
Mapping
Embedding
GPT-3 model
Decode
Positional Encoding (parallel to step 1,2)
Adding Sequence Embedded, Sequence
Positional Encoded matrices
Across hidden layers
Note: n_layer=96, d_model=12288, n_heads=96, d_head=128, cntxt_win=2048, vocab=50257, response_length=512, batch=64, d_ff=4d_model
Source: Brown, T. B., Mann, B., Ryder, N., Subbiah, M., et al. Language Models are Few-Shot Learners, Bernstein analysis and estimates
GPU OPPORTUNITY FROM
CHATGPT
GPU opportunity from ChatGPT and other LLMs could be in the multiple tens of
billions of dollars annually at scale
Using the knowledge of the compute required to execute a ChatGPT query, we can now size the
silicon hardware opportunity as we know the performance (in terms of operations per second)
for the relevant NVIDIA GPUs.
NVIDIA's GPUs handle a variety of calculation precisions (Exhibit 4). AI training typically uses
floating point math (typically half-precision FP16, single-precision FP32, or double-precision
FP64). For some context, the A100 operates at ~312 TFLOPs (trillion floating-point operations
per second) at FP16 precision, and double that (~624 TFLOPs) with sparsity. However, the new
Hopper architectures has "transformer engines" designed to accomplish the relevant compute at
FP8 precision, with far higher performance (in reality, the transformer engines allow the Hopper
architecture to dynamically select and optimize between FP8 and FP16 as needed). The H100
BERNSTEIN
BOTTOM-UP APPROACH TO SIZING THE LLM SILICON OPPORTUNITY (PART 1 — INFERENCE) 247
(PCIe version) can do ~750 TFLOPs at FP16 precision (~1500 TFLOPs with sparsity) and twice
that amount at FP8 (~1500 TFLOPs, or ~3000 with sparsity), more than 6x what the A100 can
accomplish (the SXM version of the H100 can go even higher) (Exhibit 5).
Additionally, when working with trained networks, 8-bit integer math (INT8 precision) is often
used for the inference process, as it reduces the computational complexity with relatively small
loss of accuracy. But the networks trained on FP16 or higher floating-point precision require a
conversion step to recast the model's parameters into INT8 (a process known as "quantization").
However, networks trained at FP8 (using the Hopper architecture) can skip that conversion step,
with consequently significant improvements in throughput.
For our purposes, we will calculate inference compute intensity using INT8 precision on the
H100.
Additionally, key to inference is how efficiently the computation power of the GPU is used,
a concept called FLOPs utilization. The bottleneck in leveraging peak computation is the
communication bandwidth between chips, which is needed to share memory and compute.
Efficiently Scaling Transformer Inference, a research paper by Google, details the relationship
between latency (time taken to serve the query) and FLOPs utilization for test LLM models
with specific input, output token setups (Exhibit 6), with lower utilization associated with lower
latency, given the presence of greater compute headroom. We assume a ~50% FLOPs utilization
in our calculations (in other words, requiring 2x the theoretical GPU compute capacity that
would otherwise be needed). We also assume a ~75% hardware utilization (assuming the GPUs
themselves cannot be running full-out 100% of the time).
We then combine this all together with our prior math, assuming ~3x10^17 operations per query
and the relevant TFLOPs per GPU (~3,000 TOPS for H100 PCIe at INT8 precision) in order to
estimate the total amount of GPU compute that would need to be purchased for a given query
magnitude.
Pulling it together, we estimate that executing 100 million ChatGPT queries per day would
require annual purchases of H100 PCIe GPUs worth ~US$1Bn-US$2Bn, or ~US$10Bn-US
$20Bn for ~1 billion queries per day (Exhibit 7). As a sanity check on the math, we estimate the
GPU cost per query at a few cents (Exhibit 8), which seems reasonable.
For some fun, scaling this up to Google's search volume (~10 billion searches daily?) would
require likely US$150Bn+ in annual H100 GPU sales to duplicate (plus whatever other
infrastructure, networking, cooling, etc., would be needed, and of course assuming all the work
was done on GPUs).
Presumably, the cost curve will slope downward as the scale of LLM adoption rises, but the
multifold expansion of the GPU market feels inevitable if LLM usage (whether for search or
otherwise) becomes a sustained thing. But this math does suggest that an at-scale GPU TAM
opportunity from ChatGPT and other LLMs in the multiple tens of billions of dollars annually is
quite plausible.
Our model (which can handle a variety of different GPUs and precision levels) is available upon
request.
BERNSTEIN
248 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
EXHIBIT 4: NVIDIA's Ampere and Hopper GPUs are the preferred solutions for LLM compute loads
Peak performance (TFLOPS/ TOPS)
FP64
FP64 Tensor Core
FP32
Tensor Float32 [TF32] 156.0 312.0* 378.0 756.0* 494.7 989.4*
BFLOAT16 Tensor Core 312.0 624.0* 756.0 1513.0* 989.4 1978.9*
FP16 Tensor Core 312.0 624.0* 756.0 1513.0* 989.4 1978.9*
INT8 Tensor Core 624.0 1248.0* 1,513.0 3026.0* 1,978.9 3957.8*
FP8 tensor Core 1,513.0 3026.0* 1,978.9 3957.8*
33.5
66.9
66.9
H100 PCIe
H100 SXM
A100 PCIe / A100 SXM
9.7
19.5
19.5
25.6
51.2
51.2
*Sparsity
Source: Company reports, Bernstein analysis
EXHIBIT 5: Modern GPUs can provide several thousand TFLOPs of performance
312
756 989
1,513
1,979
624
1,513
1,979
3,026
3,958
0
500
1,000
1,500
2,000
2,500
3,000
3,500
4,000
4,500
A100 PCIe/ SXM H100 PCIe H100 SXM H100 PCIe H100 SXM
FP16 Tensor Core FP8 Tensor Core
TFLOPS
GPUs performance improvement with sparsity
No sparsity Sparsity
Source: Company reports, Bernstein analysis and estimates
BERNSTEIN
BOTTOM-UP APPROACH TO SIZING THE LLM SILICON OPPORTUNITY (PART 1 — INFERENCE) 249
EXHIBIT 6: FLOPs utilization during inference is a key input to GPU capacity needed, with lower latency at lower
utilization (i.e., more GPU headroom); we assume ~50% utilization for our calculations
Note: The observation in the exhibit is for running a 60 input token, 20 output token inference.
Source: Pope, R., Douglas, S., Choudhary, A., Devlin, J., Bradbury, J., Levskaya, A., Heek, J., Xiao, K., Agrawal, S., Dean, J. Efficiently Scaling Transformer Inference
2022, Bernstein analysis
EXHIBIT 7: ChatGPT would require ~US$1Bn-US$2Bn in annual H100 GPU sales for 100 million queries per day, and
~US$10Bn-US$20Bn for ~1 billion queries per day
Annual $B GPUs sales for LLM search (assuming H100 PCIe, INT8 Tensor Core* data type, lifetime 3 years)
1.3 100,000,000 500,000,000 1,000,000,000 5,000,000,000 10,000,000,000
$15,000 $1.0 $4.8 $9.6 $48.1 $96.3
$20,000 $1.3 $6.4 $12.8 $64.2 $128.4
$25,000 $1.6 $8.0 $16.0 $80.2 $160.4
$30,000 $1.9 $9.6 $19.3 $96.3 $192.5
$35,000 $2.2 $11.2 $22.5 $112.3 $224.6
Number of Search queries/ day
Cost of GPU ($)
*with sparsity
Source: Company reports, press searches, Bernstein analysis and estimates
BERNSTEIN
250 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
EXHIBIT 8: We estimate GPU cost per query to be in single digit cents range
Assumptions
Tokens per response 512
FLOPs Utilization (%) 50%
Hardware Utilization (%) 75%
Number of Queries per day 1,000,000,000
Latency 2
Batch of queries 64
Data type used in the model INT8 Tensor Core (sparsity)
GPU used H100 PCIe
Peak Performance (TOPS) 3,026
Cost of GPU ($) $25,000
Lifetime 3
Calculation of GPU Inference infrastructure cost
Number of compute operations per token inference 7.37E+14
Number of compute operations per response inference 3.78E+17
Compute capacity required for a response (#ops) 1.01E+18
Compute capacity required per batch of responses (#ops) 6.44E+19
Number of batches of queries to be inferenced per second 181
Compute capacity required (#ops) 5.83E+21
Number of GPUs required 1,925,303
Total GPU investment ($B) $48.1
Annual GPU investment ($B) $16.0
Number of queries serviced by the GPUs in lifetime 1.10E+12
GPU cost per query ($) $0.04
Source: Company reports, Bernstein analysis and estimates
VALUATION METHODOLOGY See Disclosure Appendix for valuation methodology.
EXHIBIT 9: Ratings and target prices
Ticker Rating Currency
25-May-2023
Closing Price
Target
Price
NVDA O USD $379.80 $475.00
SPX $4,151.28
Source: Bloomberg, Bernstein analysis and estimates
RISKS See Disclosure Appendix for risks.
Stacy A. Rasgon stacy.rasgon@bernstein.com +1 212 756 4403
Akhilesh Kumawat akhilesh.kumawat@bernstein.com +1 212 969 1308
Alrick Shaw alrick.shaw@bernstein.com +1 212 969 1458
BERNSTEIN
BOTTOM-UP APPROACH TO SIZING THE LLM SILICON OPPORTUNITY (PART 1 — INFERENCE) 251
BERNSTEIN
252 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
BOTTOM-UP APPROACH TO SIZING THE
LLM SILICON OPPORTUNITY
(PART 2 — TRAINING)
HIGHLIGHTS •"Training" a neural network model is relatively simple in concept but very complex
in practice. Simplistically, it involves passing input data forward and backward through the
model, changing the model parameters each time while measuring the output against a real
(expected) output, and minimizing the delta through careful selection of those parameters.
In practice, however, this requires huge amounts of compute. Training ChatGPT requires
~3x1023 operations (~300 sextillion) with ~175 billion parameters; newer (larger) models
would require far more. We have generalized our calculations to be applicable to other
transformer models once specs are known (for example OpenAI has not released GPT-4
specs, but is rumored to use a trillion parameters). Our sizing model is available on request.
•However, the LLM training TAM does not appear all that big today (tens to hundreds of
millions of dollars annually?), quite a bit smaller than LLM inference (which we sized in
the multiple tens of billions) as training does not scale the same (i.e., the model, once trained,
works, and if retraining is needed, the infrastructure is already in place). But models are getting
bigger, and we are early on the adoption curve; over time, we believe the TAM could potentially
reach into the billions as model complexity grows and adoption widens.
•While the debate is still on, we believe that small language models will democratize
and accelerate AI deployments. But we also believe broader adoption of AI, in all its
forms, is likely to be a net positive for NVIDIA.
INVESTMENT IMPLICATIONS We rate NVIDIA Outperform with a target price of US$475. Opportunities around datacenter, SW,
and auto remain early, and large.
HOW IS LLM TRAINING
DIFFERENT FROM
INFERENCE?
LLM inference is about leveraging a model to predict the next word in a response based on the
previous words in the response, building off a large dataset that was used to "train" the model.
In other words, the model already knows what to do. On the other hand, "training" the model is
about helping it learn what to do, i.e., we need to find settings (parameters) that make the model
learn a "generalization rule" from a given set of training examples.
What does it mean to "train" a model?
The idea of "training" a model is simple in concept, but can be devilishly complex in practice. But at
its highest level, it can be thought of as an optimization. Imagine a black box with many (billions?)
of knobs ("parameters" in our parlance) on the side, a slot on one end to feed in information
("inputs"), and a screen on the other side to display results ("outputs"). Training involves feeding
in numerous samples of known inputs and observing the outputs, and then turning the knobs
(i.e., changing the parameters) until the output itself coincides with expectations. At that point
the model is "trained." We lock down the knobs and use the box to generate outputs based on
BERNSTEIN
BOTTOM-UP APPROACH TO SIZING THE LLM SILICON OPPORTUNITY (PART 2 — TRAINING) 253
any input we wish to feed in as the model (through the training process) has now extracted and
generalized the rule, and outputs the correct answers for an out-of-sample input.
The structural form of these types of models tends to be a neural network. These networks
are typically represented schematically by a set of circles, connected with arrows and lines,
with each circle representing a single neuron in the network (Exhibit 1). In layman's terms, these
neurons can be thought of as calculators, which take in a number of inputs, and produce an output
depending on what those inputs are (for example an artificial neuron could be programmed to
output a "1" if the sum of its inputs is above a certain threshold, and a "0" if the sum of the inputs
is below that threshold).
The process through which a neuron translates inputs to outputs is known as the "activation
function." In the above example, this might be a simple binary step function ("0" or "1" depending
on the threshold selected), though in practice different types of activation functions can be used.
The output of a given neuron can change depending on the inputs received to it. But we don't
need to treat each input as holding equal importance; in other words, we can weight the relative
contribution of each individual input into a given neuron (known as the "weights" of the network),
and these weights can vary. We can also change the threshold required for the neuron to make
its decision (this threshold is sometimes known as the "bias" of the neuron) (Exhibit 2).
It is common to utilize more than one layer of neurons, with the output of one layer becoming the
inputs to the next. These extra layers in a network are known as "hidden" layers, and the layout
and design of these hidden layers is one of the more complex pieces of designing a network.
The parameters of the model are these various "weights" and "biases," and individual model
structures can grow staggeringly complex (the GPT-3 model used for ChatGPT has 175 billion
different parameters). "Training" the model involves varying and fine-tuning these individual
weights and bias of the network until the desired output is achieved.
By way of example, the network shown in Exhibit 1 depicts a neural network with an input
layer with three neurons, three hidden layers with four neurons each, and an output layer with
three neurons. The outputs of each neuron are connected to the neurons in the next layer. Each
input into each layer will have a "weight" assigned to it (therefore, at each layer there will be
a vector of weights corresponding to the importance of each input), and each neuron has an
activation function associated with it. Mathematically, each layer of neurons has a vector of
outputs associated with it (the outputs in layer n are represented by the vector An). An itself
depends on the activation functions of the neurons in layer n (Zn), which itself is a function of
the output vector of the n-1 layer of neurons, and the weights and biases in the current layer n
(mathematically, we can write Zn=(An-1*Wn) + Bn, where A is the output vector of a neuron layer,
W is the weight matrix between two neuron layers, and B is the bias vector of neuron layer).
Training involves passing calculations forward and backward through the network, changing the
parameters each time while measuring the final output (output vector A4 in this case) against a
real output. One can think about this measurement as a cost function, with the goal of the training
to minimize it (i.e., make the difference between the final output and real sample as small as
possible) by careful selection of the parameters in the network.
BERNSTEIN
254 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
In practice, this works by first using a forward pass, feeding a sample (with three tokens in
the example) to the model, ultimately resulting in an output vector from the fourth and final
layer of neurons (A4). This output is measured against a real output, Y in terms of a loss/cost
function (think of it simplistically as absolute value of the difference Y-A4). This loss/cost is then
minimized by calculating gradients (i.e., change in loss versus change in the weight and bias
parameters) by working backward through the model (a backward pass), and then changing
the parameters (by a factor called learning parameter) to minimize the loss (this process is called
"backpropagation"). The process (forward and backward) is then repeated, again and again,
until the model converges.
EXHIBIT 1: Schematic of a simple neural network
𝑎1
1
𝑎2
1
𝑎3
1
𝑎4
1
𝑎1
2
𝑎2
2
𝑎3
2
𝑎4
2
𝑎1
3
𝑎2
3
𝑎3
3
𝑎4
3
𝑎1
4
𝑎2
4
𝑎3
4
W1W2W3W4
B1B2B3
x1
x2
x3
Note: x: input, Wn: weights from layer n-1 to n, Bn: Biases in layer n, an: outputs of activation functions in layer n
Source: Bernstein analysis
EXHIBIT 2: Artificial neuron
Source: Bernstein analysis
BERNSTEIN
BOTTOM-UP APPROACH TO SIZING THE LLM SILICON OPPORTUNITY (PART 2 — TRAINING) 255
Calculating the number of operations during training of a transformer-based
LLM
Training a neural network begins with a training dataset. Datasets (~300 billion tokens in GPT3)
are broken into samples of context window length of the model (2,048 tokens in GPT-3). These
samples are passed through the model (forward pass), and the model output is then compared
with the expected output. This error or loss is minimized through a gradient descent optimization
algorithm (GPT-3 uses ADAM), which needs a back propagation algorithm (backward pass). After
a certain number of passes (batch size), the parameters (weights and biases) are changed by a
factor (called the learning parameter), and the loss (the difference between the model output and
a known desired output) is measured, all utilizing a certain compute capacity (in FLOPs), a dataset
and model hyper parameters (learning rate, batch size, model width, and model depth).
In GPT-3, training begins with a batch size of 32K tokens from the training dataset, which is
gradually increased to 3.2 million tokens over the first ~12 billion tokens used. The learning
parameter is linearly increased to 6×10-5 over the first 375 million tokens and then decreased
(by a cosine decay) to ~6*10-6 by 260 billion data tokens (Exhibit 3).
Loss (i.e., the difference between the model output and the expected output) is calculated during
the learning process, growing smaller and smaller as training progresses. Using more and more
compute (i.e., training longer and longer) reduces the measured loss (i.e., creates a better trained
model), but there is a limit to everything. In fact, loss has been found to eventually approach
a power-law versus compute as more compute is added (Exhibit 4), giving a measure of the
appropriate efficient stopping point for a given model. The loss is calculated through the batches
until this efficient compute frontier is reached.
Hence, we can break down the problem of calculating the number of operations needed to
train a transformer-based LLM in three steps: (1) calculate the number of passes (forward and
backward) to train the model, (2) calculate the number operations in a forward pass, and (3)
calculate the number of operations in a backward pass.
Number of passes through the transformer
If we know the training datasets available, the dataset size (in tokens) and the number of times
a particular dataset is used (known as "epochs"), and the input context window of the model (in
tokens), we can estimate the number of passes made to train the model.
For GPT-3, an overall training dataset of ~500 billion tokens was passed through ~0.6 epochs (in
other words, the model saw ~60% of the full training dataset). This results in training the model
over an effective ~300 billion token dataset (500B x 0.6) using ~143 million (~1.4x108) passes
(~300 billion tokens taken in windows of 2,048 tokens each, with each pass done forward and
backward) (Exhibit 5).
Number of operations in a forward pass
A forward pass entails the number of operations needed to predict the next token given a query
of input tokens. In fact, we already did this calculation (which corresponds to the number of
operations required to pass a single token through the model during a query response). Details
are in our "NVIDIA (NVDA): A bottoms-up approach to sizing the ChatGPT opportunity" research
note. As a reminder, for GPT-3, we calculate the number of operations in a single forward pass
BERNSTEIN
256 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
as 7.4 × 1014 operations (Exhibit 6).
Number of operations in a backward pass
A backward pass entails using a backpropagation algorithm to calculate the gradients of the
desired cost function with regard to our parameters (in fact we can think of cost function as the
mean squared error of the model results).
The math gets somewhat complex here, so please feel free to skip down a few paragraphs if
vector differential equations are not your thing. But the backpropagation algorithm comprises
four key equations:
1. dZ[L] = (dA[L])(.*) g'[L](Z[L]), where dZ[L] is the error in neurons in last layer of network/
output layer, dA[L] is the gradient of cost function w.r.t. output of last layer L, g'[L] is the
derivation of activation function in last layer, and (.*) is element wise product or Hadamard
product.
2. dZ[l-1] = (dZ[l]*W[l]T) (.*) g'[l](Z[l-1]), where dZ[l] is the error in Z of each neuron in
layer l, W[l] is the weights matrix connecting neuron layer l-1 to neuron layer l, g'[l] is
the derivative of activation function in layer l, (.*) is element wise product or Hadamard
product.
3. dW[l] = ( A[l-1]T* (dZ[l] )/m, where dW[l] is gradient of cost function w.r.t. weights
connecting layer l-1 to layer l, the A[l] is the output of layer l, m is the batch size, i.e., the
number of training samples before updating the parameters.
4. dB[l] = sum (dZ[l])/m, where dB[l] is the gradient of cost function w.r.t. biases in neuron
layer l.
After each batch size (m training samples), parameters are updated as:
5. W[l] = W[l-1] - e*dW[l]; B[l] = B[l-1] - e*dB[l], where e is the learning rate. This iteration
is done for multiple batches with each batch containing m training samples.
In GPT-3, we use cross entropy (type of log loss function) as the cost function and softmax (type of
sigmoid function) as the activation function giving probabilities of potential output tokens. These
further simplify equations 1 and 2 to:
6. dZ[L] = A[L] - Y, where Y is the known result of a training sample.
7. dZ[l-1] = (dZ[l]*W[l]T) (.*) A[l-1](.*)(1-A[l-1])
Using these equations, we can estimate the number of operations needed for one backward pass
of a GPT-3 model; we come out at ~1.4×1015 operations (Exhibit 7). This ties in well with multiple
research papers implying backward operations as ~2x of forward operations in the model.
BERNSTEIN
BOTTOM-UP APPROACH TO SIZING THE LLM SILICON OPPORTUNITY (PART 2 — TRAINING) 257
EXHIBIT 3: Learning rate in GPT linearly ramps up to base rate (6x10^-5) until ~375 million tokens and then
decreases by cosine decay until ~260 billion tokens
0.0E+00
1.0E-05
2.0E-05
3.0E-05
4.0E-05
5.0E-05
6.0E-05
7.0E-05
0 20,000 40,000 60,000 80,000 100,000 120,000 140,000
Learning rate
Cumulative number of batches
Learning rate vs Cumulative number of batches
Source: Brown, T. B., Mann, B., Ryder, N., Subbiah, M., et al. Language Models are Few-Shot Learners, Bernstein analysis and estimates
EXHIBIT 4: Models are trained with examples until they hit the efficient compute frontier (where loss begins to
scale as a power function with the amount of compute)
Note: One PetaFLOP/s-day is the amount of compute one computer running at 1 petaFLOP per second for one day could handle; in other worlds 10^15 FLOPS/
s*60 s/min*60min//hr*24 hr/day =8.64*10^19 FLOPS = 86.4 exaFLOPS.
Source: Brown, T. B., Mann, B., Ryder, N., Subbiah, M., et al. Language Models are Few-Shot Learners, Bernstein analysis and estimates
BERNSTEIN
258 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
EXHIBIT 5: GPT3 uses ~300 billion tokens with ~143 million passes made through the model in each direction
(forward and backward passes)
Dataset
Quantity available
(Bn tokens)
Epochs
Quantity used
(Bn tokens)
Passes made
through model
(Mn)
Common Crawl 410 0.4 180 88
WebText2 19 2.9 55 27
Books1 12 1.9 23 11
Books2 55 0.4 24 12
Wikipedia 3 3.4 10 5
Other 0 0.0 0 0
Total 499 0.6 292 143
Note: Input context window = 2,048 tokens
Source: Brown, T. B., Mann, B., Ryder, N., Subbiah, M., et al. Language Models are Few-Shot Learners, Bernstein analysis
EXHIBIT 6: Training GPT-3 requires ~7.4x1014 operations for a forward pass (assuming no inference efficiency
methods like sparsing)
Input Output Operations
Input Matrix (2,048X1)
Vocabulary Encoded Matrix
(2,048X50,257)
2.1E+08
Vocabulary Encoded Matrix
(2,048X50,257)
Sequence Embedded
Matrix (2,048X12,288)
2.5E+12
Input Matrix (2,048X1)
Sequence Positional
Encoding Matrix
(2,048X12,288)
5.0E+07
Sequence Embedded
Matrix (2,048X12,288)
Sequence Positional
Positional Embedded
Encoded Matrix
(2,048X12,288)
2.5E+07
Multi Head
Attention
Positional Embedded
Encoded Matrix
(2,048X12,288)
Concatenated Attention
Matrix (2,048X12,288)
2.6E+14
Feed Forward
Concatenated Attention
Matrix (2,048X12,288)
Feed Forward Output
Matrix (2,048X12,288)
4.8E+14
Feed Forward Output
Matrix (2,048X12,288)
Output Matrix (2,048X1) 2.5E+12
Input Matrix (2,048X1) Output Matrix (2,048X1) 7.4E+14
Steps
Mapping
Embedding
GPT-3 model
Decode
Positional Encoding (parallel to
step 1,2)
Adding Sequence Embedded,
Sequence Positional Encoded
matrices
Across hidden
layers
Note: n_layer=96, d_model=12288, n_heads=96, d_head=128, cntxt_win=2048, vocab=50257, response_length=512, batch=64, d_ff=4d_model
Source: Brown, T. B., Mann, B., Ryder, N., Subbiah, M., et al. Language Models are Few-Shot Learners, Bernstein estimates and analysis
BERNSTEIN
BOTTOM-UP APPROACH TO SIZING THE LLM SILICON OPPORTUNITY (PART 2 — TRAINING) 259
EXHIBIT 7: Training GPT3 requires ~1.4x1015 operations for a single backward pass (~2x the requirements of a
forward pass)
Steps Operations
Calculate backpropagation errors
across model
7.7E+14
Calculate gradients of cost function
w.r.t. parameters
5.9E+14
Other (updating moments (means and
variances), parameters)
Negligible (much
lower order)
Total
1.4E+15
Source: Bernstein analysis and estimates
Total operations to train the model? ~300 sextillion
We can now estimate the total operations needed to train ChatGPT by multiplying the total
number of passes through the model by the total operations involved in a single forward and
backward pass. Therefore, we estimate total operations needed to train at 1.4x108 passes x
(7.4x1014 forward operations per pass +1.4x1015 backward operations per pass) = ~3.0x1023
(~300 sextillion) total operations (Exhibit 8).
This ties in well with the actual number (3.1×1023 operations) OpenAI released. However, our
methodology can be generalized to other models if we know or can estimate their characteristics.
We note that GPT-1 required ~2.1x1020 operations to train with ~124 million parameters; GPT-2
required ~2.8x1021 operations to train with ~1.6 billion parameters, and GPT-3 at ~3x1023
operations with ~175 billion parameters (Exhibit 9). We do not know the specifications of GPT-4
yet (OpenAI has not released them publicly), but it has been rumored to contain ~1 trillion
parameters, which is easily possible (Exhibit 10) and, indeed, would represent only a ~6-fold
increase versus GPT-3, small relative to prior transitions. But a trillion parameters would likely
require something on the order of ~2x1024 operations to train (~2 septillion) (Exhibit 11).
BERNSTEIN
260 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
EXHIBIT 8: GPT3 was trained with ~3x1023 operations
GPT 3
Number of operations in forward pass 7.4E+14
Number of operations in backward pass 1.4E+15
Number of passes made during training 1.4E+08
Total number of operations 3.0E+23
Source: Bernstein analysis and estimates
EXHIBIT 9: OpenAI's parameters have expanded from ~120 million parameters in the original GPT to ~175 billion
parameters in GPT3
Model Layers
Model
dimension
Parameters Total ops
GPT 1 12 768 1.24E+08 2.12E+20
GPT 2 48 1600 1.55E+09 2.79E+21
GPT 3 96 12288 1.75E+11 3.00E+23
Source: Radford, A., Narsimhan, K., Salimans, T., and Sutskever, I. Improving Language Understanding by Generative Pre-training. 2018, Bernstein analysis and
estimates
EXHIBIT 10: Model parameters are largely a function of model size and number of layers
Model parameters (B) are largely a function of model dimension and number of layers
175 3,072 6,144 12,288 24,576 49,152
10 1 5 19 74 292
50 6 23 91 364 1,452
100 11 46 182 726 2,902
500 57 227 907 3,625 14,498
1,000 113 453 1,813 7,249 28,993
Model dimension (dmodel = 12,288 for GPT3)
Number of layers in the
model (n=96 for GPT3)
Source: Bernstein analysis and estimates
BERNSTEIN
BOTTOM-UP APPROACH TO SIZING THE LLM SILICON OPPORTUNITY (PART 2 — TRAINING) 261
EXHIBIT 11: Training operations would also scale as parameters increases
Training operations as a function of model size and number of layers
############## 3,072 6,144 12,288 24,576 49,152
10 2.2E+21 8.2E+21 3.2E+22 1.3E+23 5.0E+23
50 1.0E+22 4.0E+22 1.6E+23 6.2E+23 2.5E+24
100 2.0E+22 7.9E+22 3.1E+23 1.2E+24 5.0E+24
500 1.0E+23 3.9E+23 1.6E+24 6.2E+24 2.5E+25
1,000 2.0E+23 7.9E+23 3.1E+24 1.2E+25 5.0E+25
Number of layers in the
model (n=96 for GPT3)
Model sizes (dmodel = 12,288 for GPT3)
Source: Bernstein analysis and estimates
How big is the GPU market for training LLM?
Unlike inference, sizing the LLM training market is less straightforward even once we have the
operation count. This is because inference (i.e., actually running queries) scales directly; in other
words, the more queries we serve, the more compute (and hence GPUs) we need.
In contrast, the training market does not scale the same way. Rather, once a model is built, it
works, and if it needs to be retrained, the infrastructure to do so is already there. One could
imagine, for instance, a cloud vendor building a single LLM, and retraining or rebuilding it once a
month on updated data. In that case, the market value would still be for only a single LLM (they
would not need to repurchase the compute every month as it would already be installed), and
those assets would remain suitable until the increasing complexity of new models overextends
them. It therefore makes sense to us to instead try to size this market as a function of the number
of companies that might build LLMs, and how many trainings per year they might need to do;
equivalently, the latter can be couched as asking how many days it takes to train a given LLM.
As a simple example, imagine we have one entity building one LLM, retraining or rebuilding it
every month (or 12 trainings per year), requiring some baseline of compute to accomplish this. It
would take 2x this amount of compute for this LLM to train their model every 15 days instead (i.e.,
24x per year). This would also be equivalent to the entity building two LLMs, retraining each of
them every month (again, 24 trainings per year). And it would be equivalent to two entities, each
building one LLM, each retraining it every month (24 trainings per year). Overall, we could think
about the market scaling with the total number of "trainings" performed globally on an annual
basis.
For some context, ChatGPT was rumored to have been trained on ~10,000 NVIDIA V100 GPUs in
~15 days. Our rough modeling (given the known specifications for the V100) suggests something
similar, ~3500 V100's required assuming the GPUs were running flat-out for the entire 15-day
period at peak performance at FP16 precision (so 10K GPUs in real life, which is never as perfect
as theory, seems reasonable to us). Of course, the V100 is fairly old, and newer GPUs (A100 and
H100), while more expensive, are also far more efficient (the A100 added support for sparsity,
and the transformer engine on the H100 allows training at FP8 precision). This should enable far
fewer newer GPUs to be used to train a ChatGPT-like LLM, and correspondingly lower cost (with
the H100 likely 90% cheaper, or more, versus the V100; Exhibit 12). But, depending on the days
BERNSTEIN
262 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
to train and the CPU cost, we estimate a single organization's single ChatGPT-like LLM could be
trained for under US$10Mn (perhaps even under US$5Mn) using the latest H100 GPUs (Exhibit
13).
Overall, we are inclined to conclude the GPU training market for LLMs is likely not all that huge
as it stands today (probably in the tens to hundreds of millions of dollars annually at the current
model sizes) (Exhibit 14), significantly smaller than our estimate for the likely LLM inference TAM
(which we believe could be in the multiple 10's of billions of dollars annually).
However, that is where we stand today, and over time we think training TAM will increase
as well, as the models are getting bigger and bigger, and adoption overall grows. Whereas a
single ChatGPT model (with 175 billion parameters) might be trainable for a few million dollars
on current state-of-the-art hardware, a single trillion-parameter model (GPT-4?) might require
5-10x that investment, and it seems probable that training costs will increase commensurate
with growth in model size (Exhibit 15). Given those trends, we believe the LLM training market
could grow into the billions of dollars over time (Exhibit 16) as model complexity grows and
adoption increases.
Our models (for both training and inference of LLMs) are available for interested clients.
EXHIBIT 12: ChatGPT was rumored to have been trained on ~10K V100 GPUs; our theoretical calculation is in the
right order of magnitude; additionally, higher efficiencies on newer GPUs could result in far lower training costs
GPU V100 A100 A100 H100 H100 H100 H100
Data Type FP16 FP16 FP16 FP16 FP16 FP8 FP8
Sparsity? N N Y N Y N Y
Peak Performance (TOPS) 112 312 624 756 1513 1513 3026
Days to Train 15 15 15 15 15 15 15
Number of GPUs 3444 1236 1236 510 255 255 127
Cost/GPU $8,000 $10,000 $10,000 $30,000 $30,000 $30,000 $30,000
GPU Cost per Model ($M) $27.6 $12.4 $12.4 $15.3 $7.7 $7.7 $3.8
Note: Assumes 60% FLOPS utilization and 100% GPU utilization (assuming a single model being trained one time)
Source: Company reports, press searches, Bernstein analysis and estimates
EXHIBIT 13: A single organization's single ChatGPT-like LLM could likely be trained for under US$10Mn
$M GPUs sales needed to train a single LLM model (assuming H100 PCIe, FP8 Tensor Core* data type)
4 $15,000 $20,000 $25,000 $30,000 $35,000
1$29 $38 $48 $57 $67
5$6 $8 $10 $11 $13
10 $3 $4 $5 $6 $7
15 $2 $3 $3 $4 $4
30 $1 $1 $2 $2 $2
Days to Train
Cost of GPU ($)
*with sparsity, 60% FLOPSs utilization rate, and 100% hardware utilization rate (assuming a single model being trained one time)
Source: Company reports, press searches, Bernstein analysis and estimates
BERNSTEIN
BOTTOM-UP APPROACH TO SIZING THE LLM SILICON OPPORTUNITY (PART 2 — TRAINING) 263
EXHIBIT 14: LLM training opportunity with GPT3-like models today is likely in 10s-100s of US$Mn (not huge)
Annual $M GPUs sales for LLM search (assuming H100 PCIe, $30K, FP8 Tensor Core* data type, lifetime 3 years)
26 1 5 10 50 100
1$38 $191 $383 $1,913 $3,825
5$8 $38 $77 $383 $765
10 $4 $19 $38 $191 $382
15 $3 $13 $26 $128 $255
30 $1 $6 $13 $64 $127
Days to train the model
Number of entities building LLM models
*with sparsity, 60% FLOPs utilization rate, and 50% hardware utilization (half the utilization of a single model training as we figure resources are unlikely to remain
constantly fully utilized)
Source: Company reports, press searches, Bernstein analysis and estimates
EXHIBIT 15: However, GPU training infrastructure costs are likely to go to higher as models get bigger
$M GPUs sales needed to train a single LLM mode (assuming H100 PCIe, $30K, FP8 Tensor Core data type)
$15,000 $20,000 $25,000 $30,000 $35,000
100 $1 $1 $2 $2 $3
500 $5 $7 $9 $11 $13
1,000 $11 $15 $18 $22 $26
5,000 $55 $73 $91 $109 $127
10,000 $109 $146 $182 $218 $255
Number of parameters (Bn)
in the model
Cost of GPU ($)
*with sparsity, 15 days to train the model, 60% FLOPS utilization, and 100% hardware utilization (a single model being trained one time)
Source: Company reports, press searches, Bernstein analysis and estimates
EXHIBIT 16: As models grow bigger and adoption increases, we believe LLM training market could potentially grow
into billions of dollars over time
Annual $M GPUs sales for LLM search (assuming H100 PCIe ($30K), FP8 Tensor Core* data type, lifetime 3 years)
3 310 25 50 100
25 $3 $9 $22 $44 $87
200 $21 $70 $175 $349 $699
500 $52 $175 $437 $873 $1,747
2000 $210 $699 $1,747 $3,493 $6,986
5000 $524 $1,747 $4,366 $8,733 $17,465
Number of trainings completed
per year
Number of operations needed to train the model (x10^23)
*with sparsity, 60% FLOPS utilization, and 50% hardware utilization. For some context, ~24 trainings per year would be equivalent to one entity, with one model,
retrained every ~15 days.
Source: Company reports, press searches, Bernstein analysis and estimates
BERNSTEIN
264 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
An aside — the advent of small language models?
OpenAI concluded in its research Scaling Laws for Neural Language Models (Kaplan et al.) that
training performance has a power law relationship with each of the scale factors viz. parameters,
training dataset, and the compute allocation for training (Exhibit 17). Hence, the problem for
developers/builders of these models breaks down to: Given the compute budget, how should
one scale the training dataset and/or the model size (parameters)?
As per the OpenAI research, it was argued that, given a 10x increase in compute facility, the model
should scale by ~5x parameters and training dataset by ~2x. However, DeepMind retested the
hypothesis in their research Training Compute-Optimal Large Language Models (Hoffman et al.)
and found that the model parameters and training dataset should scale equally, given a particular
compute capacity. This research provides a strong basis for building smaller models trained on
larger datasets that may need a similar training compute, but significantly reduce the inference
compute needed, thereby democratizing inference hardware deployment, be it in cell phones
(say Apple) or automobile (say your Tesla car) or the broader IoT (all things NVIDIA can target).
Meta (covered by Mark Shmulik) further leveraged this idea of the largest training dataset and
fewer parameters to train its LLaMA models, demonstrating a better performance on certain
benchmarks than the conventional LLMs (Exhibit 18 and Exhibit 19). Moreover, a new model type
called the retrieval augmented language model outsources large scale memorization, typically
implemented through parameters in LLMs. These models retrieve relevant documents based on
the input prompt and produce the output. Atlas from Meta and Retro from DeepMind are two
examples of such retrieval-based models.
While the debate is still on, we believe small language models will democratize and accelerate
AI deployments. But, we also believe broader adoption of AI in all its forms is likely to be a net
positive for NVIDIA.
A bit of housekeeping
We rate NVIDIA Outperform with target price of US$475. Everyone has been looking for ways to
play AI that aren't as expensive as NVIDIA, given the run this year. However, perhaps NVIDIA itself
is the best way to accomplish that (while still undeniably pricey, it is clearly not quite as expensive
as it looked), and the narrative, backed by actual products and sales, still has legs, in our opinion.
BERNSTEIN
BOTTOM-UP APPROACH TO SIZING THE LLM SILICON OPPORTUNITY (PART 2 — TRAINING) 265
EXHIBIT 17: Language models' performance has a power law relationship with scale factors viz. compute, training
dataset, and parameters
Source: Scaling Laws For Neural Language Models, Kaplan et al., Bernstein analysis
EXHIBIT 18: Meta's LLaMA models are smaller in size, but trained on larger datasets
Source: LLaMA: Open and Efficient Foundation Language Models, Touvron et al., Bernstein analysis
EXHIBIT 19: LLaMA small models have fared better than LLMs on multiple benchmarks
Source: LLaMA: Open and Efficient Foundation Language Models, Touvron et al., Bernstein analysis
BERNSTEIN
266 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
VALUATION METHODOLOGY See Disclosure Appendix for valuation methodology.
EXHIBIT 20: Ratings and target prices
Ticker Rating Currency
25-May-2023
Closing Price
Target
Price
NVDA O USD $379.80 $475.00
SPX $4,151.28
Source: Bloomberg, Bernstein analysis and estimates
RISKS See Disclosure Appendix for risks.
Stacy A. Rasgon stacy.rasgon@bernstein.com +1 212 756 4403
Akhilesh Kumawat akhilesh.kumawat@bernstein.com +1 212 969 1308
Alrick Shaw alrick.shaw@bernstein.com +1 212 969 1458
BERNSTEIN
BOTTOM-UP APPROACH TO SIZING THE LLM SILICON OPPORTUNITY (PART 2 — TRAINING) 267
BERNSTEIN
268 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
WILL AI KILL GOOGLE SEARCH?
HIGHLIGHTS •We started the year with a doozy diving into whether AI chatbots such as ChatGPT
are on the fast-track to disrupt and displace Google Search. If you had spent any time
on Twitter in the first few months of the year, the end of Google Search was a forgone
conclusion. Those ominous armchair analyst predictions only got louder with the New York
Times reporting that Google held an impromptu "Code Red" meeting over Christmas week to
deal with ChatGPT's exponential growth.
•Compounding the issue at hand is that the buzz around a search killer is happening
right as the digital ad sector faces a major secular crisis on whether the market has
matured. A mature digital ad market and a search replacement? Explains why Google's stock
was down -10% by mid-January since the launch of ChatGPT (Meta's stock was up +14%
over this same period, so let's just agree that it's mostly ChatGPT). Then an 800-lb gorilla in
Microsoft launched a slick PR campaign announcing its intention to go after Google's
search empire with an all new Bing, complete with a conversational AI assistant and plans
to integrate this new feature across the Microsoft application suite. Google quickly tried to
respond with its own AI product, Bard, in a hastily put together Paris event that failed to
impress investors, causing a -8% stock move. Since then, the panic has somewhat subsided,
Bing share gains have been negligible, and Google's own IO event more recently showed off
its slick AI presentation.
•It would be sadly ironic if Google — who was busy telling anyone who'd listen during
the metaverse bubble that it was an AI company — ends up getting disrupted by AI. But
precisely the point, Google hasn't been caught with its pants down here. For years the
company has been stockpiling AI talent (a very expensive endeavor), collecting and
enriching datasets, and building the architecture, tools, and AI models in anticipation
of ushering in the era of machines when needed. It is certainly not the first time Google has
had to face down would-be search killers. There was the desktop-to-mobile migration, the
rise of vertical-specific search engines such as Amazon, and even the very short-lived voice
assistant disruption. So is this just another stone in the road for Google'S Search tank
to run over, or will AI push Google's Search kingdom to the brink of collapse?
INVESTMENT IMPLICATIONS We rate Google Outperform with a target price of US$125.
“THIS IS GOING TO CHANGE
EVERYTHING!”
In the few years I've been in this, we've already gone through a few hype cycles. I had a front row
seat (and a few dollars) into whatever the crypto/web3/NFT movement was, I moved my family
to Miami at peak NYC exodus, and you may recall a certain note my colleagues and I published
toward the end of 2021 signaling the peak on The Metaverse (Bernstein enters the metaverse:
A primer on what it is, the size of the prize, and why you should care) — which is the second
most read Bernstein note in the last five years.
The fallout is real. We've now got a Meta instead of Facebook, with investors rightly skeptical
of that company's investment level in an idea few believe in. Valuations across the Internet
BERNSTEIN
WILL AI KILL GOOGLE SEARCH? 269
landscape have been slashed with investors promising greater discipline to avoid similar greed-
fueled bubbles. Whether nature or nurture, tech companies must have really short memories
as Microsoft (covered by Mark Moerdler) prepares to invest ~US$10Bn into the buzzy ChatGPT
creator, OpenAI. The deal would value OpenAI at ~US$29Bn, rich for a company whose
breakthrough product is months old with maybe a few hundred million in revenues and losses
multiples of that number.
Disruption feels inevitable in the Internet space, yet you can easily go broke betting on all the
"next big thing" false starts. AI as an investment theme is once again en vogue in venture capital
land, replacing the crypto/web3 movement, autonomous vehicles, micro-mobility, and "x-as-a-
service" before that. Where there's a flame, there's venture capital fuel waiting to be poured.
Some of these investment themes are ridiculous in hindsight — the rush into quicker delivery
turnaround windows (15 minutes?) reminds me of the one-upping we saw in the 1990s for how
little time you needed to get rock hard abs (8-minute abs!).
Unlike the hype cycles above, ChatGPT (text), DALL-E (images), and more broadly the generative
AI underpinnings that these models rely upon have an impressive proof of concept that anyone
can try out. I was blown away by DALL-E since I suck at art, and impressed with ChatGPT's
creativity. These applications have certainly been fun to play around with, but when I tried to query
classic structured and monetizable searches, the shortcomings were glaringly obvious. Bob's
job is safe for now (Bernstein Energy & Power: Why I (and not ChatGPT) should be your trusted
analyst).
Now, that's not to say that generative AI isn't the next big thing. I don't want this chapter shared
in 20 years as an example of what idiots equity research analysts are. I can see plenty of end
markets that can be created around increasing productivity, disrupting and streamlining creative,
and even another go at the "smart" home by improving those voice assistant devices to be used
for more than checking the weather and playing your favorite song. But there's a long way to go
before we declare this the end of Google's search dominance. My counter-intuitive hypothesis
so far is that it is just as likely that generative AI opens up new end-markets expanding the
search TAM as it is that is disrupts the existing search ecosystem.
In Google's case, the buzz around a search killer is happening right as the digital ad sector faces a
major secular crisis on whether the market has matured. A mature digital ad market and a search
replacement? Partly explains why Google's stock is up 16% since ChatGPT's launch, while Meta's
stock is +98% over the same period, so let's just agree that ChatGPT plays a part here (Exhibit 1).
BERNSTEIN
270 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
EXHIBIT 1: Google's stock price has declined -10% since ChatGPT's launch, while Meta's has increased 14%
16%
98%
1%
8%
-40%
-20%
0%
20%
40%
60%
80%
100%
120%
Stock Price Change Indexed to November 30
GOOGL META S&P500 NASDAQ
Note: Updated on 5/12/23
Source: Bloomberg, Bernstein analysis
WHAT IS CHATGPT? THE
BRIEFEST OF PRIMERS
This isn't a primer note (yet), and I'm assuming if you're reading this you've probably heard of
ChatGPT even in passing. But for a bit more context and to help navigate the plethora of new-
to-me acronyms out there, GPT stands for generative pre-trained transformer, a type of LLM
developed by OpenAI that actually sits on Google's open-sourced Transformer architecture (NNI
for those interested in even more acronyms).
ChatGPT is trained on large datasets such as Common Crawl's repository of the open web and
then fine-tuned for tasks such as text generation, language translation, and question answering.
ChatGPT is a specific implementation of GPT-3 that is fine-tuned for conversational AI tasks.
DALL-E, for example, is also a GPT-3 implementation; however, it converts the text parameters
into images, allowing you to suddenly become an artist using only text prompts.
It's not often in tech you get that "wow" moment when experiencing something new, and I
certainly felt it when I tried out ChatGPT for the first time. And some of you will kill me for the
comparison I'm about to say next, but it's the same "wow" moment I felt when I put on Meta's
Quest 2 device for the first time. In both cases I was able to do something I've never been able to
do before, in one case feel like I am having a conversion with something nearing the Turing Test,
and in the other felt transported to a different world. And yet in both cases, I haven't felt the need
nor the urge to go back and try it again…yet.
We're working with limited data, but according to Similarweb, ChatGPT exited April with 30
million DAUs and 60 million daily visits (Exhibit 2). Of course, it's early, and OpenAI itself has been
throttling access, given the costs of running these query types. We'll probably see traffic tick up
once more use cases emerge, funding is secured, and maybe we even get an app.
BERNSTEIN
WILL AI KILL GOOGLE SEARCH? 271
EXHIBIT 2: ChatGPT exited April with 30 million DAUs and 60 million daily visits
0
10
20
30
40
50
60
70
Millions
ChatGPT Traffic (7 days trailing average)
Visits Unique Visitors
SimilarWeb and Bernstein analysis
MOST SEARCH IS SAFE…FOR
NOW
If you've read my work, or spoken with me, you'll know that my north star for disruption always
comes back to user behavior. The hardest way to disrupt something in tech is to try and create
new behaviors, though mountains of free money certainly helped the likes of Uber and TikTok
grease those wheels of behavioral change. As I write this, I'm seeing news break that Meta is
getting rid of the "Shop" tab from Instagram. It is one thing to scroll through your Feed and see
an ad that inspires you to click and purchase, but almost no one has ever thought to themselves
"I need to go buy some new shoes" and proceeds to open Instagram to go shopping. The "Shop"
tab was designed to create that exact behavior, and it never materialized.
Reports that Microsoft is planning on integrating ChatGPT into Bing suggests that the folks
in Redmond believe they've finally got a new mousetrap to take share from Google from their
current ~3% (7% in the US) share of search queries. But not all search queries are created equal.
Turns out I'm not quite the reformed consultant I thought, and offer the following framework to
think about where search is likely, and perhaps more importantly unlikely, to be disrupted.
BERNSTEIN
272 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
EXHIBIT 3: Unstructured informational queries are right in the crosshairs of ChatGPT
Source: Bernstein analysis
The first thing I'd note is that there's an ocean of difference between a general information search
query and a monetizable one. Back to my behavioral lens, in an effort to improve the search
experience, Google has slowly been training users to get answers from informational searches
without having to click on a link — a zero click search. Ask Google about the weather, a person,
or for a definition like "what is ChatGPT" — what I'm calling structured searches with a correct
answer — and you'll notice that Google already leverages AI by showing a snippet at the top of
your search results page. These snippets are usually pretty good, and if they're not you still get
the normal page listings, but in an ironic twist they've been slowly training users to trust a singular
answer.
There are a few critical issues with current generative AI responses:
1. The answer isn't always correct, and there are limited ways of tracking down the source; and
2. The answer removes all choice — if you don't like the answer, your best bet is to go back and
tweak your search query.
Simplistically, Google Search indexes websites and returns a list of websites based on a derived
quality score to the questions asked. But the result is still coming from a third-party website. One
can simply go check the source — say weather.com — to see if they really believe that it's getting
down to the 40s this weekend in Miami. Or head over to an OTA website and see hotel prices
based on the specific parameters entered.
When it comes to earning ad dollars, the current flow appears ideal if the user roughly knows
what they're looking for. We get high-quality results and, at the very least, the illusion of choice.
ChatGPT, on the other had, works a lot better with unstructured queries (Exhibit 3) — asking it to
BERNSTEIN
WILL AI KILL GOOGLE SEARCH? 273
write specific code to run a function offers a vastly improved result. One can easily see an entire
industry being born disrupting services across creative (create a video ad), financial (build me a
model), and especially homework.
But ChatGPT doesn't offer that validity, and Bob Brackett's recent note1 highlights what happens
when AI gets it wrong. It is not to say that it can't overcome this obstacle, as the model gets to
know who's to say that one day it can't plan the perfectly tailored trip along the Amalfi coast
based on your preferences inclusive of flight and preferred hotel choices. Maybe in this future
world airlines and hotels bid for the right to be included in the default answer. Maybe. That was
the promise of the smart speaker boom that fizzled out. It is hard to change consumer behavior,
especially if you start taking away their freedom to choose.
GOOGLE: THE AI COMPANY
AND THE INNOVATORS'
DILEMMA
To quote the quotable Todd Juenger's old media initiation, a funny thing happened on the way to
the graveyard. Google was born a disruptor in the land grab battle to map the Internet. And with
that DNA comes a constant fear that if they don't build the thing that disrupts Google, someone
else will. And so with blinders on they got to work, quietly acquiring DeepMind, hiring the brightest
minds in AI, and developing the foundation upon which the disruptive era of AI was to be built.
Imagine the discussion we'd be having if the current investor shift to cost-cutting took place
five years ago and Google had cut its AI investment because investors wanted another 25bps
in operating margins. Instead, Google has steadily been integrating their own LLM AI efforts
into Google Search for years. Around 20% of search queries in any given year have never been
searched for before, and so Google has leaned on internal LLM efforts such as BERT and now
MUM (more acronyms!) to try to better understand what it is you're searching for and improve the
quality of search results.
LaMDA is perhaps Google's most comparable offering to GPT-3, which is specifically trained to
understand dialog and carry out a natural and fluid conversation with humans. You may remember
not too long ago a certain Google engineer (later fired) claiming that the AI was so good it was
sentient. Google also has PaLM, which is a part of its larger Pathways AI architecture (introduced
way back in October 2021) that can quickly learn and excel at multiple tasks in contrast to all
these other systems trained to do just one thing. PaLM uses more than 540 billion parameters
in its training set, roughly 3x that of GPT-3. And lastly, Google Brain is out there with a trillion-
parameter AI model — 6X GPT-3.
I'm sure we've missed a bunch of things, but all this is to say that Google's been busy. It won't
get beat on technology, and theoretically it'll always have a data advantage (upon which these
models are trained), given its treasure trove of first-party data.
Google's long-standing issue is commercialization. Stadia, Google's cloud gaming service, was
by all accounts a technical marvel. It's since joined the ever-growing graveyard of failed Google
services. And that was trying to build a new business outside of core. Launching something that
could potentially disrupt core search would be the ultimate test of the innovator's dilemma.
1 https://www.bernsteinresearch.com/brweb/ViewResearchStreamer.aspx?cid=ejXD
%2faT8KVZNMtQX9QwTn8YrQRqfWe5B%2fA3YeRM2CA%2fbCqUnsRhyzwmbHrsIfZkt
BERNSTEIN
274 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
Let's take a hypothetical scenario where these LLMs are a perfect substitute for search. By
definition these models learn and improve from search volumes, which can only happen out in
the wild. In a closed environment, it is extremely difficult to train the model on all those search
queries that have never been searched before. Yet, today most Google searches are structured,
and results from ChatGPT, and I'd imagine from LaMDA, are inferior to the indexed results Google
currently delivers. They could be wrong, there's no choice, users need to get the search query
exactly right…in short, for most people it is probably a terrible experience versus the
standard search today, yet without those reps the model can't learn and leaves the door
open for a new entrant such as OpenAI that has no baggage and can afford those "funny"
wrong answers.
GROSS MARGIN RISK >
REVENUE RISK
At the risk of sounding Luddite, I stand by the view that technological disruption, and more
commonly channel disruption (e.g., Netflix > Blockbuster), is sometimes accompanied by some
shiny new tech. In this case, we see two margin-specific risks to Google: (1) the cost to own the
channel likely goes up, and (2) the cost to deliver said generative AI queries also goes up.
On 1, there has been a trickling of news flow/leaks tied to Microsoft aggressively looking to
win default search placement on Samsung, Mozilla, and I'd imagine Apple — where Google pays
out ~$20Bn for ~$60Bn in revenues — is the big prize. But status quo is incredibly hard to
break, especially with an inferior product that Bing has with informational search, that as far as
I'm concerned still makes up the lion's share of how people search. Why would a web browser
or device worsen the consumer experience to chase a buck? Occam's razor tells us that they
wouldn't look to switch out the search engine but instead simply go back to Google and ask for
a few more basis points of TAC (fees Google pays to be the default).
On 2, these AI models are expensive to run! It is no wonder that OpenAI is out there throttling
usage, launching paid tiers, and seeking a cash infusion. AI-based queries run on costly GPUs
instead of the more economical CPUs used to power Google's search indexing of the Internet. In
a response to a Twitter query by Elon Musk, Sam Altman replied that average cost to ChatGPT
per chat is "probably single-digit cents." Some estimates show that Google runs ~100k search
queries per second or ~3.1 trillion annually. Even assuming the low-end cost of just US$0.01,
if all the Google's search queries are migrated to an AI model overnight, the bill would run up
an extra ~US$31Bn per year and a cost of US$0.025 would double Google's entire Other Cost
of Revenue of ~US$76Bn. Suffice to say, running an AI-based engine to power its searches will
easily dig a large hole in Google's pockets.
But it is unclear just how much incremental costs Google would incur since a healthy portion of
these web indexed searches are already using AI, in this case Google's MUM, to deliver snippets,
auto-fill your query, and offer up additional questions they think you might be interested in.
Add in that running these queries should get cheaper over time, and perhaps the incremental
infrastructure costs aren't so great.
I'M MORE SCARED OF TIKTOK ChatGPT gets all the buzz, but it might not be the most imminent threat to Google's Search
dominance. I even caught myself searching for a restaurant review on TikTok when Yelp let
me down. Back to my behavior litmus test, searching on TikTok doesn't require a significant
behavioral shift — it simply changes where that user behavior takes place. YouTube already
trained us to search for videos, the same Maps trained us on how to search for locations.
BERNSTEIN
WILL AI KILL GOOGLE SEARCH? 275
The danger here is the same one facing Meta, in that younger Gen Z users don't have the baggage
and memory of opening the browser and searching on Google. They're born on TikTok and have
built their search behavior right in that app. In an offhanded comment at Fortune's Brainstorm
Tech conference last year, Google's SVP Knowledge & Information referenced Google's own
studies showing that ~40% of 18-24 year olds go to TikTok or Instagram instead of Google
Search (or Maps) when looking for a place for lunch. But here again we've seen Google fight back
from these disruptions when Amazon stole product search share. And on the TikTok front, most
Google searches now include more visuals with images and videos meshing with the old blue
links.
There's more to come, but for now let me conclude by saying that it is ultimately too early to
tell how ChatGPT, Microsoft, and conversational AI more broadly will impact the Search
landscape. Generative AI has a place in our Internet future, and I suspect we'll see the
technology underpin entirely new end markets that will ultimately be accretive to the
current ad-supported search universe. While some of us may see the "outline of Search's
peak," we'll get closer only to discover larger peaks in the background all ruled over by
the king of Search, Google.
VALUATION METHODOLOGY See Disclosure Appendix for valuation methodology.
EXHIBIT 4: Ratings and target prices
25-May-2023 Target
Ticker Rating Currency Closing Price Price
GOOGL O USD 123.48 125.00
SPX 4,151.28
Source: Bloomberg, Bernstein analysis and estimates
RISKS See Disclosure Appendix for risks.
Mark Shmulik mark.shmulik@bernstein.com +1 212 823 3237
BERNSTEIN
276 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
CHINA AI: GETTING SMARTER...SCORING
THE GENERATIVE AI CONTENDERS
HIGHLIGHTS •Gold rush! Since the ChatGPT launch in November 2022, generative AI tech has moved at
a dizzying pace. In China, Baidu, SenseTime (not covered), and Alibaba have now published
large language models (LLMs). The excitement among investors — and desire to find ways to
invest — has been palpable. In this chapter, we've outlined thoughts on the progress to date,
and some mental frameworks around how the benefits might accrue to LLM developers.
•Winning the long game. While the scrutiny over performance at launch is understandable,
we argue that speed of iteration and the surrounding ecosystems will be equally important. In
our view, consumer success will accrue to developers with the largest pools of high-quality,
in-depth user and content data, and whose app ecosystems support efficient monetization.
On the enterprise side, one outcome we're partial to is a market split between Alibaba, Baidu,
SenseTime…and perhaps Huawei (private). We suspect prevailing issues around enterprise
software adoption and willingness to pay will remain relevant.
•A five-point framework for assessing the competition. In this chapter, we've scored the
top LLM contenders in China along a framework focused on: (1) hardware layer scale; (2)
access to high-quality data and AI tech pedigree — together influencing a developer's ability
to develop a model; (3) strength of consumer lock-in and existence of organic monetization
pathways; (4) industry structure — which we expect will inform the distribution of gains from
generative AI; and (5) agility of management execution.
INVESTMENT IMPLICATIONS In recent months, excitement has exploded globally around generative AI, and the potential for
LLMs to impact the tech sector. The desire among investors to discuss the technology and ways
to invest has been enormous. In China, a gold rush of sorts has ensued, with Baidu, SenseTime,
and Alibaba publishing LLMs to date — arguably despite underinvesting in reinforcement
learning time. In general, our instinct is to remind investors that these models — and broader
generative AI — remain in the very earliest stages of their development. Further, we argue
that user reach and lock-in, monetization pathways, and industry structures will be at least as
important for commercial AI success as the underlying technology.
Within our China Internet coverage, we consider Tencent (Outperform, target price HK$455)
and Alibaba (Outperform, target price US$130/HK$128) the most likely to emerge as long-term
winners from generative AI development. Within our China Software coverage, we rate Baidu
(OutPerform, target price US$160/HK$157) as we expect the competition in the enterprise
space to be more intensive than expected.
BERNSTEIN
CHINA AI: GETTING SMARTER...SCORING THE GENERATIVE AI CONTENDERS 277
OUR THOUGHTS ON
GENERATIVE AI IN CHINA
In recent months, OpenAI's ChatGPT and GPT4 launches have taken the world by storm. Initially
what struck us above all else was the child-like wonder and excitement the technology invoked
in both hardened tech veterans and users. Product development has evolved at an astonishing
pace since. In China, a flurry of activity followed the ChatGPT debut. Baidu's Ernie took a month
to go from China's answer to OpenAI to facing generally mixed reviews from users, to now being
considered merely "one of many." Sensetime and Alibaba have since launched their own LLM
products. Both Tencent and Bytedance have announced that they are developing "ChatGPT-like"
services, though neither have made more specific comments on timing.
In our recent discussions with investors, generative AI has quickly become a favorite "side topic."
As enthusiasts on the underlying technology, the temptation to explore all the various technical
rabbit holes has been high. For investors looking to generate a financial return from the evolution
of this technology, we think commercial success will require contenders to possess a variety of
capabilities in addition to owning stacks of semiconductors and training large models. In this
chapter, we've outlined a five-point framework, which we think will be useful when assessing
generative AI contenders in China. For brevity, we've focused our attention on five companies —
Tencent, Alibaba, Bytedance (private), Baidu, and SenseTime.
A five-point framework for assessing generative AI competition in China
In the spirit of putting the conclusion first, in our framework we've scored the main LLM
developers in China along the axes of: (1) hardware layer strength — essentially cloud hosting
capabilities and compute capacity; (2) platform layer capability — e.g., access to high-quality,
in-depth user datasets, and visible progress to date in fields such as NLP, machine vision, and
LLMs; (3) the strength of their commercial ecosystems — including strength of user lock-in
and user willingness to pay; (4) the intensity and nature of competition developers face, which
determines whether gains related to generative AI can be retained or will be competed away; and
(5) management track record on complex new projects and agility of execution in a broad sense.
Exhibit 1 shows our assessment of the top LLM developers in China, scoring each of the
companies along five axes. To state the obvious, these scores incorporate subjectivity on our part,
and leaves out nuances in our views by company. But we're hopeful that this at least offers a
starting point for discussions with investors.
EXHIBIT 1: We've scored the top five LLM developers in China on five metrics we believe will prove important;
Tencent, Bytedance, and Alibaba seem most likely to win
Hardware Platform Application Industry structure Execution
Tencent ★★★★ ☆★★★★ ★★★★★ ★★★★ ☆★★★★
Alibaba ☆★★★★ ★★★★ ★★★★ ★★★ ☆★★★
Bytedance ★★★★ ★★★★ ★★★★ ★★★★ ☆★★★★
Baidu ★★★★ ★★★★ ★★★ ★★★★ ★★★
SenseTime ★★★★ ☆★★★ ★★ ☆★★★ ★★★
Source: Bernstein analysis and estimates
BERNSTEIN
278 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
THE WORLD IS SPINNING
TOO FAST…SOME LESSONS
FROM PAST CYCLES OF AI
EXCITEMENT
The development speed of the underlying LLM technology and new APIs both has been dizzying.
Trying to keep up with the flurry of media announcements, Arxiv papers, and new developments
(e.g., AI agents) has been an all-hands-on-deck exercise. The speed of news flow creates the
obvious risk that any technical observations made in this chapter could soon become obsolete.
But we think some general principles we've applied to assess the leading developers will have a
longer shelf life. Helpfully, unlike past episodes of AI excitement (the autonomous driving boom
and bust, and Metaverse hype come to mind), generative AI has arguably already found ways to be
commercially valuable. Compared with autonomous driving, generative AI runs less risk of getting
bogged down by an endless stream of new edge cases. The cost of failure (incorrect responses,
strangely conceived pictures, etc.) also feels much lower than in the autonomous driving example
(grandma gets hit by a Chevy Bolt).
Does having a head start matter? In our view, not in the Chinese case to date. A wide range of
China Internet and software companies have announced plans to develop LLMs and generative
AI functionality — mostly for fear of being left out of the excitement, or being considered less
advanced than peers. So far, we would describe all the launches to date as sub-GPT-3 levels of
performance, and suffering from insufficient fine-tuning, reinforcement learning, and alignment
training — making them limited facsimiles of what they might eventually become. ChatGPT
ostensibly required six months of such training before seeing the light of day. Our inclination is
that a year from now the field of competition will have narrowed, leaving a smaller cohort of better,
more refined models.
Let's get this out of the way…the semiconductor challenge is real
The US government's strategic blockade on China in leading-edge semiconductors creates real
problems for China's efforts to cultivate an AI champion. Industry contacts we've spoken with
have mostly preferred to leave this topic unaddressed, but it is clear that both the pre-training
of LLMs and subsequent refinement of LLMs and generative AI products require vast amounts
of GPU chips such as Nvidia's A100 and H100. The introduction of Nvidia's A800 and H800
offered Chinese buyers a diminished, but still credible alternative. Tencent, for example, recently
announced a hardware stack designed to facilitate LLM trading, featuring arrays of H800 chips.
Baidu has claimed a larger stockpile of GPUs than peers. Regardless, industry contacts we spoke
to expect the limited supply of GPUs to represent a bottleneck for how quickly generative AI
development can progress both in China and possibly elsewhere.
Commercially, it feels logical that Nvidia will want to provide Chinese buyers with future
generations of leading-edge semiconductors, even if they will need to be constrained in some
way like the A800/H800. In this age of confrontational geopolitics, there's no guarantee that
the US government will continue to allow such workarounds forever. In the event it does not, we
would expect AI development in China to be negatively impacted — possibly in a very severe way.
The Chinese government's efforts to brute force an indigenous leading-edge semiconductor
industry have led to mix results to date.
On a global level, the above also raises questions about whether a "fast follower" strategy among
Chinese players can be viable. OpenAI founder Sam Altman notably argued that the future
of AI development belonged not in ever larger models, but more reinforcement learning and
alignment, and novel approaches to the AI problem. On a relative basis in a China-only context,
the main contenders' access to distributed GPU computing facilities becomes an important
differentiator. Said differently, large cloud players that have stockpiled advanced Nvidia chips will
have an advantage over newcomers that have not. The former likely also have more experience
BERNSTEIN
CHINA AI: GETTING SMARTER...SCORING THE GENERATIVE AI CONTENDERS 279
in linking GPUs for high-end compute.
LLMs have gotten exponentially more complex, but training dataset scale is as
important, if not more so, as parameter scale
Over time, LLMs have become bigger and more complex, requiring a growing amount of
compute to train (see Exhibit 2 to Exhibit 4). One common theme across the recent LLM launch
presentations in China has been the celebration of model parameter scale. Alibaba's Tongyi
Qianwen claimed 10 trillion parameters, while Sensetime's SenseChat product is said to have
been based on "hundreds of billions" of parameters. Baidu's Ernie 3.0 Titan model (a 2022
ancestor to the latest Ernie) had 100 billion parameters.
While the promotion of parameter scale in the media is understandable, DeepMind's Chinchilla
model serves as a reminder that training dataset scale can be of greater importance for model
performance. DeepMind's paper argued that even models such as GPT-3 were essentially under-
trained relative to their parameter scale. Our discussions with Chinese AI contacts make us
think a similar issue likely afflicts the latest Chinese LLMs. The extent to which training datasets
could be fully utilized during training — and the extent to which these developers spend time
on reinforcement learning from human feedback (RLHF) — were also cited as key performance
drivers to watch as Chinese developers seek to close the gap with OpenAI.
Limitations to LLM development in China: the CAC might be a bigger challenge
than Chinese language datasets
We heard interesting push back on the notion that the Chinese language imposed limitations on
training dataset scale, given there is far more English language content on the Internet than in
Chinese. Some Chinese AI experts we know have argued that the total size of the latter is only
a single-digit percentage of the former, which creates issues for LLM development for Chinese
language applications. Baidu's Ernie notably seems to generate images by first translating the
prompt into English. That said, it is notable that GPT-4 sees comparable performance in a variety
of languages (e.g., Afrikaans) that presumably have smaller natural datasets than Chinese.
We think a practical issue for Chinese developers relates to the CAC's requirement that training
datasets contain information that's "truthful, accurate, and objective." Developers we met
expected the industry push back against this requirement. As it stands, the rule was expected to
raise the technical bar for LLM development in China, meaning a much smaller field of players
will be able to compete, but also slow down development even across the top Chinese players.
BERNSTEIN
280 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
EXHIBIT 2: Number of parameters of large language and multimodal models has increased significantly in recent
years
GPT-2 Grover-Mega
Megatron-LM
(Original, 8.3B)
T5-3B
T5-11B Turing NLG
Meena
GPT-3 175B (davinci)
ERNIE-GEN (large)
DALL-E
Wu Dao - Wen Yuan GPT-Neo
GPT-J-6B
PanGu-α
CogView
HyperClova
ERNIE 3.0
Jurassic-1-Jumbo
Megatron-Turing
NLG 530B
Gopher
GPT-NeoX-20B
Stable Diffusion (LDM-KL-8-G)
Chinchilla
OPT-175B
PaLM (540B)
Minerva (540B)
GLM-130B
BLOOM
DALL·E 2
Jurassic-X
Wu Dao 2.0
Codex
GPT-4
Yuan 1.0
M6-10T Tongyi Qianwen
SenseNova
1.0E+08
1.0E+09
1.0E+10
1.0E+11
1.0E+12
1.0E+13
1.0E+14
Dec-18 Jul-19 Jan-20 Aug-20 Feb-21 Sep-21 Mar-22 Oct-22 May-23
Number of parameters (Log scale)
Date published
2019-2023: Large Language and Multimodal Models - number of parameters
Source: AI Index Report 2023, Stanford University, Epoch, Bernstein analysis
EXHIBIT 3: Training datasets have grown alongside parameter scale, though recent studies (e.g., DeepMind's
Chinchilla exercise) have argued that the former could be more important for model perfomance (1/2)
GPT-2 Megatron-LM (Original,
8.3B)
T5-3B T5-11B
Turing NLG
Meena
GPT-3 175B (davinci)
DALL-E
GPT-Neo
GPT-J-6B
PanGu-α
CogView
HyperClova
ERNIE 3.0
Jurassic-1-Jumbo
Megatron-Turing NLG 530B
Gopher
GPT-NeoX-20B
Stable Diffusion (LDM-KL-8-
G)
Chinchilla
OPT-175B
PaLM (540B)
Minerva (540B)
DALL·E 2
Codex
Yuan 1.0
M6-10T
1E+7
1E+8
1E+9
1E+10
1E+11
1E+12
1E+13
1E+9 1E+10 1E+11 1E+12 1E+13
Training dataset size (datapoints)
Number of parameters
2019-2023: Large Language and Multimodal Models - Training dataset size (datapoints) vs.
number of parameters
Source: AI Index Report 2023, Stanford University, Epoch, Bernstein analysis
BERNSTEIN
CHINA AI: GETTING SMARTER...SCORING THE GENERATIVE AI CONTENDERS 281
EXHIBIT 4: Training datasets have grown alongside parameter scale, though recent studies (e.g., DeepMind's
Chinchilla exercise) have argued that the former could be more important for model perfomance (2/2)
Model Developers Date published Parameters FLOPS
Tongyi Qianwen Alibaba Apr-23 Est. 10tn n.a.
SenseNova SenseTime Apr-23 100bn 5E+18
ERNIE (Wen Xin Yi Yan) Baidu Mar-23 Est. 260bn n.a.
GPT-4 OpenAI Mar-23 Est. 1tn 2.2E+25
BLOOM Hugging Face, BigScience Nov-22 176bn 1.8E+23
GLM-130B Tsinghua KEG Aug-22 130bn 4.6E+22
Minerva (540B) Google Jun-22 540bn 2.7E+24
Jurassic-X AI21labs May-22 7bn n.a.
OPT-175B Meta AI May-22 175bn 7.6E+23
Hun Yuan Tencent Apr-22 Est. 1tn n.a.
Stable Diffusion (LDM-KL-8-G) Stability AI, Runway Apr-22 1.5bn 5.0E+22
DALL·E 2 OpenAI Apr-22 3.5bn n.a.
PaLM (540B) Google Apr-22 540bn 2.5E+24
Chinchilla DeepMind Mar-22 70bn 5.8E+23
GPT-NeoX-20B EleutherAI Feb-22 20bn 9.3E+22
AlphaCode DeepMind Feb-22 n.a. 4.1E+22
Gopher DeepMind Dec-21 280bn 6.3E+23
Yuan 1.0 Inspur Oct-21 245bn 4.1E+23
Megatron-Turing NLG 530B Microsoft, NVIDIA Oct-21 530bn 1.4E+24
Tong Yi - M6-10T Alibaba Oct-21 10tn 5.5E+21
Jurassic-1-Jumbo AI21 Labs Aug-21 178bn 3.7E+23
Codex Open AI Jul-21 12bn n.a.
ERNIE 3.0 Baidu Jul-21 10bn 2.4E+18
Wu Dao 2.0 BAAI Jun-21 1.75tn n.a.
CogView Tsinghua, Alibaba May-21 4bn 2.7E+22
HyperClova Naver May-21 204bn 6.3E+22
GPT-J-6B EleutherAI May-21 6.1bn 1.5E+22
PanGu-αPanGu-α team Apr-21 207bn 5.8E+22
GPT-Neo EleutherAI Mar-21 2.7bn 7.9E+21
Wu Dao - Wen Yuan BAAI Mar-21 2.6bn 6.5E+20
DALL-E OpenAI Jan-21 12bn 4.7E+22
ERNIE-GEN (large) Baidu Aug-20 340mn n.a.
GPT-3 175B (davinci) OpenAI May-20 175bn 3.1E+23
Turing NLG Microsoft Feb-20 17bn 1.6E+22
Meena Google Jan-20 2.6bn 1.1E+23
T5-3B Google Oct-19 3bn 1.0E+22
T5-11B Google Oct-19 11bn 4.1E+22
Megatron-LM (Original, 8.3B) NVIDIA Sep-19 8.3bn 9.1E+21
Grover-Mega University of Washington May-19 1.5bn n.a.
GPT-2 OpenAI Feb-19 1.5bn 1.5E+21
Note: Models published by Chinese developers highlighted in blue.
Source: AI Index Report 2023, Stanford University, Epoch, public reports, Bernstein analysis
What it takes to win: LLM tech, but also content generation, app ecosystems,
user lock-in, willingness to pay…
We think it is important to consider that commercial AI success will require more than simply
developing and refining a model. However, we also argue that a steady source of high-quality user
data — and pathways to organic monetization — will be at least as important. The earliest paying
customers for generative AI services such as Midjourney have included prospective developers
and or tech enthusiasts like ourselves. In the fullness of time though real commercial success will
require such tech to be made available to much broader audiences.
On the consumer side, we think it is likely that over time generative AI will become part of
how online platforms and media outlets generate engagement. Alibaba's LLM launch featured
repeated mentions of how Tongyi Qianwen will be integrated with its ecosystem of apps.
BERNSTEIN
282 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
Microsoft has spoken of introducing ads as part of its Bing Chat service. In our view, gaming
represents another obvious arena where generative AI can be used to improve user experience.
NetEase, for example, recently spoke of incorporating immersive NPC chatbots into its upcoming
Justice Mobile game. On the enterprise side, the most attainable monetization opportunities
will likely come from the licensing of productivity tools and time savers, and developer tools for
downstream commercial use. The obvious examples here include office assistants and customer
service chatbots, but could potentially evolve to include AI-generated imagery, video, or beyond.
We expect the prevalence of walled gardens in China to have important implications for
data collection as well as AI commercialization. For players without access to an established
application layer, we would expect the path to building a new one to prove difficult. Most top
apps in China — think WeChat, Douyin, Taobao, Meituan — have built user bases that have shown
considerable stickiness. The history of China Internet is littered with examples of companies
offering lavish promotions or new products and services — only to be confronted with the issue
that users locked in to larger competing apps simply remained oblivious to them.
Industry structure and competition will drive value capture (i.e., profits) as much
as technology leadership
Longer-term followers of AI will be familiar with the printing press analogy. Movable type printing
was either invented by the Song dynasty of China (c. 1040 CE) or shortly before 1450 by
Johannes Gutenberg, depending on the language in which one reads Wikipedia. Compared with
the state-of-the-art at the time, the new technology catalyzed a revolution in the generation and
distribution of content — similar to our expectations for generative AI. As far as we can tell though,
the revolution did not bring wild riches to the inventors of the press (in either version of the
inventory story). Gutenberg notably was bankrupted by a co-financier of his press.
The point of bringing this up is that efficiency gains and consumer surplus are generally
necessary for widespread adoption of a new technology. We doubt many investors will question
the direction of generative AI adoption from here. For all the excitement, we still expect industry
structure and competitive intensity to determine how the value associated with the new tech is
distributed. In the printing press example, when new iterations of printing presses were brought
to market that lowered production costs, printing houses invariably bought the newest version
so they could print at lower cost. But in short order, every other printing house would also buy
the new press, resulting in lower prices for and greater distribution of printed materials. But print
house margins quickly reset to an equilibrium not dissimilar from prior levels.
In the Internet sector, the nature of competition means we expect verticals such as advertising
and gaming to retain the gains from generative AI to a greater extent than say e-commerce, where
periods of price competition are routine.
ChatGPT can help Buzzfeed generate content, but then why do we still need
Buzzfeed?
One of the more memorable early examples of AI excitement we can remember was Buzzfeed
announcing it would replace human-generated content with AIGC, causing its shares to
essentially triple over two trading days. There are several ways one might choose to explain the
excitement (e.g., shiny new toy effect, the meme stock phenomenon, etc.), but presumably the
assumption was that: (1) the novelty of AIGC might drive increased traffic to Buzzfeed; or (2)
Buzzfeed might benefit from a reduction in its human content generation cost. Since then, much
BERNSTEIN
CHINA AI: GETTING SMARTER...SCORING THE GENERATIVE AI CONTENDERS 283
more reasonable commentators have claimed that the incremental cost of content will trend
toward zero over time. Cheap energy + AIGC = free digital crack for all!
In general, we consider investing in the generative AI application layer a hazardous endeavor.
For a start, the underlying tech remains in considerable flux. Another problem we have with
calling users of generative AI tech "AI winners" is: If ChatGPT could produce articles of a standard
comparable with Buzzfeed's existing team of human content generators, why would users not
simply use ChatGPT for such needs going forward? Why, in this instance, would Buzzfeed still
have a right to exist? User habits can be a valid defense over short-to-medium-term horizons.
Over long-term horizons, we would expect the latter's terminal value to be called into question.
In the online platform and content generation domains, outside a limited number of apps that
enjoy absolute ubiquity (e.g., WeChat, Douyin, etc.), we expect the debate of whether generative
AI is friend or foe to be a long one. Video games strike us as an area where generative AI can
create value without introducing meaningful "disruption risk," given the importance of storytelling
and other features of games that go beyond the generation of visual or audio content. The full
range of market cap implications from this kind of disruption is beyond the scope of this chapter
— but is clearly vast.
Feedback from the downstream software industry in China
Most downstream companies we know acknowledge that generative AI adoption will not be
an overnight phenomenon, but one that will require investment of considerable resources and
expertise in the coming years. The emergence of ChatGPT took our downstream companies by
surprise, which are now pursuing the challenging task of defining their strategy, designing their
app requirements, developing initial pilot products, and testing in phases in response. We expect
this process will take a good chunk of 2023. On potential partnerships with LLM developers,
most downstream companies have retained a "wait and see" approach — before it becomes
clear which upstream companies can bring the most compelling product later in 2023. From a
commercial tie up perspective, many have cited the importance of user-friendly SDKs and PaaS
connectivity layers as part of their selection criteria for potential LLM partners.
Several software companies we spoke with argued that the Chinese enterprise LLM market
will become a multi-model market, where applications would route queries to different models
based on functionality and costs. While we might be far away from that type of reality, suffice
to say nobody upstream today has become sufficiently differentiated from the perspective
of downstream software developers. Over time, we think upstream developers with the best
enterprise support, routing tools, and optimization software will win most traction on the
enterprise front — not too dissimilar to competition across PaaS platforms.
A FIVE-POINT ANALYTICAL
FRAMEWORK FOR
ASSESSING AI IN CHINA
In our view, generative AI tech stack follows a similar hardware-platform-application
configuration as do many popular tech stacks common today (Exhibit 5). The hardware layer
essentially consists of servers and semiconductors, cloud hosting, data warehousing, and
compute capacity. We would like to think of the LLM itself as part of the platform layer. Also, part
of the platform are APIs that allow data owners to connect proprietary datasets to the inference
capabilities of LLMs. End applications that interface with human users (e.g., ChatGPT, Midjourney,
Stable Diffusion, etc.) then represent the application layer.
BERNSTEIN
284 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
In the following sections, we've scored the main LLM developers in China along the axes of:
(1) hardware layer strength — essentially cloud hosting capabilities and compute capacity;
(2) platform layer capability — e.g., access to high-quality, in-depth user datasets, and visible
progress to date in fields such as NLP, machine vision, and LLMs; (3) the strength of their
commercial ecosystems — including strength of user lock-in and user willingness to pay; (4) the
intensity and nature of competition the developers face, which determines whether gains related
to generative AI can be retained or will be competed away; and (5) management track record on
complex new projects and agility of execution in a broad sense. We think this represents a set of
necessary rather than sufficient conditions for commercial success in AI.
Hardware layer: generative AI is an expensive hobby
The hardware layer is the most straightforward part of our analytical framework. The large cost of
training LLMs — coupled with the fact successive generations of LLMs have grown exponentially
in complexity — means it is an exercise feasible only for the largest tech companies. By most
third-party estimates, Alibaba and Tencent rank first and either second or third in terms of IaaS/
PaaS market share in China, while Baidu's AI Cloud business is smaller. Bytedance has begun to
build out a cloud vendor business in the last 18 months, but our understanding is this remains
small. Our channel checks suggest that each of these companies has stockpiled GPUs in recent
quarters. We expect the limited availability of GPUs to represent a limiting factor for new entrants
hoping to participate in the industry — notably VC funded start-ups.
Platform layer: datasets and model design capabilities
We define the platform layer in the generative AI tech stack as the LLM itself, plus APIs used
to facilitate communication with the application layer and end users. In our view, the nature
of generative AI as a creature dependent on the ingestion of new data (both to extend the
knowledge cut-off and to refine responses through techniques such as RLHF) means the access
to high-quality, in-depth data represents an integral component of the generative AI platform
layer. We think this applies to LLM developers globally. But the fact that China's consumer
Internet is organized around walled gardens means that data is even more siloed among different
players and more inaccessible if one does not have access to certain groups of customers. As
such, there is an added premium on the ability to capture data across user life cycles and usage
scenarios. Tencent and Bytedance both have access to almost everybody in China through
WeChat and Douyin. In contrast, SenseTime might have a solid database of image data from its
computer vision exploits, but has much lesser user-specific data.
Application layer: properties and user lock-in (search, productivity, content
generation, gaming)
For individual developers of LLMs, we expect their respective endeavors to be heavily influenced
by existing arenas where these companies already have domain expertise, and where the
strength of consumer lock-in will enable the developer to benefit from incremental monetization
without having to retrain user behavior. Alibaba's new LLM, for example, will be foremost used
to recommend shopping ideas on Taobao and Tmall, and as a part of productivity suites for
merchants, as well as for Alicloud and DingTalk. Tencent's generative AI efforts will presumably
be focused on accelerating game development, capturing eyeballs on Tencent's various media
properties, or as chatbot assistants within the WeCom/WeChat Mini Program/Video Accounts
ecosystems. This should be obvious, but the flip side is also true in that lack of domain expertise
in the application layer will put fences around what individual developers will struggle to do.
BERNSTEIN
CHINA AI: GETTING SMARTER...SCORING THE GENERATIVE AI CONTENDERS 285
A critical view of our emphasis on existing ecosystems might argue that we're underestimating
the forces of disruption. Over very long-time horizons, this is probably fair. But we note that
Microsoft's Bing search engine had a 6.6% share of the US search market in March 2023, flat
versus a year ago, and lower than November 2022 when ChatGPT excitement drove Bing's share
to a high of 7.8%.
Industry structure: where will the value accrue?
This axis of our framework borrows from the printing press example earlier. In competitive end
markets with many players, we expect LLM developers to find it more difficult to retain the
value that accrues from generative AI development. We also differentiate between markets
where competition is prone to happening on the basis of user subsidies and counter-subsidies
(e.g., e-commerce), compared with markets where competition takes place in a more oblique
fashion (e.g., gaming). On the enterprise side, competition happens more on the basis of building
relationships with large customers. We see from other software categories that once formed,
relationships are generally sticky. Companies need to push the boundaries of technology and
refrain from the temptation of providing endless service commitments to drive pricing power and
product margin growth over time. A subscription model here is the holy grail every enterprise
tech provider aims for, but few in China have achieved to date.
Management execution: agility of execution and track record on
commercializing major new projects
Given and the speed of technological evolution and how early we are in the adoption cycle
for generative AI, we think that investments in the space implicitly place a heavy premium on
management teams' ability to "figure things out" over time. It is clear that OpenAI has opened a
large early lead on the competition globally in GPT model technology. But our discussions with
AI experts have the raised the possibility of future "forks" that can shake up industry leadership.
In China's context, we think Tencent and Bytedance have demonstrated the highest levels of
"agility of execution" in recent years, along with PDD, which to date has yet to enter the generative
AI conversation. Alibaba's prowess in cloud is clear, but its execution in core e-commerce in
recent years (and broader sense of organizational inertia) has left more to be desired. Baidu's
past history with moonshots has been a factor that has colored our perception of its odds of
generative success.
BERNSTEIN
286 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
EXHIBIT 5: We expect the generative AI tech stack to follow a typical hardware-platform-application layer
schematic, surrounded by user and content ecosystems, which will be equally important for commercial success
Indicative generative AI tech stack
Cloud hosting services, database, etc.
Large language model
API
API
API
API
API
API
API
API
API
API
User applications
Platform layer
Hardware layer
Application layer
Source: Bernstein analysis
FURTHER THOUGHTS ON THE
MAIN CONTENDERS
In our view, it is hard to look past Tencent and Bytedance as eventual AI winners in China. The
slight problem with this conclusion is that neither has published an LLM product for public
consumption yet, so a leap of faith is required that recent promises of ongoing development
bear fruit. But the combination of considerable compute capacity, massive content ecosystems
— which can be used both as a source of data and as a venue to put AIGC on display — represent
important structural advantages, in our view. Both occupy considerable user time spent in China,
and can monetize users in a variety of ways — including digital advertising, but also beyond. Baidu
has spoken about generative AI in the context of bringing the company closer to the bottom of
the ads funnel, and in the context of enterprise productivity tool kits.
On the enterprise side, one of the eventual outcomes we're partial to is a divided market split
between Alibaba, Baidu, Sensetime and, perhaps, Tencent. Huawei is also a player to remember
— it has stated plans to lead in AI, and its vision to have a full tech stack with local chips means
that LLM is an area it would need to participate in somehow. While Alibaba leads in scale and
resources, Baidu and SenseTime are essentially staking their future on making enterprise work.
Tencent's hold on merchants keen to tap into its WeChat ecosystem means it should also have a
part to play on the enterprise front — in a multi-model, multi-cloud future. Huawei, with its deep
pockets and full tech capabilities, could also emerge as a solution tailored for complex SOEs.
Unlike monetization on the user side though, we suspect prevailing issues related to enterprise
software adoption, pricing power and willingness to pay (especially a subscription model) will
remain relevant.
Tencent
The reach of Tencent's WeChat ecosystem is unparalleled, and acts as both an environment
which can contribute to the training of Tencent's AI model(s), and a venue where AIGC can be
put on display. AI chatbots are an obvious application of generative AI tech which could be
introduced within WeChat and QQ, and through WeCom and Qidian (CRM tool) on the enterprise
side. Generative AI tech could also contribute to Tencent's gaming business, and the Video
Accounts ecosystem. On the former, tools such as Midjourney and Stable Diffusion have already
BERNSTEIN
CHINA AI: GETTING SMARTER...SCORING THE GENERATIVE AI CONTENDERS 287
started making waves in the graphics design space — it would surprise us if AI-generated 3D
models don't follow in short order. The applicability of AIGC within the WeChat Video Accounts
ecosystem also feels obvious (at present curiosity about AIGC alone seems sufficient to attract
eyeballs — over time we expect quality to replace novelty as a driver of engagement). The rabbit
hole on AI-powered v-tubers runs deep too. Did you know for example that Tencent's Honor of
Kings has an associated virtual boy band?
On the enterprise side, WeCom and Qidian could potentially benefit from AI-enabled chatbots
as part of their service. Tencent's Voov-Docs-WeCom suite of software meanwhile is probably
the closest thing in China that has the potential to follow a sort of Office-Teams-Copilot strategy,
potentially sold to both individual users and enterprises. To the extent enhanced functionality
drives incremental demand for cloud hosting services, we would expect Tencent's broader cloud
business to benefit too. Culturally, we think these types of growth areas align well with Tencent's
historic preference for being the toolmaker.
The obvious knock against Tencent is lack of product. In contrast with Baidu and Alibaba, Tencent
has yet to formally publish the latest iteration of its Hunyuan model. The company has also
historically been less visible than, say, Baidu or Alibaba within the Chinese AI scene. On the other
hand, it is known that Tencent owns the largest number of AI-related patents out of the field of
competition (all manner of nuances likely exist, meaning a literal interpretation of the headline
number creates issues, but even so). Its Hunyuan model achieved a record score on the CLUE
test for Chinese language proficiency. Tencent management recently spoke of advancements in
AI tech having a positive impact on digital ads growth. Given the vast array of multi-modal content
data Tencent has access to, we think it is reasonable to expect Tencent to bring a competitive
product to the table.
Alibaba
We attended Alibaba's LLM launch in person — the event added to our view that Alibaba has
the tools necessary to be an AI winner in China, especially on the enterprise side. There was
no live demo of the new Tongyi Qianwen product on stage, and the day’s presentations were
light on technical detail on the model. But management presentation indicated the possible
product map — connecting user prompts with shopping recommendations, travel planning,
event planning ideas, and more. On the enterprise side, the company touted joint development of
propriety enterprise models based on Alibaba's foundational model. DingTalk and smart speaker
integration were said to be "coming soon."
Media testing of Tongyi Qianwen has generally looked solid, and in our view painted it in
a reasonably positive light compared with, say, Baidu's Ernie. Alibaba spoke candidly about
comparisons with OpenAI's ChatGPT and GPT-4. On some level though we think the real proof
in the pudding will be in Alibaba's speed to market with LLM integration in its various apps,
rate of enterprise adoption, and rate of performance improvement between now and, say, in
several months as the company has spent more time on fine-tuning (e.g., with RLHF). On the
consumer side, we're generally open-minded about the possibility that the novelty of Tongyi
Qianwen could drive user engagement — neither JD nor PDD have made much visible progress
toward a competing product.
Our main concerns with Alibaba's ability to monetize LLM development relate to willingness to
pay — both among consumers and enterprise users. What happens, for example, when Tongyi
Qianwen recommends a product, but PDD sells a similar SKU for cheaper? On the enterprise
BERNSTEIN
288 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
side, questions over the extent of enterprise digitization and willingness to pay for software are
not new, but will remain relevant too.
Bytedance
Our belief that Bytedance will rank among the eventual AI winners in China is rooted in: (1) the
size of the Douyin-Toutiao content ecosystem — which features both news and other text-based
information and China's largest short video UGC ecosystem; (2) Bytedance's history as an early
adopter of AI-based recommendation algorithms and considerable investments in AI over the
years; and (3) the company's track record as one of the sector's most formidable executors.
Douyin (and Toutiao to a lesser extent) offers a natural outlet for short videos based on AIGC.
Many of the arguments we've made above for Tencent (e.g., AI-powered livestreamers) ought to
apply to Bytedance too. Bytedance recently hired Yang Hongxia, development lead of Alibaba's
M6 LLM, to join its effort to develop an LLM.
Bytedance's capabilities on the enterprise side, on the other hand, are more limited. Bytedance's
attempts to enter the gaming market have also been met with limited success to date. Pico, its
virtual reality unit, remains very much a work in progress. Bytedance has gradually internalized
its cloud workloads in recent years, and has harbored ambitions to build its own cloud vendor
business in the not too distant past. But, we understand the business has remained small. It is
also notable that Bytedance's enterprise productivity suite Lark — despite being well-regarded
— has only attracted a fraction of the users of WeChat's WeCom or Alibaba's DingTalk.
Baidu
Baidu has continuously invested in LLM, upgrading ERNIE from 1.0 to 3.0, making it one of
the few truly multi-modal LLMs in China that offers full-suite NLP, CV, Cross-Modal, and Life
Science models. Baidu launched ERNIE to external partners in mid-March 2023 for initial trial,
while all other competitors are at best in the internal testing phase. On paper, it seems Baidu is
taking all the right steps to be a leading contender and capitalize on any head start it might have.
However, from the reviews and testing, it seems Baidu still is quite far away from an eventual
model — recent news of other players' models reveal that the stage of progress/quality is not that
distinguishable. While one could argue around parameters and metrics, from a user experience
perspective (and if the Turing Test matters), it is arguable whether Baidu has a lead at this point.
Still, from the investment it has made, the vast amounts of curated data, and indexed web search
data, we believe its model will be competitive with the rest, and could emerge as one of the
earliest players. It might not be the only leader, but there could be a handful of players, e.g.,
Alibaba and SenseTime, with a GPT level 4 version within a year's time.
The question with Baidu is around how would it commercialize its technology. In terms of Search,
it is the market incumbent, and Search is inherently fragmented in China. Would GPT enable
Baidu to gain share in Search that heretofore has been taken by Alibaba (ecommerce), Meituan
(food delivery), etc.? It feels to us that Search is based on user journeys. So, instead of aiming
to take share from others, we expect Search volumes could see an increase if Baidu can open
up Search queries that can't be done to date with indexed search, i.e., more complicated multi-
query searches. Incorporating its productivity assistant into its ecosystem of tools could also help
incrementally gain some traffic to its Search — though the productivity area would be competitive
as we question what edge Baidu would have here compared with, say, Tencent (offered on
WeCom) or Alibaba (offered in Dingtalk ecosystem). Our base case is: GPT helps improve usage
of its Search, though it might not necessarily gain share from others — incrementally, it still is a
BERNSTEIN
CHINA AI: GETTING SMARTER...SCORING THE GENERATIVE AI CONTENDERS 289
net positive for Search, which has been priced with a 3-5% ongoing growth rate.
Regarding its Enterprise endeavors, as we argued above, we envision the eventual market
unlikely to display winner-takes-all dynamics and will probably evolve in a more fragmented
fashion. While Baidu has talked about having more than 650 ecosystem partners, from our
channel checks it looks like everyone has connected with an API to test rather than having chosen
Baidu's product at this time. Enterprise customers are more likely to try different AIGC products,
thereby, one might argue the first mover advantage here will be limited. However, we believe that
Baidu will still have an edge here because it allows ERNIE LLM to receive more user feedback and
get updated before competitor products kick in. We have already heard from some enterprises
which started using ERNIE, commenting that the quality of ERNIE's services have significantly
improved over a short two to three weeks' time.
One of the key concerns we keep hearing is about Baidu's content ecosystem and management
execution capabilities. The root cause here is management strategies focused too much on tech
innovation without providing a clear monetization pathway with the result being technology ends
up a cash sink in the long run. The risk here is that the same story repeats itself, where Baidu
gains some traction but not sufficient scale/share to justify the costs it has put into the tech in
hopes of chasing the leader status it never obtains (ends up in a vicious cycle as it invests more
in tech but neglects the application side).
SenseTime
SenseTime has set its strategy to be the leading enterprise provider, choosing to differentiate on
its model-as-a-service proposition (MaaS) — similar to Alibaba. Its recent launch of SenseNova
LLM was notable for the number of developer tools it focused on, promising to meet 80% of use
cases, thereby cutting down coding time by 65%, and touting the platform's ease of use. The fact
that it is not a large Internet platform also stands in its favor, as application providers don't have
to worry about the platform competing with their own products.
On top of LLM offering, SenseTime also showcased SenseCore AI infrastructure, its proprietary
distributed computing platform, which could be a key asset in an environment where computing
power is likely a bottleneck. Leverage SenseTime's Artificial Intelligence Data Centre (AIDC),
SenseCore integrates AI chips and AI sensors to offer a strong computing power foundation,
supporting enterprise users' needs on model development and training. SenseTime AIDC has
27,000 GPUs and can support 5,000 petaFLOPs of computational power, enabling concurrent
training of 20 models, each comprising100 billion parameters.
However, the enterprise-only focus could also be an Achilles heel. SenseTime has chosen to
focus on providing LLM and AI infrastructure to enterprise users, empowering them to develop
and train customized models. The downside of this approach is that limited data will be retained
by SenseTime — given its lack of enterprise application scale, at least compared to the likes
of Alibaba, Tencent, Bytedance, and Baidu. Additionally, its strategy of gaining traction in the
enterprise space is to offer customized use cases, which will require implementation work with
end users. In a landscape where there would be two or three enterprise providers (Alibaba, Baidu,
etc.), our learnings from other software areas is this type of work quickly ends up seeing margin
squeeze, and the road to eventually gaining pricing power might be long.
BERNSTEIN
290 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
VALUATION METHODOLOGY See Disclosure Appendix for valuation methodology.
EXHIBIT 6: Ratings and target prices
05/25/23
Ticker Rating Currency Closing Price Target Price
BABA O USD 78.78 130.00
9988.HK O HKD 77.80 128.00
700.HK O HKD 322.40 455.00
BIDU O USD 118.67 160.00
9888.HK O HKD 116.50 157.00
MXAPJ 503.58
SPX 4,151.28
Source: Bloomberg, Bernstein estimates and analysis.
RISKS See Disclosure Appendix for risks.
Robin Zhu robin.zhu@bernstein.com +852 2918 5733
Boris Van boris.van@bernstein.com +852 2918 5753
Mark L. Moerdler mark.moerdler@bernstein.com +1 212 756 1857
Ronald Ma ronald.ma@bernstein.com +852 2918 7870
Ke Li ke.li@bernstein.com +65 6230 2354
Xuan Ji xuan.ji@bernstein.com +852 2918 5342
BERNSTEIN
CHINA AI: GETTING SMARTER...SCORING THE GENERATIVE AI CONTENDERS 291
BERNSTEIN
DISCLOSURE APPENDIX
I. REQUIRED DISCLOSURES
Autonomous Research US is a unit within Sanford C. Bernstein & Co., LLC , a broker-dealer registered with the U.S. Securities and
Exchange Commission and a member of the Financial Industry Regulatory Authority (www.finra.org) and the Securities Investor
Protection Corporation (see www.sipc.org). When this report contains an analysis of debt securities, such report is intended for
institutional investors and is not subject to all the independence and disclosure standards applicable to debt research for retail
investors under the FINRA rules.
VALUATION METHODOLOGY
European Food Retail
We value stocks in our coverage in the following stages: 1. We use a market-based approach to valuation. We take data for a
set of comparable companies and assess how multiples relevant to the sector (PE, EV/EBITDA, EV/Sales, EV/EBIT, FCF Yield)
change relative to expected growth rates, creating a regression of each multiple versus expected growth. 2. We generate earnings
forecasts for the company. We compare those forecasts to consensus expectations, and seek to reflect events that may happen
during the 12 months that are likely to move consensus expectations. 3. We value the stock by applying the relevant multiple (as
determined by our industry valuation regressions) to our earnings forecast. 4. Where appropriate we break down the company
into its parts (e.g., by geography) and value it as a sum of those parts. 5. Note that we make a number of adjustments to our
valuation analysis: (1) for company-specific tax rates, habits of recurring one-off charges, or other company-specific traits, (2) to
separate non-operating assets if we feel their inclusion is distorting the valuation multiples, and (3) to include pension deficits,
non-operating provisions, and seasonality of debt in our net debt calculation.
US Emerging Internet
We value companies in our coverage by triangulating a combination of our long-term DCF models (across all names) and a comps-
based approach on forward EV/Sales, EV/EBITDA, and P/E multiples (metric varies by stock).
European Medical Devices & Services
Our valuation analysis is based on two primary approaches – relative valuation based on price to forward earnings (forward P/
E) metrics, and a discounted cash flow (DCF) analysis. For the relative P/E valuation, we apply a sector specific growth adjusted
price-to-2022E earnings multiple (P/2022E EPS), derived from the relationship between price and the forecast 2021-2024E
earnings per share (EPS) growth for comparable medical device stocks.
U.S. Semiconductors
We value companies in our coverage using a combination of Enterprise Value to Sales, Enterprise Value to EBITDA, Price to EPS,
and/or Price to Book multiples.
India Technology, Media & Internet
The India technology services business has multiple coverage companies with different characteristics. Large sized players with
a complete portfolio. Mid-sized players with focus on certain segments like engineering services. We also include media and
internet companies in this sector. To arrive at our price targets, we use a combination of discounted cash flow and price to earnings
multiples and benchmark P/E to historical averages.
European Capital Goods & Industrial Technology
We believe that over time, valuation for EU Capital Goods is driven by the ability of a company to generate cash flow and grow its
business. Several names in our sector are also used by the market as bond proxies, and so their valuation moves with 10-yr bond
yields. We find the market considers both attributes in reference to the wider economy, and so it is the relative performance of
292 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
BERNSTEIN
each that appear to be the principal driver of valuation.
We find the market's preferred valuation metric in this sector is EV/EBITDA, relative to the MSCI Europe, and 24-months forward
(NTM+1). Our target multiple is based upon this metric. We use our proprietary holistic valuation model to derive a fair value target
multiple for each stock, based upon their cash generation relative to the economy and to their sector peers.
We apply our target multiple to our forecast for the company's EBITDA 24m-forward to give enterprise value. We deduct our
expectation for net debt and minority interests in the period for which the multiple is applied. This generates our target price,
typically reflecting the latest FX rates. We do not use our discounted cash flow analysis to derive target price, albeit it is calculated
alongside as a reference point.
U.S. Machinery
We calculate twelve-month target prices for our coverage using a mix of P/E and EV/EBITDA methodologies based on each
company's mode of value creation. We use multiples from the appropriate place in the cycle to triangulate our valuations.
Asian Industrial Technology
We use EV/EBITDA multiple as the primary valuation method. We set the target multiple referencing previous cycle peaks but
adjust for specific situations of the current cycle, apply it on the upcoming cycle peak to get the enterprise value, and discount
it back to derive our price target. We use DCF as reference for the company's long-term intrinsic value. As we move along the
different stages of a cycle, the time-dependent target price may temporarily deviate from the DCF-implied value. Currently,
because the 12-month target price date sits in a solid upcycle and approaches the cycle peak, our price targets are higher than
the DCF-implied value for most companies.
Asia Software Technologies
We value our coverage universe of China Software companies with a combination of methods namely: 1) DCF to act as a reference
to ground our valuation especially for longer dated incomes 2) Forward P/E multiples and P/S multiples, and also cross check to
EV/EBITDA multiples where appropriate. For the multiples approach. we evaluate our coverage companies for their stability of the
operating margin and growth profile and for those companies that are on a steady trajectory of earnings, we use a P/E multiple
and for those with limited earnings capability to-date, we apply a P/S multiple. 3) sum-of the-parts analyses. On a relative basis
we also compare our coverage stocks with US and China software peers on the basis of forward P/E multiples and P/S versus
the forward revenue growth and free cash flow margins.
China Internet
We value our coverage stocks using a combination of methods, including (1) forward valuation multiples including PE, EV/sales,
and P/GMV; (2) DCF; (3) sum-of-the-parts analyses; and (4) top-down estimates for medium-term market share and profitability.
On a relative basis we also compare our coverage stocks with US and China internet peers on the basis of forward EV/sales
multiples versus the sum of forward revenue growth and free cash flow margins.
Global Software
For Global Software, we value our companies using a mix of relative P/FE, DCF and sum-of-the-parts methodologies. We value
shares based on our estimate for 12-month NOPLAT in 1 years' time and apply an adjusted P/FE multiple. We then add back in
the net cash per share, discounted at 15% to account for potential tax costs and other “friction” to repatriate all cash, arriving at
our price target.
ARTIFICIAL INTELLIGENCE: GETTING SMARTER 293
BERNSTEIN
Asian Insurance
We value the Asian insurance companies using a forward price-to-book (PB) approach, reflecting the balance sheet-driven nature
of the insurance business. Our PB approach is driven by IFRS operating profit after-tax earnings, Return on Equity (RoE) and
adjusted Cost of Equity (CoE), explicitly balancing growth, return and risks. When applicable, we have also applied discounts
to listed companies' market value to reflect our views on earning sustainability, given 1) ongoing regulatory risks 2) intensified
competition, and 3) uncertainties of each company's business transformational reform or execution.
Pan-Asia Healthcare
We use sum of the parts valuation approach with DCF to value the specialty & biosimilar businesses and 1-year forward PE for
the generics business.
European Logistics
We value the stocks in our coverage principally on 13-24m P/E, with the multiple chosen using a combination of historic rates
and multivariate regression models.
Global Hotels & Leisure
We primarily value our companies using a combination of EV/EBITDA, relative P/E and discounted cash flow analysis. Our target
price is a subjective combination of the approaches. We benchmark our PE and EV multiples against peer companies adjusting,
where appropriate, for cost of capital, relative growth and ROIC. For our DCF we do 5 years of fully detailed estimates, a further
5 years of estimates where we only consider changes to revenue growth, NOPAT margin and ROIC and then calculate a terminal
value beyond that.
Asia Logistics and Travel
For the airlines in our coverage we apply a consistent framework of EV/EBITDA and use MSCI ACWI Airlines Index as our
benchmark. We maintain dual A- and H-share rating when stocks have both categories of shares listed on the relevant exchanges.
For airlines listed on multiple exchanges of Hong Kong and China, we derive our A-share target prices by translating the H-share
target prices from HKD to RMB, and apply a trading value difference based on historical trend.
We value the Chinese express delivery companies using forward price-to-earnings (P/E) (or EV/EBITDA if the company do not
have earnings in the forecasted period), backed by conservative discounted cash flow analysis (DCF). Valuation based on future
earnings reflects our view that the value creation of this group is mainly driven by future growth potential, which cannot be
adequately captured with near term earnings, or is reflected in the P/E of the same industry companies from other regions.
We value the Chinese duty-free companies using their 5-year forward earning forecast and target P/E multiple, and discount back.
We also back our valuation by a conservative discounted cash flow analysis (DCF).
Global Luxury Goods
Luxury goods stocks tend to trade short-term on organic growth positive / negative surprises. Longer-term, we believe there is
value on taking a more structural stance. We have a multi-pronged proprietary methodology to ascertain structural appeal. We
use target relative PEs to establish our price targets, and gear those target relative PEs to our structural assessment scores. We
make an exception for Farfetch, where we use a target EV / Sales multiple, using a correlation of EV/Sales to “take rate” with a
number of other platforms.
294 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
BERNSTEIN
China Consumer
We value our companies in China Household appliances, China cosmetics and China grocery based on target next-twelve-month
price-to-earnings (NTM P/E) multiples.
We select the target NTM P/E based on company's profit growth and return on invested capital (ROIC). We believe that stocks
with higher long-term growth rates and higher ROIC deserve higher multiples and so we apply incremental company premiums
or discounts to individual stocks to reflect their outlook for growth and returns.
We use a blended forward EPS estimates of FY2023 and FY2024 to set our 1-year target prices.
U.S. Apparel & Specialty Retail
We generally use three valuation approaches to reach our target prices, with some differences across our four sub-sectors. First,
we use regression-based multiples (either P/E or EV/Sales) based on sub-sector regressions with strong predictive power, using
a basket of selected comps for each sub-sector. Second, we compare the historical premium or discount versus the sub-sector
or versus key comp stocks. Third, we use a 10-year DCF. In some cases, we add additional metrics such as SOTP or cross-sector
regressions to supplement the three approaches above.
U.S. Internet
We value companies in our coverage using a combination of DCF, and forward Enterprise Value to EBITDA and Price to EPS
multiples.
U.S. Healthcare Services
For the following 6 companies, i.e. ELV, CI, CNC, CVS, UNH and HUM our preferred valuation methodology is relative (to S&P)
price-to-forward-earning (P/FE) due to the predictive NTM results in quantile analysis across time periods, as well as the relatively
strong and stable earnings generating capability of the companies' mature business. We base the companies' valuation on our EPS
estimates 12-months forward, multiply it by the respective absolute P/FE ratio for each company to arrive at our target prices.
For OSH, our preferred valuation methodology is relative (to S&P) price-to-forward-earning (P/FE) and Relative Price /Revenues
in Year 10 that is discounted back to establish a price target for 12 months out. In this approach we forecast the next ten years of
revenues for OSH with our published model for five years (annual revenue growth rates range from 50% to 39%) and projected
growth rates for years 6-10 (declining to 25%).
For HCA, our preferred valuation methodology is relative (to S&P) EV-to-forward-EBITDA (EV/FEBITDA) due to the predictive NTM
results in quantile analysis across time periods, its high degree of financial leverage, as well as the relatively strong and stable
earnings generating capability of the companies' mature business. We base HCA's valuation on our EBITDA estimate 12-months
forward, multiply it by the absolute EV/FEBITDA ratio to arrive at our target price.
In addition to P/FE and EV/FEBITDA metrics, we also consider other valuation metrics including sum-of-the-parts, PEG, FCF Yield,
and discounted cash flow, when arriving at the target price across our coverage. In addition, we acknowledge that our coverage
companies generate healthy amounts of cash and often maintain relatively conservative balance sheets, suggesting potential
further upside through effective capital allocation over the investment horizon.
U.S. Restaurants
We generally use two valuation approaches to reach our target prices: (1) multiple approach, where we apply NTM P/E multiples
based on relative performance of the company versus the S&P500 index, and (2) DCF approach, where we run a 15-year projection
DCF.
ARTIFICIAL INTELLIGENCE: GETTING SMARTER 295
BERNSTEIN
South & SE Asia Consumer Tech
We value companies on discounted cash flow methodology with WACC in the 11.5% range. The terminal year is considered 2030,
and terminal growth of 5% is assumed. We then compare with a peer group based on EV/sales and EV/GMV multiples from the
valuation we arrive at for each company in our coverage.
RISKS
European Food Retail
There are certain risks that are common to the all companies in our coverage: Prevailing economic conditions – In each of the
territories our coverage companies operate in, the food retail spend is correlated to the prevailing economic conditions. Thus any
unexpected deterioration or improvement in the macroeconomic conditions in these countries will impact the growth assumptions
applied to those operations. New Entrants – All companies in our coverage are at risk from new entrants either at a local/regional
level (i.e. a new supermarket opening locally to an incumbent) or national level (a new entrant entering a whole market). Currently
the greatest expansion is being seen at the lower (Lidl/Aldi in the discount sector) and higher (Waitrose/Wholefoods) ends of the
market or online (Amazon). These companies may continue to outpace the sector and impact the growth of the companies in our
sector. Similarly successful operators in certain regions/countries, e.g. E.Leclerc in France, could expand beyond their current
boundaries. As a lot of the non-coverage companies are privately held, it can be difficult to assess the ability and willingness of
these companies to expand further.
US Emerging Internet
•Global macro conditions: Our sector's revenues are primarily derived from consumer discretionary markets (e.g., retail/
eCommerce, mobility services, out-of-home restaurant consumption). Any sustained decline in economic conditions and
consumer spending can have a material negative impact on revenue growth across the sector.
•Rising rates: We cover growth stocks where much of the value is derived from expected cash flows in the terminal year. To
the extent that rates continue to rise, it can have an outsized impact on the multiples applied to our coverage given long-date
discounting.
•Regulation: While most of our companies are sheltered from the regulatory actions facing big tech, this increased level of
oversight can become an issue in the future as our coverage companies scale up. It can also restrict M&A in the near-term (both
as acquirers and targets). Gig economy marketplaces face pressing regulatory activity around the classification of drivers and
commission rate caps on the delivery side.
•Global competition: The Internet, more than any other industry, is susceptible to new and emerging competitive threats that
seemingly disrupt entire ecosystems and value pools. With emerging fast-growing tech companies domestically and abroad, it
stands the reason that new competitors will emerge that could reduce short-term revenue growth and destroy entire revenue
pools long-term. Stiff competition can also come from the big tech companies that have superior scale, distribution, and CAC
advantages; the mega-caps continue to diversify their revenue streams today.
•Cyber attacks: These attacks can severely impact the trust and engagement of platform users, resulting in a significant impact
to stock price.
European Medical Devices & Services
The risks to the European medical device stocks in our coverage include: the impact of healthcare reform, tax code reform, or
other policy initiatives which could negatively impact product utilization, pricing, and competitiveness. The risk of deteriorating
macro-economic conditions that may impact spending on healthcare which could cause an unexpected drop in product sales or
demand for healthcare services. Companies could be subject to product recalls, FDA warning letters, or government enforced
296 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
BERNSTEIN
actions which could negatively impact sales and operations. Unexpected fluctuations in foreign currency could impact earnings
in a positive or negative manner.
U.S. Semiconductors
The greatest sector-wide risk that could affect all of the stocks in our coverage is the macroeconomic environment and resulting
impact on revenues and sentiment. Upside risk to our targets exist if global GDP growth is quicker than we currently anticipate,
which would result in stronger semiconductor / semicap industry growth than we currently forecast. Conversely, if GDP growth
is slower than expected, this would result in slower growth for the industry and semiconductor / semicap companies. Recent
increasingly negative geopolitics, and of course the coronavirus pandemic, represent further potential risks to our broad coverage.
India Technology, Media & Internet
The downside risks to India Technology Services sector include any macroeconomic downturn that could impact demand
environment. Currency headwinds from rupee appreciation could impact margins. Immigration related issues or protectionist
measures in US or Europe could significantly increase operational complexities.
European Capital Goods & Industrial Technology
As industrial staples, the key risk to European Capital Goods is a slowdown in manufacturing, industrial production and the wider
economy. The majority of our names would be negatively impacted by such a slowdown. Our stocks are also valued relative to
the wider economy, and so their valuation moves up and down with general sentiment on equities. In both cases our target prices
would be significantly impacted by any material move in the wider stock market.
U.S. Machinery
Upside/Downside risks include the following: 1) a better/worse than expected cyclical recovery; 2) higher/lower market share
gains/losses; 3) higher/lower products penetration; 4) better/worse cost structure management; 5) more/less aggressive
deployment of balance sheet.
Asian Industrial Technology
The risks to our coverage names are mainly associated with the global macro economy, including industrial capex cycles, trade
frictions and currency. US companies' share prices are sensitive to their quarterly results relative to management guidance and
consensus forecasts. Japanese and Chinese companies are much less so. For IPGP and Harmonic Drive, as they have >50% of
global share in their respective industries, potential change in competitive landscape would be a bigger risk to them than to other
companies.
Asia Software Technologies
The target price for our coverage companies are subject to a number of macroeconomic and company specific risks that include:
1) changes in Chinese macroeconomic growth which impacts both financial and human resources that customers are able to
deploy; 2) changes in degree of competitive activity within China and overseas market; 3) changes in Chinese government strategy
and industry priority; 4) Changes in the nature of our covered companies' relationships with their key customers, partners and/or
suppliers; 5) Changes in our covered companies' ability to deliver on anticipated growth and/or margin improvement opportunities,
whether due to internal or external causes (including the unsuccessful integration of acquired companies)
China Internet
The risks to our views on our China internet stocks and our price targets include (1) macroeconomic risks, including liquidity in the
Chinese economy, and retail consumption trends; (2) changes in consumer preferences and engagement with specific brands and
online platforms; (3) competition – both between other internet companies and offline peers; and (4) regulatory risk, for example
ARTIFICIAL INTELLIGENCE: GETTING SMARTER 297
BERNSTEIN
related to China's anti-monopoly regulations. Tensions between the US and China could create political risks which may affect
our coverage companies.
Global Software
Our price targets for ADBE, CRM, CTXS, MSFT, ORCL, SABR, SAP, SPLK, VMW, and WDAY are subject to a number of
macroeconomic and company specific risks that include: The potential of a recession Changes in the degree of competitive activity
within any key market Foreign exchange fluctuations Changes in the nature of our covered companies' relationships with their
key customers, partners and/or suppliers Changes in our covered companies' ability to deliver on anticipated growth and/or
margin improvement opportunities, whether due to internal or external causes (including the unsuccessful integration of acquired
companies) Changes in our companies' stances toward M&A or the prioritization of cash in general Adverse situations in one of
its key markets
Asian Insurance
The predominant downside risk to our sector outlook is a region-wide economic slowdown leading to insurance premiums
contraction. A prolonged period of stagnant economic growth would reduce premium growth and delay the demand shift towards
protection products. Market risks are significant and interest rate volatility would also place distortion on both new business sales
and in-force book return. Investment volatility, especially short-term equity market fluctuations, could create a negative impact
on both the IFRS earnings and balance sheet stability of the life insurers. Near-term headwind remains on the local government's
stringent zero-COVID policies. Lastly, regulatory overhauls on the broad financial and property industries in each of the key Asia
markets could trigger further business slowdown or increase life insurers' risk exposure.
Pan-Asia Healthcare
Risks to the pharmaceutical industry include a) risk of pipeline products failing or getting delayed due to FDA actions, b) possibility
of adverse litigation outcomes delaying key generic launches, c) cGMP non-compliance in manufacturing facilities leading to FDA
actions like Warning Letters or Import Alerts to plants, d) product recalls or other product safety issues, e) pricing pressure from
market factors or price control regulations, f) supply and logistics disruptions and f) healthcare regulations and reforms.
European Logistics
European logistics stocks are highly sensitive to the macroeconomic cycle. Most stocks in our coverage are geared to global trade
and demand for goods, both as inputs for manufacturing and for final consumption. Since the 2010s, global trade has largely
grown in line with global GDP.
The end industry matters for different transport modes. On ocean, retail dominates with c. 50% of containerized flows supporting
this industry: broadlines, electronics and apparel are the largest. A further c. 25% of flows support homeware: furniture and DIY.
Among other sectors, food and beverage and automotive are also impactful.
In air, different products dominate. Perishables are significant by volume but only relevant for some of the air freight forwarders
in our coverage. Otherwise, complex manufactured goods and high-value items with short lead times will often travel by air:
machinery and electronics are important users of airfreight. Ecommerce penetration also tends to benefit airfreight as short lead
times may be requred to dispatch from the other side of the world.
From a unit profitability perspective, freight forwarders' earnings tend to rise in an environment of supply chain complexity.
Forwarders exist to help clients solve their supply chain problems, and can bring significant logistics expertise to bear. Greater
complexity for clients thus translates to greater demand for forwarding services. The converse is also true: where logistics is easier,
forwarders may see demand dip.
Finally, ecommerce penetration is important for ecommerce-exposed stocks: particularly Deutsche Post and IDS. As shopping
shifts online, more parcels are required to be delivered, supporting volumes and earnings. Ecommerce penetration also has a
298 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
BERNSTEIN
positive impact on the contract logistics operations of freight forwarders.
Global Hotels & Leisure
The leisure sector is reliant on consumer spending and therefore is susceptible to changes in consumer spending and the broader
macroeconomic environment. Any slowdown in these trends will affect revenues and earnings and market sentiment towards
our coverage. For the Hotel and Travel stocks particularly, there is a risk of terrorism or other geo-political events changing the
demand for international and domestic travel. There are a wide range of disruptors who pose a potential risk (Airbnb, Expedia,
Uber Eats) to our coverage and any increase in their in roads into our segments could result in market share losses and revenue/
earnings declines.
Asia Logistics and Travel
ASIA LOGISTICS AND TRAVEL
The Asia Logistics and Travel companies that we cover are subject to macroeconomic risks, including exposure to overall economy
growth, trade volume, interest rates, foreign exchange rates, etc.; as well as competitive landscape changes, brought by new
entrants and new technology that may disrupt the market game. CHINESE AIRLINES INDUSTY The airline industry is highly
regulated and is subject to the risks of changes in government regulations or policies. Fuel prices are critical to the profitability of
airlines, especially to the covered companies that do not have any oil price hedge programs. Airlines' earnings are also sensitive
to foreign exchange fluctuations, apart from a rather significant amount of debt dominated in foreign currency from the purchase
of aircrafts and other fixed assets, passenger travel sentiments and fuel costs are also subject to RMB exchange rate against
foreign currency. CHINESE TRAVEL INDUSTRY The Chinese duty-free industry is highly regulated, all duty-free shops must obtain
licenses from the government and the sale of goods should comply with customs’ supervision and regulations. The operators are
therefore subject to the risks of changes in government regulations or policies.
Global Luxury Goods
Covid-19 triggers at least five of the ten risks of luxury (below), precipitates a material downward correction to GDP growth and
leads to a sharp decline in consumer demand - and possibly medium-term damage to consumer confidence and propensity to
spend. We are on “terra incognita” in terms of duration, impact and effectiveness of measures, as this scenario has become worse
than 2008. More uncertainty remains regarding the plummeting oil price, upheavals in Hong Kong and the Sino-American trade
confrontation. Luxury is cyclical and would suffer a triple whammy blow in a recession: 1) slower or negative top line growth would
cause: 2) operating de-leverage – as luxury is a fixed cost industry. Valuation multiples would typically contract in that environment;
Luxury sales thrive on customers feeling affluent and secure in their wealth. A higher interest rate environment would dampen
asset prices and cause the richer to feel poorer: this would be a severe blow to luxury. Asset price trends are important to support
confidence of luxury consumers. The Chinese real estate market and the US stock market are the bellwethers. Higher taxation
of upper income brackets, higher property taxes or other government actions to reduce the Gini coefficient would be a sector
headwind.
Luxury thrives on people travelling, and on a limited number of global cities. Terrorist attacks (e.g. 9/11), tighter custom controls
(especially in China) and epidemics (e.g. SARS) would be a risk for luxury – as fewer consumers would travelling and spending
money abroad. Luxury is dependent on a small number of cities: 25 of them account for more than 2/3 of luxury spend - Paris,
Hong Kong and New York being the top three. Serious problems in any of the top luxury cities would be a sector headwind –
partially compensated by consumers shopping elsewhere/ and – increasingly – online;
FX would also be a risk for the sector. European luxury goods companies thrive on a weaker EUR and stronger USD. American
luxury goods companies are the mirror image to that. A weaker CNY causing Chinese consumers to spend more in Mainland China
would be a headwind: prices in China are higher, price elasticity would reduce overall spend – all else being equal.
ARTIFICIAL INTELLIGENCE: GETTING SMARTER 299
BERNSTEIN
China Consumer
China Grocery
The China Grocery sector is subject to macroeconomic risks including overall economy growth, disposable income growth,
inflation, etc. The industry competitive landscape is exposed to disruptors such as e-commerce and technology. Other factors
such as natural disasters or weather conditions can also impact the supply side of the grocery sector.
China Household Appliances
The China Household Appliances sector is subject to macroeconomic risks including overall economy growth, disposable income
growth, inflation, etc. Additional factors that may impact the supply side include: fluctuations in raw material prices, channel
staffing and inventory issues caused by industry players, price war competition caused by industry players, and threats from
foreign brands becoming more aggressive in their competitive strategies. Additional factors that may impact the demand side
include: government policy, unfavorable weather conditions impacting sales particularly for air conditioners, real estate market
lower than anticipated, market sentiment downturn leading to consumers purchasing lower-end products. The export business
segment is also exposed to US-China trade war, which may result in slower growth in exported appliances.
China cosmetics
The China cosmetic sector is one the highest levels of opening-up consumer sectors to foreign brands. The competition between
domestic and foreign players could become increasingly intense if more international brands enter China. The fast development
of cosmetics e-commerce gives Chinese brands opportunities to grow revenue significantly, but it also pushes up the cost of
online marketing, platform capex, and consumer subsidies, thereby possibly reducing the overall sector profitability. China's overall
cosmetic market growth also depends on the disposable income growth and per capita spending on cosmetics, both of which are
sensitive to changes in macroeconomic conditions.
U.S. Apparel & Specialty Retail
Across the Apparel and Specialty Retail sector, we note a number of macro risks that may impact our coverage. Any material
changes in consumer preferences or behavior towards discretionary spending is a major risk factor to our sector – these include
changes in economic outlook, consumer sentiment, income inequality, tax rates, or government restrictions on consumer behavior
(including related to COVID-19). Our stocks are also impacted by fiscal and economic policy changes including interest rates
and inflation, corporate tax rates, tariffs, sanctions and import / export rules. Finally, as a discretionary shopping sector, we are
impacted by natural elements including weather, seasonality and natural disasters that may make it difficult or unreasonable for
consumers to purchase goods.
U.S. Internet
- Global macro conditions: our sector's revenues are primarily generated from advertising dollars and consumer spend. Any
sustained decline in economic conditions, economic outlook, or burdens from a potential trade war can have a material negative
impact on revenue growth potential across the sector. - Anti-trust regulations & litigation: Most of our sector is currently being
investigated by the DOJ, FTC, or international regulatory bodies for anti-competitive, anti-trust behavior. Regulating big tech has
become a bi-partisan initiative in the United States with reasonable expectations that some type of new regulation will prevail.
Outsized risk remains if new regulations result in compounding cost of compliance, severely limiting revenue growth, and full or
partial break-up of the companies all together. - Privacy regulations: Almost every company in our coverage sector is involved
in on-going litigious lawsuits surrounding the capturing and usage of personal data. Any negative outcomes can set challenging
precedents resulting in a materially different data collection and usage practices. Most exposed are ad supported businesses
where data collection is the primary value contributor to providing desired ad targeting and attribution capabilities to advertisers. -
Cyber attacks: similarly, almost all of our companies have recently experienced some type of cyber attack. Continued cyber attacks
and/or a major attack can severely impact the trust and engagement of platform users, resulting in a significant impact to stock
300 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
BERNSTEIN
price. - Global competition: The Internet, more than any other industry, is susceptible to new and emerging competitive threats
that seemingly disrupt entire ecosystems and value pools. With emerging fast-growing tech companies domestically and abroad,
it stands the reason that new competitors will emerge that could reduce short-term revenue growth and destroy entire revenue
pools long-term.
U.S. Healthcare Services
Price targets for all our covered companies are subject to full range of domestic U.S. macro-economic risks, such as GDP growth,
unemployment rate, the pace of population aging, inflation and interest rate dynamics to fiscal spending, especially on healthcare,
on both federal and state levels. As some of our covered companies continue to increase international presence outside of the U.S.,
currency fluctuations will become a more substantial risk. A number of industry specific factors will have significant impacts on the
companies' future earnings, including medical cost trends, premium rate trends for government businesses and public exchange,
industry-wide health insure tax, government spending on healthcare, and government regulations on healthcare costs, such as
pharmaceuticals. That said, in most cases, the key drivers to outperformance against industry peers and attractive shareholder
return is each company's ability to generate organic growth, achieve market share gains, execute on margin expansion plans (and
integration initiatives post mergers for covered companies), and allocate capital efficiently and effectively. Finally, the valuation of
the broader market has recovered but is subject to higher growth expectations and market volatilities. The valuation of the broader
market might contract if we don’t see quality growth meeting market expectations and this would also impact the valuation of our
covered companies.
U.S. Restaurants
Our view on the U.S. Restaurant sector is subject to a number of macro and idiosyncratic risks that could impact our earning
estimates and price target. Macro risks include unexpected strengthening/weakening in consumer demand, higher/lower than
expected commodity costs, and, if relevant, strengthening/weakening of the dollar against other major currencies. Idiosyncratic
risks include better/worse than expected performance by product launches, failure to anticipate/respond to competition and
transformations.
South & SE Asia Consumer Tech
The sector is still in the early stages of growth, with most companies unprofitable. Hence, the risk could be led by higher than
expected cash burn, leading to a requirement for more funding and equity dilution. As Super Apps get created, and companies
move into adjacent verticals, higher competition is a risk.
RATINGS DEFINITIONS, BENCHMARKS AND DISTRIBUTION
Bernstein brand
The Bernstein brand rates stocks based on forecasts of relative performance for the next 6-12 months versus the S&P 500 for
stocks listed on the U.S. and Canadian exchanges, versus the Bloomberg Europe Developed Markets Large & Mid Cap Price
Return Index (EDM) for stocks listed on the European exchanges (except for Russian companies), versus the Bloomberg Emerging
Markets Large & Mid Cap Price Return Index (EM) for Russian companies and stocks listed on emerging markets exchanges
outside of the Asia Pacific region, versus the Bloomberg Japan Large & Mid Cap Price Return Index USD (JP) for stocks listed on
the Japanese exchanges, and versus the Bloomberg Asia ex-Japan Large & Mid Cap Price Return Index (ASIAX) for stocks listed
on the Asian (ex-Japan) exchanges -unless otherwise specified.
The Bernstein brand has three categories of ratings:
•Outperform: Stock will outpace the market index by more than 15 pp
•Market-Perform: Stock will perform in line with the market index to within +/-15 pp
•Underperform: Stock will trail the performance of the market index by more than 15 pp
ARTIFICIAL INTELLIGENCE: GETTING SMARTER 301
BERNSTEIN
Coverage Suspended applies when coverage of a company under the Bernstein research brand has been suspended. Ratings and
price targets are suspended temporarily. Previously issued ratings and price targets are no longer current and should therefore
not be relied upon.
Not Rated: The stock Rating, Target Price and/or estimates (if any) have been suspended temporarily.
Autonomous brand
The Autonomous brand rates stocks as indicated below. As our benchmarks we use the Bloomberg Europe 500 Banks And
Financial Services Index (BEBANKS) and Bloomberg Europe Dev Mkt Financials Lrg & Mid Cap Price Ret Index EUR (EDMFI)
index for European banks, the Bloomberg Europe 500 Insurance Index (BEINSUR) for European insurers, the S&P 500 and S&P
Financials for US banks coverage, S5LIFE for US Insurance, the S&P Insurance Select Industry (SPSIINS) for US Non-Life Insurers
coverage, and Ibovespa Brasil Sao Paulo Stock Exchange Index (IBOV) for Brazil and Hang Seng H-FIN (HSHFI-HK) index for China
banks and insurers. Ratings are stated relative to the sector (not the market).
The Autonomous brand has three categories of ratings:
•Outperform (OP): Stock will outpace the relevant index by more than 10 pp
•Neutral (N): Stock will perform in line with the relevant index to within +/-10 pp
•Underperform (UP): Stock will trail the performance of the relevant index by more than 10 pp
Coverage Suspended (CS) applies when coverage of a company under the Autonomous research brand has been suspended.
Ratings and price targets are suspended temporarily. Previously issued ratings and price targets are no longer current and should
therefore not be relied upon.
Not Rated: The stock Rating, Target Price and/or estimates (if any) have been suspended temporarily.
Those denoted as ‘Feature’ (e.g., Feature Outperform FOP, Feature Under Outperform FUP) are our core ideas. Not Rated (NR) is
applied to companies that are not under formal coverage.
For both brands, recommendations are based on a 12-month time horizon.
DISTRIBUTION OF RATINGS/INVESTMENT BANKING SERVICES
Rating Market Abuse Regulation (MAR) and
FINRA Rule 2241 classification
Count Percent Count* Percent*
Outperform BUY 387 49.87% 0 0.00%
Market-Perform (Bernstein Brand)
Neutral (Autonomous Brand)
HOLD 267 34.41% 1 0.37%
Underperform SELL 122 15.72% 0 0.00%
* These figures represent the number and percentage of companies in each category to whom Bernstein and Autonomous
provided investment banking services.
As of May 31 2023. All figures are updated quarterly and represent the cumulative ratings over the previous 12 months.
PRICE CHARTS/ RATINGS AND PRICE TARGET HISTORY
This research publication covers six or more companies. For price chart and other company disclosures:
Please visit: https://bernstein-autonomous.bluematrix.com/sellside/Disclosures.action.
Or, you can also write to the Director of Compliance, Sanford C. Bernstein & Co. LLC, 1345 Avenue of the Americas, New York,
N.Y. 10105.
CONFLICTS OF INTEREST
Nikhil Devnani maintains a long position in Reliance Industries.
302 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
BERNSTEIN
Nithya Balasubramanian and her spouse maintain long positions in Cipla Ltd. and Lupin Ltd. Ms. Balasubramanian was employed
by Cipla from September 2013 through August 2019.
All statements in this report attributable to Gartner represent Bernstein's interpretation of data, research opinion or viewpoints
published as part of a syndicated subscription service by Gartner, Inc., and have not been reviewed by Gartner. Each Gartner
publication speaks as of its original publication date (and not as of the date of this report). The opinions expressed in Gartner
publications are not representations of fact, and are subject to change without notice.
Stacy A. Rasgon maintains long positions in various crypto currencies.
An associate contributing to this report maintains a long position in Infosys Ltd.
Nikhil Nigania maintains a long position in Reliance Industries.
Bernstein and/or its affiliates exercise investment discretion over accounts or otherwise beneficially own 1% or more of the
outstanding common stock of the following companies: Etsy Inc, Cognex Corp, Adobe Inc, Burberry Group PLC, Elevance Health,
UnitedHealth Group Inc and Chipotle Mexican Grill Inc.
Bernstein provided non-investment banking-securities related services and received compensation for such services during the
past twelve months for the following clients: Microsoft Corp and Ping An Insurance Group Co of China Ltd.
Affiliates of Bernstein provided non-investment banking-securities related services and received compensation for such services
from the following clients during the past twelve (12) months: companies: Prudential PLC, China Life Insurance Co Ltd, AIA Group
Ltd and Yum! Brands Inc.
Rahul Malhotra maintains a long position in Tata Consultancy Services Ltd (TCS.IN).
Ronald Ma maintains a long position in Tencent Holdings Ltd and Alibaba Group Holding Ltd (700.HK, BABA and 9988.HK).
Richard J. Clarke maintains a long position in Koninklijke Philips NV (PHG and PHIA.NA).
Cherry Leung maintains a long position in Tencent Holdings Ltd (700.HK).
Kate Xiao maintains a long position in AIA Group Ltd (1299.HK).
Pearl Xu maintains a long position in NVIDIA Corp and Tencent Holdings Ltd (NVDA and 700.HK).
Aneesha Sherman maintains a long position in Alibaba Group Holding Ltd (9988.HK and BABA).
Firoz Valliji maintains a long position in NVIDIA Corp, Alphabet Inc and Elevance Health (NVDA, GOOGL and ELV).
Stacy A. Rasgon maintains a long position in Microsoft Corp, Salesforce.com Inc, Snowflake Inc and VMware Inc (MSFT, CRM,
SNOW and VMW).
Akhilesh Kumawat maintains a long position in AGCO Corp (AGCO).
Bill He maintains a long position in Airbnb Inc (ABNB).
Jonathon Unwin maintains a long position in Ocado Group PLC (OCDO.LN).
Melinda Hu maintains a long position in Alibaba Group Holding Ltd and Alphabet Inc (9988.HK, BABA and GOOGL).
Nikhil Devnani maintains a long position in Microsoft Corp (MSFT).
Eva Zhang maintains a long position in NVIDIA Corp (NVDA).
Alex Irving maintains a long position in Alphabet Inc (GOOGL).
Nikhil Nigania maintains a long position in HCL Technologies Ltd, Tech Mahindra Ltd, Infosys Ltd and Wipro Ltd (HCLT.IN,
TECHM.IN, INFO.IN, INFY, WPRO.IN and WIT).
Robin Zhu maintains a long position in Microsoft Corp (MSFT).
Ronald Ma maintains a long position in Hangzhou Hikvision Digital Technology Co Ltd, Alibaba Group Holding Ltd, Tencent Holdings
Ltd and Restaurant Brands International Inc (002415.CH, 9988.HK, BABA, 700.HK and QSR).
Boris Van maintains a long position in NVIDIA Corp, Snowflake Inc and Workday Inc (NVDA, SNOW and WDAY).
Amir Farahani maintains a long position in Alibaba Group Holding Ltd and Alphabet Inc (9988.HK, BABA and GOOGL).
William Robbins maintains a long position in NVIDIA Corp (NVDA).
Parth Shah maintains a long position in Larsen & Toubro Ltd (LT.IN).
Luca Solca maintains a long position in Alibaba Group Holding Ltd and Lonza Group AG (9988.HK, BABA and LONN.SW).
Renny Shao maintains a long position in AIA Group Ltd (1299.HK).
OTHER MATTERS
It is at the sole discretion of the Firm as to when to initiate, update and cease research coverage. The Firm has established,
maintains and relies on information barriers to control the flow of information contained in one or more areas (i.e. the private side)
within the Firm, and into other areas, units, groups or affiliates (i.e. public side) of the Firm.
The legal entity(ies) employing the analyst(s) listed in this report can be determined by the country code of their phone number,
as follows:
+1 Sanford C. Bernstein & Co., LLC
ARTIFICIAL INTELLIGENCE: GETTING SMARTER 303
BERNSTEIN
+44 Bernstein Autonomous LLP
+353 Sanford C. Bernstein Ireland Limited
+91 Sanford C. Bernstein (India) Private Limited
+852 Sanford C. Bernstein (Hong Kong) Limited 盛博香港有限公司
+65 Sanford C. Bernstein (Singapore) Private Limited
+81 Sanford C. Bernstein Japan KK
CERTIFICATION
Each research analyst listed in this report, who is primarily responsible for the preparation of the content of this report, certifies
that all of the views expressed in this publication accurately reflect that analyst's personal views about any and all of the subject
securities or issuers and that no part of that analyst's compensation was, is, or will be, directly or indirectly, related to the specific
recommendations or views in this publication.
II. OTHER IMPORTANT INFORMATION AND DISCLOSURES
References to "Bernstein" or the “Firm” in these disclosures relate to the following entities: Sanford C. Bernstein & Co., LLC,
Bernstein Autonomous LLP, Sanford C. Bernstein Limited (for dates prior to January, 1, 2021), Autonomous Research LLP (for
dates between April 1, 2019 and December 31, 2020), Sanford C. Bernstein (Hong Kong) Limited 盛博香港有限公司, Sanford
C. Bernstein (Canada) Limited, Sanford C. Bernstein (India) Private Limited (SEBI registration no. INH000006378) and Sanford C.
Bernstein (Singapore) Private Limited, which is a licensed entity under the Securities and Futures Act and registered with Company
Registration No. 20213710W.
Separate branding is maintained for “Bernstein” and “Autonomous” research products.
•Bernstein produces a number of different types of research products including, among others, fundamental analysis and
quantitative analysis, under both the “Autonomous” and “Bernstein” brands. Recommendations contained within one type of
research product may differ from recommendations contained within other types of research products, whether as a result of
differing time horizons, methodologies or otherwise. Furthermore, views or recommendations within a research product issued
under one brand may differ from views or recommendations under the same type of research product issued under the other
brand. The Research Ratings System for the two brands and other information related to those Rating Systems are included
in the previous section.
•Each operates as a separate business unit within the following entities: Sanford C. Bernstein & Co., LLC, Sanford C. Bernstein
(Hong Kong) Limited 盛博香港有限公司 and Bernstein Autonomous LLP. For information relating to “Autonomous” branded
products (including certain Sales materials) please visit: www.autonomous.com. For information relating to Bernstein branded
products please visit: www.bernsteinresearch.com.
Information related to the acquisition of Autonomous Research:
•On and as of April 1, 2019, AllianceBernstein L.P. acquired Autonomous Research. As a result of the acquisition, the research
activities formerly conducted by Autonomous Research US LP and Autonomous Research Asia Limited were assumed by
Sanford C. Bernstein & Co., LLC and Sanford C. Bernstein (Hong Kong) Limited盛博香港有限公司, respectively. Both entities
continue to publish research under the Autonomous brand.
•References to “Autonomous” in these disclosures relate to the Autonomous Research LLP and, with reference to dates prior
to April 1, 2019, to Autonomous Research US LP and Autonomous Research Asia Limited, and, with reference to April 1, 2019
onwards, the Autonomous Research US unit and separate brand of Sanford C. Bernstein & Co., LLC and the Autonomous
Research Asia unit and separate brand of Sanford C. Bernstein (Hong Kong) Limited 盛博香港有限公司, collectively.
Information related to the reorganization of Sanford C. Bernstein Limited and Autonomous Research LLP:
304 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
BERNSTEIN
•On and after close of business on December 31, 2020, as part of an internal reorganisation of the corporate group, Sanford C.
Bernstein Limited transferred its business to its affiliate Autonomous Research LLP. Subsequent to this transfer, Autonomous
Research LLP changed its name to Bernstein Autonomous LLP. As a result of the reorganisation, the research activities formerly
conducted by Sanford C. Bernstein Limited were assumed by Bernstein Autonomous LLP, which is authorised and regulated by
the Financial Conduct Authority (FRN 500498) and now publishes research under the Bernstein Research Brand.Please note
that all price targets, recommendations and historical price charts are unaffected by the transfer of the business from Sanford
C. Bernstein Limited and have been carried forward unchanged to Bernstein Autonomous LLP. You can continue to find this
information on the Bernstein website at www.bernsteinresearch.com.
Analysts are compensated based on aggregate contributions to the research franchise as measured by account penetration,
productivity and proactivity of investment ideas. No analysts are compensated based on performance in, or contributions to,
generating investment banking revenues.
This report has been produced by an independent analyst as defined in Article 3 (1)(34)(i) of EU 296/2014 Market Abuse
Regulation (“MAR”).
Where this material contains an analysis of debt product(s), such material is intended only for institutional investors and is not
subject to the independence and disclosure standards applicable to debt research prepared for retail investors.
This document may not be passed on to any person in the United Kingdom (i) who is a retail client (ii) unless that person or entity
qualifies as an authorised person or exempt person within the meaning of section 19 of the UK Financial Services and Markets Act
2000 (the "Act"), or qualifies as a person to whom the financial promotion restriction imposed by the Act does not apply by virtue
of the Financial Services and Markets Act 2000 (Financial Promotion) Order 2005, or is a person classified as an "professional
client" for the purposes of the Conduct of Business Rules of the Financial Conduct Authority.
This document may not be passed onto any person in Canada unless that person qualifies as "permitted client" as defined in
Section 1.1 of NI 31-103.
To our readers in the United States: Sanford C. Bernstein & Co., LLC, a broker-dealer registered with the U.S. Securities and
Exchange Commission (“SEC”) and a member of the U.S. Financial Industry Regulatory Authority, Inc. (“FINRA”) is distributing this
publication in the United States and accepts responsibility for its contents. Where this report has been prepared by research
analyst(s) employed by a non-US affiliate, such analyst(s), is/are (unless otherwise expressly noted) not registered as associated
persons of Sanford C. Bernstein & Co., LLC or any other SEC-registered broker-dealer and are not licensed or qualified as
research analysts with FINRA or any other US regulatory authority. Accordingly, reports prepared by such analyst(s) may not be
prepared in compliance with FINRA’s restrictions regarding (among other things) communications by research analysts with a
subject company, interactions between research analysts and investment banking personnel, participation by research analysts
in solicitation and marketing activities relating to investment banking transactions, public appearances by research analysts, and
trading securities held by a research analyst account.
To our readers in the United Kingdom: This publication has been issued or approved for issue in the United Kingdom by
Bernstein Autonomous LLP, authorised and regulated by the Financial Conduct Authority and located at 60 London Wall, London
EC2M 5SH, +44 (0)20-7170-5000. Registered in England & Wales No OC343985.
To our readers in Ireland and the member states of the EEA: This publication is being distributed by Sanford C. Bernstein
Ireland Limited, which is authorised and regulated by the Central Bank of Ireland.
To our readers in Hong Kong: This publication is being distributed in Hong Kong by Sanford C. Bernstein (Hong Kong) Limited
盛博香港有限公司, which is licensed and regulated by the Hong Kong Securities and Futures Commission (Central Entity No.
AXC846) to carry out Type 4 (Advising on Securities) regulated activities and subject to the licensing conditions mentioned in
the SFC Public Register (https://www.sfc.hk/publicregWeb/corp/AXC846/details)). This publication is solely for professional
investors only, as defined in the Securities and Futures Ordinance (Cap. 571).
To our readers in Singapore: This publication is being distributed in Singapore by Sanford C. Bernstein (Singapore) Private
Limited, only to accredited investors or institutional investors, as defined in the Securities and Futures Act (Chapter 289).
Recipients in Singapore should contact Sanford C. Bernstein (Singapore) Private Limited in respect of matters arising from, or in
connection with, this publication. Sanford C. Bernstein (Singapore) Private Limited is a licensed entity under the Securities and
Futures Act and registered with Company Registration No. 20213710W. It is regulated by the Monetary Authority of Singapore
and located at One Raffles Quay, #27-11 South Tower, Singapore 048583, +65-62302300.
ARTIFICIAL INTELLIGENCE: GETTING SMARTER 305
BERNSTEIN
To our readers in the People’s Republic of China: The securities referred to in this document are not being offered or sold and
may not be offered or sold, directly or indirectly, in the People's Republic of China (for such purposes, not including the Hong Kong
and Macau Special Administrative Regions or Taiwan), except as permitted by the securities laws of the People’s Republic of China.
To our readers in Japan: This document is not delivered to you for marketing purposes, and any information provided herein
should not be construed as a recommendation, solicitation or offer to buy or sell any securities or related financial products.
For the institutional client readers in Japan who have been granted access to the Bernstein website by Daiwa Securities
Group Inc. (“Daiwa”), your access to this document should not be construed as meaning that Bernstein is providing you with
investment advice for any purposes. Whilst Bernstein has prepared this document, your relationship is, and will remain with, Daiwa,
and Bernstein has neither any contractual relationship with you nor any obligations towards you.
To our readers in Australia: Sanford C. Bernstein & Co., LLC., Bernstein Autonomous LLP, Sanford C. Bernstein Ireland Limited,
Sanford C. Bernstein (Hong Kong) Limited 盛博香港有限公司, Sanford C. Bernstein (Singapore) Private Limited, and Sanford C.
Bernstein (India) Private Limited ("Bernstein Affiliates") are regulated, respectively, by the Securities and Exchange Commission
under U.S. laws, by the Financial Conduct Authority under U.K. laws, by the Central Bank of Ireland, by the Hong Kong Securities
and Futures Commission under Hong Kong laws, by the Monetary Authority of Singapore under Singapore laws, and Securities
and Exchange Board of India, all of which differ from Australian laws. The Bernstein Affiliates are exempt from the requirement to
hold an Australian financial services license under the Corporations Act 2001 in respect of the provision of the following financial
services to wholesale clients:
•providing financial product advice;
•dealing in a financial product;
•making a market for a financial product; and
•providing a custodial or depository service.
To our readers in Canada: If this publication pertains to a Canadian domiciled company, it is being distributed in Canada by
Sanford C. Bernstein (Canada) Limited, which is licensed and regulated by the Investment Industry Regulatory Organization of
Canada ("IIROC"). If the publication pertains to a non-Canadian domiciled company, it is being distributed by Sanford C. Bernstein
& Co., LLC, which is licensed and regulated by both the SEC and FINRA, into Canada under the International Dealers Exemption.
To our readers in India: This publication is being distributed in India by Sanford C. Bernstein (India) Private Limited (SCB India)
which is licensed and regulated by Securities and Exchange Board of India ("SEBI") as a research analyst entity under the SEBI
(Research Analyst) Regulations, 2014, having registration no. INH000006378 and as a stock broker having registration no.
INZ000213537. SCB India is currently engaged in the business of providing research and stock broking services.
•SCB India is a Private limited company incorporated under the Companies Act, 2013, on April 12, 2017 bearing corporate
identification number U65999MH2017FTC293762, and registered office at Level 6, 4 North Avenue, Maker Maxity, Bandra
Kurla Complex, Bandra (East), Mumbai 400051, Maharashtra, India (Phone No: +91-22-68421401).
•For details of Associates (i.e., affiliates/group companies) of SCB India, kindly email scbincompliance@alliancebernstein.com.
•SCB India does not have any disciplinary history as of the date of this report.
•Except as noted above, SCB India and/or its Associates (i.e., affiliates/group companies), the Research Analysts authoring this
report, and their relatives
•do not have any financial interest in the subject company
•do not have actual/beneficial ownership of one percent or more in securities of the subject company;
•is not engaged in any investment banking activities for Indian companies, as such;
•have not managed or co-managed a public offering in the past twelve months for the subject company;
306 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
BERNSTEIN
•have not received any compensation for investment banking services or merchant banking services from the subject
company in the past 12 months;
•have not received compensation for brokerage services from the subject company in the past twelve months;
•have not received any compensation or other benefits from the subject company or third party related to the specific
recommendations or views in this report;
•do not currently, but may in the future, act as a market maker in the financial instruments of the companies covered in the
report; and
•do not have any conflict of interest in the subject company as of the date of this report.
•Except as noted above, the subject company has not been a client of SCB India during twelve months preceding the date
of distribution of this research report. Neither SCB India nor its Associates (i.e., affiliates/group companies) have received
compensation for products or services other than investment banking, merchant banking or brokerage services from the
subject company in the past twelve months.
•The principal research analyst(s) who prepared this report, members of the analysts' team, and members of their households
are not an officer, director, employee or advisory board member of the companies covered in the report.
•Our Compliance officer / Grievance officer is Ms. Rupal Talati, who can be reached at +91-22-68421451, or
scbincompliance@alliancebernstein.com.
•Disclaimer : Registration granted by SEBI, and certification from NISM, is in no way a guarantee of performance of the
intermediary or provide any assurance of returns to investors.
LEGAL
This publication has been published and distributed in accordance with the Firm's policy for management of conflicts of interest
in investment research, a copy of which is available from Sanford C. Bernstein & Co., LLC, Director of Compliance, 1345 Avenue
of the Americas, New York, N.Y. 10105. Additional disclosures and information regarding Bernstein's business are available on
our website www.bernsteinresearch.com.
This publication is not directed to, or intended for distribution to or use by, any person or entity who is a citizen or resident of, or
located in any locality, state, country or other jurisdiction where such distribution, publication, availability or use would be contrary
to law or regulation or which would subject any of the entities referenced herein or any of their subsidiaries or affiliates to any
registration or licensing requirement within such jurisdiction. This publication is based upon public sources we believe to be
reliable, but no representation is made by us that the publication is accurate or complete. We do not undertake to advise you of
any change in the reported information or in the opinions herein. This publication was prepared and issued by entity referred to
herein for distribution to eligible counterparties or professional clients. This publication is not an offer to buy or sell any security,
and it does not constitute investment, legal or tax advice. The investments referred to herein may not be suitable for you. Investors
must make their own investment decisions in consultation with their professional advisors in light of their specific circumstances.
The value of investments may fluctuate, and investments that are denominated in foreign currencies may fluctuate in value as a
result of exposure to exchange rate movements. Information about past performance of an investment is not necessarily a guide
to, indicator of, or assurance of, future performance.
This report is directed to and intended only for our clients who are “eligible counterparties”, “professional clients”, “institutional
investors” and/or “professional investors” as defined by the aforementioned regulators, and must not be redistributed to retail
clients as defined by the aforementioned regulators. Retail clients who receive this report should note that the services of the
entities noted herein are not available to them and should not rely on the material herein to make an investment decision. The result
of such act will not hold the entities noted herein liable for any loss thus incurred as the entities noted herein are not registered/
authorised/ licensed to deal with retail clients and will not enter into any contractual agreement/arrangement with retail clients.
This report is provided subject to the terms and conditions of any agreement that the clients may have entered into with the entities
noted herein . All research reports are disseminated on a simultaneous basis to eligible clients through electronic publication to
our client portal. The information is private and confidential and for the use of the clients only.
ARTIFICIAL INTELLIGENCE: GETTING SMARTER 307
BERNSTEIN
This report has been prepared for information purposes only and is based on current public information that we consider reliable,
but the entities noted herein do not warrant or represent (express or implied) as to the sources of information or data contained
herein are accurate, complete, not misleading or as to its fitness for the purpose intended even though the entities noted herein
rely on reputable or trustworthy data providers, it should not be relied upon as such. Opinions expressed are the author(s)’
current opinions as of the date appearing on the material only. The information in this report does not constitute a personal
recommendation, as defined by any of the aforementioned regulators, or take into account the particular investment objectives,
financial situations, or needs of individual investors. The report has not been reviewed by any of the aforementioned regulators
and does not represent any official recommendation from the aforementioned regulators.
The analysis contained herein is based on numerous assumptions. Different assumptions could result in materially different
results. The information in this report does not constitute, or form part of, any offer to sell or issue, or any offer to purchase or
subscribe for shares, or to induce engage in any other investment activity. The value of any securities or financial instruments
mentioned in this report can fall as well as rise subject to market conditions. Past performance is not necessarily indicative of
future results. Estimates of future performance mentioned by the research analyst in this report are based on assumptions that
may not be realized due to unforeseen factors like market volatility/fluctuation. In relation to securities or financial instruments
denominated in a foreign currency other than the clients’ home currency, movements in exchange rates will have an effect on
the value, either favorable or unfavorable. Before acting on any recommendations in this report, recipients should consider the
appropriateness of investing in the subject securities or financial instruments mentioned in this report and, if necessary, seek for
independent professional advice.
The securities described herein may not be eligible for sale in all jurisdictions or to certain categories of investors where that
permission profile is not consistent with the licenses held by the entities noted herein. This document is for distribution only as may
be permitted by law. It is not directed to, or intended for distribution to or use by, any person or entity who is a citizen or resident of or
located in any locality, state, country or other jurisdiction where such distribution, publication, availability or use would be contrary
to law or regulation or would subject the entities noted herein to any regulation or licensing requirement within such jurisdiction.
Source: Bloomberg Index Services Limited. BLOOMBERG® is a trademark and service mark of Bloomberg Finance L.P. and its
affiliates (collectively “Bloomberg”). Bloomberg or Bloomberg’s licensors own all proprietary rights in the Bloomberg Indices.
Neither Bloomberg nor Bloomberg’s licensors approves or endorses this material, or guarantees the accuracy or completeness of
any information herein, or makes any warranty, express or implied, as to the results to be obtained therefrom and, to the maximum
extent allowed by law, neither shall have any liability or responsibility for injury or damages arising in connection therewith.
No part of this material may be reproduced, distributed or transmitted or otherwise made available without prior consent of the
entities noted herein. Copyright Bernstein Autonomous LLP, Sanford C. Bernstein & Co., LLC and Sanford C. Bernstein (Hong
Kong) Limited 盛博香港有限公司. All rights reserved. The trademarks and service marks contained herein are the property of
their respective owners. Any unauthorized use or disclosure is strictly prohibited. The entities noted herein may pursue legal action
if the unauthorized use results in any defamation and/or reputational risk to the entities noted herein and research published under
the Bernstein and Autonomous brands.
308 ARTIFICIAL INTELLIGENCE: GETTING SMARTER
BERNSTEIN GLOBAL SALES OFFICES
AMSTERDAM
WTC Schiphol Airport, A-Tower
+31 20 201 4982
BOSTON
53 State Street
+1 617 788 3705
CHICAGO
227 West Monroe Street
+1 312 696 7800
DUBLIN
4 Earlsfort Terrace
+353 1 2463100
FRANKFURT
Bockenheimer Landstrasse 51
+49 69 5050 77 181
HONG KONG
One Island East, Taikoo Place
+852 2918 5762
LONDON
60 London Wall
+44 207 170 5000
LOS ANGELES
1999 Avenue of the Stars
+1 310 407 0027
MILAN
Via Monte di Pietà 21
+39 02 30304 400
MUMBAI
Maker Maxity, BKC
+91 22 6842 1401
NEW YORK
1345 Avenue of the Americas
+1 212 969 2204
SINGAPORE
One Raffles Quay, South Tower
+65 6230 4600
STOCKHOLM
Hamngatan 11
+46 8 535 274 80
TORONTO
161 Bay Street
+1 416 572 2466
ZURICH
Talstrasse 83
+41 44 227 7910