The Environmental Impacts of Machine Learning Training Keep Rising Evidencing Rebound Effect PDF Free Download
1 / 40/40
100%
HAL Id: hal-04839926
https://hal.science/hal-04839926v5
Preprint submitted on 29 Oct 2025
HAL is a multi-disciplinary open access
archive for the deposit and dissemination of sci-
entic research documents, whether they are pub-
lished or not. The documents may come from
teaching and research institutions in France or
abroad, or from public or private research centers.
L’archive ouverte pluridisciplinaire HAL, est
destinée au dépôt et à la diusion de documents
scientiques de niveau recherche, publiés ou non,
émanant des établissements d’enseignement et de
recherche français ou étrangers, des laboratoires
publics ou privés.
Distributed under a Creative Commons Attribution - NonCommercial - ShareAlike 4.0 International
License
The Environmental Impacts of Machine Learning
Training Keep Rising Evidencing Rebound Eect
Clément Morand, Anne-Laure Ligozat, Aurélie Névéol
To cite this version:
Clément Morand, Anne-Laure Ligozat, Aurélie Névéol. The Environmental Impacts of Machine Learn-
ing Training Keep Rising Evidencing Rebound Eect. 2025. �hal-04839926v5�
Highlights
The Environmental Impacts of Machine Learning Training Keep Rising
Evidencing Rebound Effect
Clément Morand, Aurélie Névéol, Anne-Laure Ligozat
• Our new dataset documents the environmental impact of manufacturing graphic
cards
• Rebound effect is prevalent in Machine Learning model training
• Carbon optimization strategies fail to curb ML rising environmental impacts
The Environmental Impacts of Machine Learning Training
Keep Rising
Evidencing Rebound Effect
Clément Moranda,∗
, Aurélie Névéola, Anne-Laure Ligozata,b
aUniversité Paris-Saclay, CNRS, Laboratoire Interdisciplinaire des Sciences du Numérique, Campus
Universitaire bât.507 - Rue du Belvédère, 91405, Orsay, France
bensIIE, 1 place de la Résistance, 91000, Évry-Courcouronnes, France
Abstract
Recent Machine Learning (ML) approaches have shown increased performance on
benchmarks but at the cost of escalating computational demands. Hardware, algo-
rithmic and carbon optimizations have been proposed to curb energy consumption and
environmental impacts. Can these strategies lead to sustainable ML model training?
Here, we estimate the environmental impacts associated with training notable AI sys-
tems over the last decade, including Large Language Models, with a focus on the life
cycle of graphics cards.
Our analysis reveals two critical trends: First, the impacts of graphics cards pro-
duction have increased steadily over this period; Second, energy consumption and en-
vironmental impacts associated with training ML models have increased exponentially,
even when considering reduction strategies such as location shifting to places with less
carbon intensive electricity mixes. Optimization strategies do not mitigate the impacts
induced by model training, evidencing rebound effect. We show that the impacts of
hardware must be considered over the entire life cycle rather than the sole use phase in
order to avoid impact shifting. Our study demonstrates that increasing efficiency alone
cannot ensure sustainability in ML. Mitigating the environmental impact of AI also
requires reducing AI activities and questioning the scale and frequency of resource-
∗Corresponding author
Email addresses: clement.morand@lisn.upsaclay.fr (Clément Morand),
aurelie.neveol@lisn.fr (Aurélie Névéol), anne-laure.ligozat@lisn.upsaclay.fr
(Anne-Laure Ligozat)
intensive training.
Keywords: Green AI, energy consumption, rebound effect, life cycle assessment,
carbon footprint reduction strategies, environmental impact, ML training
1. Introduction
The increasing ubiquity of Large Language Models (LLMs) is fueling debates in
the Natural Language Processing (NLP) community who is questioning the risks and
potential associated with global LLM use [5, 54, 26, 60].
NLP evaluation uses community developed tools including benchmarks and stan-
dardized performance metrics that allow for easy and reproducible comparison between
models. In recent years, these tools were leveraged by leaderboards to provide shared
results. However, this has resulted on an over emphasis on task related performance at
the expense of user oriented metrics [19]. Recent work includes guidelines to advocate
for real-world impact assessment in NLP applications [50]. The assessment of com-
pute requirements, and associated environmental impacts, should therefore be a major
component of NLP evaluation.
Research is underway to understand and develop assessment methods for the im-
pacts of carbon emissions [57], water usage [33] and metallic resource depletion [42,
6]. However these studies are limited to individual models cases and, to our knowl-
edge, no comprehensive sector-wide analysis has been performed and we lack detailed
insights into the sector’s overall evolution. Machine Learning (ML), notably including
Deep Learning (DL), impacts come from a variety of sources including: model de-
velopment (architecture search, parameter optimization...), training and use [34, 63].
Access to data for model development and use phases remains scarce so this study will
focus on trends in model training impacts, which are significant: Model training rep-
resent between 20 and 35% of the impact due to data centers for the Stable Diffusion
service [6].
Strategies have been developed to optimize the energy consumption of models
training thereby mitigating the associated environmental impacts. Hardware optimiza-
tions rely on regular hardware updates to benefit from more energy efficient recent
2
hardware. Algorithmic optimizations adjust model architecture to offer the same task
performance in smaller models. Carbon optimizations displaces computation towards
less carbon-intensive mixes performing compute (e.g., leveraging renewable energy).
Major Artificial Intelligence (AI) companies have claimed that these strategies would
mitigate [64] or reduce the carbon footprint of ML [48]. However, two important
points are being overlooked. First, frequent hardware upgrades reduce the carbon foot-
print from the use phase at the expense of the production and end-of-life phases of the
life cycle. They also increase metallic resource depletion, limiting the overall bene-
fit. This phenomenon is called impact shifting. Second, optimization may result in
smaller-than-expected reductions, or even increase the overall environmental impacts
of the sector. This phenomenon, called rebound effect, has been shown to be prevalent
in Information and Communication Technology (ICT) [24, 8].
In this study, we consider impact shifting and rebound effects to question "does effi-
ciency lead to green ML model training?" Specifically, we investigate how the impacts
of individual graphics cards have evolved in the past 12 years and how the environ-
mental impacts associated with training models have evolved over this period. We
scrutinize the characteristics and environmental impacts of the production of NVIDIA
workstation graphics cards over (2013-February 2025). We use this information to per-
form environmental assessments of ML models training, and in particular NLP models.
Finally, we question the overall effectiveness of the impact mitigation strategies. In
summary, our work makes the following primary contributions:1
• A dataset documenting the environmental impact of manufacturing graphic cards
released between 2013 and 2025;
• Evidence that the impact of training ML, notably including language models, has
increased in spite of improved hardware and algorithmic efficiency, highlighting
the prevalence of rebound effect;
• Evidence that carbon optimization strategies cannot curb the growth in the en-
1The Dataset on graphics cards impacts is available at Morand [41]. The code used in the experiments,
and all necessary information for reproducibility, is available at Morand [40]
3
vironmental impacts of training ML models. This especially holds for language
models.
Paper Outline.. We start by presenting current methods for measuring the impacts of
specific algorithms (Section 2) and present our method for comprehensively accounting
for hardware and model training over time (Section 3). We show that the impacts
of graphics cards manufacturing (Section 4) and model training (Section 5) are ever-
increasing. We argue for embracing the complexity of impact measurement (Section 6)
and conclude that reducing the environmental impact of AI can only be achieved by
reducing AI activities as well as increasing efficiency (Section 7).
2. How Can the Environmental Impact of AI be Measured?
After the high level of carbon emissions associated with training Natural Language
processing models was reported [57], researchers stated the need for a "Green AI" [53],
soon structured as an entire research field [59, 62].
Tools to Measure the Impact of an Algorithm. Methods and tools to assess the energy
consumption of AI activities have been developed, based on energy consumption mea-
sures (e.g., CarbonTracker [2]) or estimations based on hardware characteristics (e.g.,
Green Algorithms [31]). Energy consumption assessments are complemented with data
center energy efficiency indicators (typically, the Power Usage Efficiency (PUE) [3])
and information on the carbon intensity of the local electricity mix to convert hardware
energy consumption into carbon footprint assessments. Several reviews compare exist-
ing tools to assess the energy consumption and associated carbon footprint of training
language models and other AI activities [9, 29, 4]. Additional efforts address the water
usage associated with hardware energy consumption by using the Water Usage Effi-
ciency (WUE) of data centers [33].
In addition to hardware energy consumption during model training, ML related
impacts are incurred by model development (including dataset curation or architecture
search) and model use, with environmental impacts coming from each life cycle phase
of hardware used [34]. Consequently, a method was developed to account for the
4
production phase in the life cycle of hardware [37] and led to MLCA, a tool that assesses
environmental impacts associated with ML training or inferences, including carbon
footprint, energy demand and metallic resource depletion over the production and use
phase of the hardware [42]. Morrison et al. [43] build on Luccioni et al. [37] and
Li et al. [33] to assess carbon footprint and water consumption of models training,
including development and simulated inferences. However, environmental assessments
that scale beyond individual models are still needed.
Measuring at Scale Remains Complex. At the global level, technical reports from in-
dustry and academic labs suggest that the environmental impacts associated with the AI
sector have been rising [23, 39, 56]. de Vries [15] discussed the potential growth of the
ML sector with the recent surge in demand for freely accessible LLMs. Prompting the
International Energy Agency (IEA) [28] to state: "A [...] source of higher electricity
consumption is coming from energy-intensive data centres, artificial intelligence (AI)
and cryptocurrencies, which could double by 2026." Masanet et al. [38] has shown that
despite an exponential growth in computation demand between 2012 and 2018, the
total energy consumption of data centers had only increased by around 5% thanks to
scale gains in large scale data centers, that are much more energy efficient (measured
by their PUE). However, the relative decoupling between compute demand and data
center energy consumption may have evolved since 2018.
These individual reports and studies do not provide consolidated information re-
garding the global trend for environmental impacts of the AI sector. Furthermore, we
need to identify the precise sources driving this growth: Is increased impact caused by
more impactful model training? Is it the result of more frequent re-training? Or is it
mainly caused by an increase in the number of inferences?
Diverging Trends Need to be Analyzed. While optimization techniques have been used
to build less carbon intensive models [63, 48], energy demand per inference tends to
increase over time [16]. The energy efficiency of graphics cards has increased expo-
nentially leading to the claim that frequent hardware updates will shrink the carbon
footprint of training models [48]. In spite of algorithmic optimization, Sevilla et al.
[55] and Thompson et al. [58] report an exponential increase in computation require-
5
Figure 1: Hardware modeling in MLCA.
ments to train models. This could be a producer rebound effect, where the efficiency
increases fuel the creation of even larger models thereby canceling the potential im-
pact reduction [13]. Luccioni et al. [36] also warn against risks of rebound effect in
the AI sector. Optimizations are often absorbed by the growth of the Information and
Communication Technologies (ICT) sector [8, 24]. Is this also true for ML training?
Overall, previous work has led to the development of methods and tools for assess-
ing the impact of specific ML algorithms. Reports suggests that the overall environ-
mental impact of the AI sector is rising. However, mechanisms behind such an increase
remain undocumented. Focusing on model training, growing compute demand while
deploying different optimizations is consistent with rebound effect. Herein, we present
a comprehensive study of the impact of ML hardware and ML model training over a
decade, including rebound effect and hardware production.
6
3. Environmental Impact Assessment
3.1. General Methodology for Environmental Impact Assessment
We perform all environmental assessments using attributional Life Cycle Assess-
ment (LCA), as implemented in MLCA [42]. The tool aims to cover a representa-
tive set of environmental impact categories. For ICT hardware and services, carbon
footprint, metallic resource depletion, water consumption and toxicity to human and
non-human life are all important sources of impact. MLCA provides carbon footprint
and metallic resource depletion assessments, based on open-source LCA information
from Boavizta [7].2MLCA is based on a bottom-up modeling of hardware to estimate
impacts over the production and use phases of hardware. Bottom-up modeling means
that hardware is modeled as the sum of its parts. Figure 1 represents the modeling
of hardware in MLCA. Graphics cards are modeled based on the size of the GPU (in
hatched purple in the Figure), the size of on-board memory (in cross-hatched blue) and
constant impact for other card components (in green). Energy consumption is based on
an estimate using the Termal Design Power (TDP) of the card3and training duration.
We assess the energy consumption, carbon footprint and metallic resource deple-
tion associated with hardware production and with training ML systems. The car-
bon footprint is assessed according to Global Warming Potential (GWP),4measured
in kgCO2eq for the emissions of greenhouse gases such as carbon dioxide, methane
and nitrous oxide [21]. Metallic resource depletion is assessed through Abiotic De-
pletion Potential (ADPe) which is obtained by assessing the quantity of metal (e.g.,
copper, gold and rare earths) used to produce hardware. An aggregated indicator is
created based on the amount of each metal and metal rarity compared to Antimony, the
standard reference. The final indicator is expressed in kilograms antimony equivalent
(kgSb eq) [45].
2Recent open quality information on toxicity and water consumption of ICT equipment and data center
facilities is missing [33, 32]. End-of-life information is also lacking [20].
3The TDP corresponds to the heat the chip should be able to dissipate to function at maximum load; it is
used to approximate the maximum power consumption of the chip.
4We use the IPCC standard GWP100, corresponding to impact at a time horizon of 100 years.
7
Figure 2: Scope of environmental assessments of ML training in this study. The barrel represents fossil fuels,
the droplet represents water, the pickaxe represents metals, and the plug represents electricity. The top row
pictograms represent, from left to right, emissions in soil, water and air.
Figure 2 presents the scope of our study within the life cycle of hardware used for
training ML models. We consider energy consumption, carbon footprint and metallic
resource depletion over server production and usage, and data center cooling usage.
We now explore the production of graphics cards, before focusing on model training.
3.2. Graphics Cards Production Impacts
LCA of ICT equipment have shown the importance of Integrated Circuits (IC) in
the environmental impacts of ICT equipment [11]. ICs come in two forms in graphics
cards: GPU (logic type ICs) and memory (memory type ICs). The surface of the
GPU is indicated by the GPU die area. Contributors to the impact of producing ICs
include the surface of the IC (i.e., the die area of the GPU and the surface of memory
type IC for the memory chips), as well as how finely the circuits are printed on the
semiconducting material. Thus, if the die area increases we can expect an increase in
the environmental impacts of the device. Furthermore, Pirson et al. [49] have shown
that, with finer technological nodes, the environmental impacts per produced cm2of
8
die increase. Thus, as the quantity of memory increases (probable increase in memory
type IC surface), the GPU die size increases and the technological node gets finer (latest
GPUs processed at 5nm), we can estimate that the environmental impacts of graphics
cards production increases.
We study the evolution of the characteristics of graphics cards over time to test
these hypotheses. We focus on the leading provider for workstation graphics cards,
NVIDIA. We curate a dataset of the 174 workstation graphics cards models released
between 2013 and 2024 based on the TechPowerUp GPU database,5a Wikipedia page
listing NVIDIA graphics cards6and NVIDIA’s published datasheets to settle source
disagreements. The final dataset is available at Morand [41], and additional information
on data pre-processing is available in Appendix A.
Additional frequently used cards are also included in our study such as Google’s
TPU. For these models, data comes from a dataset on ML training hardware [17],
manufacturers websites, press releases and public benchmarks.
3.3. Assessing ML Model Training Impacts
Studies on ML models training have been conducted using the Epoch AI Notable
systems database [18] retrieved on February 28, 2025. Epoch AI is the most compre-
hensive database on ML systems to our knowledge. It gathers extensive information on
a large variety of notable7ML systems. Details on used software and details necessary
for exact replication of the presented experiments are available in Appendix B.
Training Duration. We estimated GPU-hours required for training models using two
methods, based on information from the Epoch AI database presented in Table 1. The
database mostly comprise data gathered from papers/technical reports presenting the
5TechPowerUp https://www.techpowerup.com/gpu-specs/ (last accessed 3/17/25) granted
authorization to use and share graphic cards data as part of our research project.
6https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_uni
t- accessed 3/6/25; content available as CC BY-SA 4.0.
7Notable systems are “models that have advanced the state of the art, had a large influence in the field’s
history, or had a large impact within the world”, per https://epochai.org/data/notable-ai-m
odels-documentation.
9
GPU-h1GPU-h2
models training
duration
card
quantity
both + card
model
training
FLOP
card
model
both
Number 897 199 168 131 122 451 266 230
Coverage (%) 100 22 19 15 14 50 30 26
Confidence
Confident 234 98 96 78 75 166 138 122
Likely 103 41 34 21 17 77 51 47
Speculative 69 21 10 9 7 51 20 16
Unknown 491 39 27 23 23 157 57 45
Table 1: Description of the Epoch AI database. The values show the number of entries for each information
type. Confidence scores are reported by database authors based on the quality of the Training compute,
Parameters, and Training data set size fields.
models and used to yield estimates with varying levels of confidence, but some param-
eters cannot be reliably estimated. GPU-hours we estimate are performed on models
majorly assessed as reporting high-quality information (74% of models).
For models specifying duration and graphics cards quantity, they can be multiplied
to obtain an estimate referred to as GPU-h1:
GPU-h1=training duration ×#cards
This estimation is the most reliable as it uses information retrieved from the papers
presenting the models. However, it only covers 15% of Epoch AI models (see Table 1).
To increase coverage, we also use models where graphics cards models and training
compute requirements (FLOP) are documented. For these models, we can estimate the
training duration by dividing the number of FLOP necessary for model training by the
peak compute power of the card used (in FLOPs, i.e., the number of FLOP the card can
perform per second). This value is referred to as GPU-h2base:
GPU-h2base =training FLOP
peak card compute power
GPU-h2base should lead to an under-estimation of the number of GPU-hours com-
10
pared to GPU-h1, as the hardware does not always operate at peak performance, for
instance due to synchronisation phases between the different nodes when compute can-
not be realised.
After checking for consistency between the two estimation methods and exclud-
ing benign anomalies, we develop a linear model to correct the underestimation from
GPU-h2base compared to GPU-h1. Appendix B.2 details procedures for ensuring
consistency and details the linear model. The obtained model corresponds to using
a quasi-constant performance ratio of ≃27%. This leads to our second estimation
method referred to as GPU-h2:
GPU-h2=training FLOP
peak card compute power ×0.27
The final estimation for GPU hours is as follows: we use GPU-h1for the 131
models (15%) where it is available, and use GPU-h2for another 103 models (11%),
covering 26% of Epoch AI models.
Server Characteristics. We made hypotheses on memory provisioning and number
of cards per server, and hardware lifespan and utilization using values consistent with
computing facility set-up at our university, documentation from graphic card and server
manufacturers and the literature [47, 63]. Exact details for replication on server char-
acteristics is available in Appendix B.3. This information serves to allocate server pro-
duction impact to a specific model training. Allocated production impacts are called
embodied impacts.
We also modeled the increase in data center efficiency estimating average PUE
from 2010 to 2018 using linear interpolation and a PUE of 1.2 from 2018 onward
based on Masanet et al. [38].
Hardware Energy Consumption.. We use a value of 100% usage both for GPUs and
CPUs during training as it was shown to yield accurate estimations [29].
Environmental Impact of Energy Usage. To evaluate the impacts of energy consump-
tion associated with training each model, we consider that models have been trained
11
using the energy mix of the countries of the ML model producers documented by Boav-
izta.8If multiple countries participate in model creation, the energy mixes of each im-
plicated country are used to create a value interval, and the mix of the country indicated
in first position is used as the reference value.
We illustrate our methodology for the GPT-4 model, released in March 2023. Train-
ing duration is estimated at 57,000,000 GPU-hours on NVIDIA A100 SXM4 40 GB
cards based on information in the EpochAI dataset and using GPU-h1. Training servers
are modeled as containing 4 cards, 2 CPUs and 512GB memory (consistently with
nvidia workstation cards), and facility PUE is estimated at 1.2 (because the model was
trained after 2018). Energy consumption impact is computed with average USA energy
mix (GPT4 authors are affiliated in the US). Table 2 presents impacts associated with
GPT-4 training estimated using MLCA. For this model, embodied impacts represent
20% of the total carbon footprint and close to 100% of metallic resource depletion.
Energy GWP ADPe
(GWh) (ktCO2eq) (kgSb eq)
Embodied - 3.3 300
Usage 27.4 10.2 2.7
Infra 5.4 2.0 0.5
Total 32.8 15 300
Table 2: Estimated production impacts for training the GPT-4 model: impacts related to the energy con-
sumption of servers (row Usage) and infrastructure (row Infra) are presented as well as embodied impacts
(row Embodied) .
Simulating Carbon Optimization Strategies. The study considers different scenarios
for reducing carbon intensity in energy mixes over time, including shifting compute
locations and decarbonizing electricity for data centers. Each scenario involves a con-
8Details on the sources used for each countries can be found at: https://github.com/blubrom
/MLCA/blob/main/boaviztapi/data/electricity/electricity_impact_factors
.csv.
12
tinuous reduction in carbon intensity, starting from 2019, with reductions of up to 25%
per year onward. The carbon intensity of the mix used for training a model is thus mul-
tiplied by (1 −ratio)nwhere n stand for the number of years since 2019 at release
date of the system.
4. Evolution of Hardware Production Impact
We analyze the characteristics and environmental impacts of graphics cards over
time, focusing on NVIDIA cards and cards used for model training.
4.1. Evolution of Graphics Cards Characteristics
Figure 3 shows the evolution of the characteristics of NVIDIA workstation graphics
cards from 2013 to 2024. Environmental impacts of graphics card production majorly
come from the GPU and memory (respectively in hatched purple and in cross-hatched
blue in Figure 1). The larger the ICs and the smaller the technological node, the greater
the impact. Information on the technological node and on die area is available for
GPUs while only raw size is available for memory. Technological node size has de-
creased over time and that average die area has increased linearly. Memory size (in
GB) has grown exponentially (around 30% Compound Annual Growth Rate). Expo-
nential growth in memory size does not lead to a proportional increase in memory IC
area as ICs have been miniaturized at a pace following Moore’s Law.
Compute efficiency is computed by dividing the peak performance of cards (max-
imum compute power in Single, Double, Half and Tensor precision) by their TDP
and corresponds to the number of operations cards can perform per second per Watt.
Compute efficiency has increased exponentially over time (see complementary Fig-
ure C.12).
Figure 4 shows the evolution of the energy consumption of NVIDIA workstation
graphics cards from 2013 to 2025, in terms of TDP. Even if the compute efficiency
of the cards has increased exponentially, the total energy consumption of a card has
slightly increased over time. This observation is consistent with rebound effect, where
the energy efficiency improvements on the cards have allowed to increase the number
13
0
200
400
600
800
2013
2014
2015
2016
2017
2018
2019
2020
2021
2022
2023
2024
2025
Release Date
Die area (mm^2)
Memory
size (GB)
50
100
Technological
node (nm)
10
20
30
40
Figure 3: Evolution of the characteristics of NVIDIA workstation graphics cards from 2013 to 2025. Dot
size represents memory size and color represents GPU technological node.
of operations performed on a card at a fixed energy consumption. The production
impacts of NVIDIA workstation graphics cards in terms of GWP and ADPe increase
over time (see complementary Figure C.13).
4.2. Evolution of the Hardware Requirements
Figure 5 shows the production impacts of graphics cards used to train ML systems
in Epoch AI. The models of graphics cards used to train ML systems have evolved
similarly to all workstation cards, confirming that the production impacts of graphics
cards used to train ML models have increased over time. Figure 6 shows that hardware
quantity has increased exponentially, even if some models are still trained using only
a few cards. More graphic cards with higher production impacts are used, suggest-
ing growing environmental impacts. Both production impact and energy consumption
during usage need to be addressed.
Computing facilities are increasingly energy efficient and can perform much more
computation in a fixed duration using less energy. This efficiency is partly obtained
by regular hardware updates in data centers, thus incurring additional equipment pro-
duction and end-of-life environmental costs. Frequently changing hardware could be
14
0
200
400
600
2013
2014
2015
2016
2017
2018
2019
2020
2021
2022
2023
2024
2025
Release Date
TDP (W)
Figure 4: Evolution of the energy consumption of NVIDIA workstation graphics cards (2013-2025).
100
200
300
400
500
2011
2012
2013
2014
2015
2016
2017
2018
2019
2020
2021
2022
2023
2024
2025
Publication date
Used graphics card individual
production GWP (kgC02eq)
(a) GWP
0.00
0.01
0.02
0.03
2011
2012
2013
2014
2015
2016
2017
2018
2019
2020
2021
2022
2023
2024
2025
Publication date
Used graphics card individual
production ADP (kgSbeq)
(b) ADPe
Figure 5: Evolution of the production impacts of the graphics cards used for training ML systems in the
Epoch AI database. Value intervals correspond to cases of ambiguous card names.
understood as a form of impact shifting. Energy consumption can be reduced (at con-
stant use) during the potentially shorter equipment lifespan. Usage impact is reduced.
However, production impact in terms of carbon footprint and metallic resource de-
pletion is increased by frequent hardware renewal. In addition, new hardware also
has higher production impact due to technological advances required to reduce energy
consumption in the usage phase.
5. Optimization Strategies
5.1. Trends in ML Model Training
We present an analysis of models training that explores the energy consumption
of hardware used to train models as well as the environmental impacts incurred by
hardware throughout life cycle.
15
100
101
102
103
104
105
2013
2014
2015
2016
2017
2018
2019
2020
2021
2022
2023
2024
2025
Publication date
Number of
graphics cards used
Figure 6: Number of graphics cards used for training models in Epoch AI over time.
10−2
100
102
104
2011
2012
2013
2014
2015
2016
2017
2018
2019
2020
2021
2022
2023
2024
2025
Publication date
Estimated energy
consumption (MWh)
Figure 7: Evolution of the energy consumption of training ML models over time. Value intervals account for
ambiguous card names.
Energy Consumption of Model Training.. Figure 7 presents the energy consumption
associated with each model training estimated using MLCA (see Section 3.1). Energy
consumption has, on average, increased exponentially over time, paralleled with an
increase in the number of models released over time. Low energy consumption models
are still produced in recent years and could result from Green AI research.
Environmental Impacts of Model Training.. Figure 8 presents the estimated training
impacts in terms of GWP (Figure 8a) and ADPe (Figure 8b). Both environmental
indicators have increased exponentially between 2012 and 2025.
Table 3 summarizes the distribution of the shares of embodied impacts. The esti-
mations for the share of embodied GWP are mostly estimated around a third the total,
16
10−4
10−2
100
102
104
2011
2012
2013
2014
2015
2016
2017
2018
2019
2020
2021
2022
2023
2024
2025
Publication date
Estimated GWP
(tCO2eq)
(a) GWP
10−4
10−2
100
102
2011
2012
2013
2014
2015
2016
2017
2018
2019
2020
2021
2022
2023
2024
2025
Publication date
Estimated ADPe
(kgSbeq)
(b) ADPe
Figure 8: Environmental impacts in GWP (a) and ADPe (b) of training ML models over time. Value intervals
account for ambiguous card names and model producers from multiple countries.
Min Q1 Q2 Mean Q3 Max
ADPe 89 100 100 99 100 100
GWP 12 19 23 23 25 58
Table 3: Share (in percent) of embodied impacts on the total impacts associated with training models.
which is consistent with previous assessments in the literature [27, 64, 37]. The vari-
ation mostly comes from variations in the carbon intensity of the electricity mixes of
countries in which models are trained. Models trained in countries with a low carbon
intensity for their electricity consumption such as France, will have a larger share of
embodied impacts than if trained in another country such as the USA. ADPe, however,
comes close to exclusively from hardware production. For both indicators, embodied
impacts represent a significant share of total impacts. Thus, solely reducing the en-
ergy impacts from ML models training will not be sufficient to solve the environmental
impacts of AI.
5.2. Are the Trends Specific to a Subfield?
Figure 9 details the evolution of training carbon footprint for models by modality.
Since too few datapoints are available, biology, image generation, games, speech and
specialty domains are grouped into a single "other" category. No clear trend emerges
from this category. Language, vision, and multimodal models (to a lesser extent) all
exhibit an exponential growth trend. Thus, growth in training impact is not the sole
product of large language models or any other subsector.
17
Multimodal
Vision
Biology
Language
2012
2014
2016
2018
2020
2022
2024
2012
2014
2016
2018
2020
2022
2024
10−2
100
102
104
10−2
100
102
104
Publication date
Estimated GWP (tCO2eq)
Figure 9: Evolution of the carbon footprint of training ML models over time, by modality. Value intervals
account for ambiguous card names.
5.3. Hardware and Algorithmic Optimization Overall Effect
The evolution of graphic cards (Figures 3-6) suggests that the environmental im-
pacts of hardware production have increased over time. The compute efficiency of the
newer cards allows hardware optimization for more recent models. In the meantime,
more efficient model architectures are continuously developed and allow to attain a
set model performance with less resources [48]. We thus hypothesize that more re-
cent models have benefited from algorithmic optimization. Model impact has however
essentially grown while hardware and algorithmic optimization have been rolled-out
(Figures 7-8). The overall effect of hardware and algorithmic optimization strategies is
consistent with producer rebound effect, where the efficiency increases fuel the creation
of larger models and cancels the potential impact reduction.
5.4. Carbon optimization can only yield short-lived benefits
We compare the impact of models trained with actual electricity mixes with a sim-
ulated annual reduction of 25% of the carbon intensity of electricity mixes (see Sec-
tion 3.3). Figure 10 shows that the carbon footprint of models released from 2019 is
increasing regardless of carbon optimizations. Trends are very similar with or without
carbon intensity reduction and regression coefficients are significantly positive. This
suggests that carbon optimization is not a sufficient reduction strategy.
18
10−4
10−2
100
102
104
2019
2020
2021
2022
2023
2024
2025
Publication date
Estimated GWP (tCO2eq)
Rate of
annual
reduction
of the
carbon
intensity
0
0.25
Figure 10: Estimated carbon footprint of training models released after 2019, with or without reduction of
the carbon intensity of the used electricity.
The carbon intensity of electricity mixes is bound to 15-20gCO2eq/kWh, based
on the current world lower intensity mixes. Thus, even if carbon intensity reductions
were quick enough to counterbalance the increase in energy consumption, the benefits
would be short lived with energy consumption trending up. Consuming lower impact-
ing electricity is a key Green AI practice as it reduces hardware usage impact for a fixed
compute demand. Our results suggest that this strategy alone cannot curb the growth of
impacts of model training. Furthermore, changing usage electricity consumption does
not mitigate hardware production impacts, which are also increasing. Worse, The geo-
graphic changes intended to reduce carbon intensity could incentivize shorter lifespan
for data center facilities [61], leading to higher impact of these facilities.
6. Discussion
To our knowledge, this is the first study proving the growth in ML training impacts
and explaining drivers of this growth. We show that carbon optimization strategies fail
to curb training impacts.
19
0
200
400
600
BLOOM−176B
GLaM
GPT−3
Llama 2−70B
Llama 2−7B
LLaMA−65B
NLLB
OPT−175B
Model
Energy related GWP
(tCO2eq)
status
Estimated
Expected
Figure 11: Estimated (herein) vs. expected (in the literature) carbon footprint from energy consumption
regarding several models training.
6.1. Study Scope Supports Reliable Trends
Production impact assessment in MLCA has limitations, including not accounting
for technological node and assuming fixed memory density, but estimates are consis-
tent with recent research [52]. The use of the PUE may lead to underestimation of
infrastructure consumption. Adjusting for dynamic ratio as proposed by Morand et al.
[42] would not change the overall findings since it would only mean using a higher
multiplicative factor for all models. Epoch AI, although extensive, does not include
all models and variants, but omissions are unlikely to alter the observed exponential
increase in training impacts.
6.2. Sparse Information Lead to Estimations
This study estimates missing parameters for models using inferences and hypothe-
ses, which may lead to under or over-estimation of certain parameters (e.g., training
time based on number of FLOP). The validity of these estimates is checked by compar-
ing our estimations to published information, with results (Figure 11) showing general
consistency but some significant differences. For example, the carbon footprint of the
BLOOM-176B model is overestimated due to assumptions about its training location
and electricity mix.
20
6.3. This Study Addresses Model Training
Morrison et al. [43] and Luccioni et al. [37] have shown that model development
includes training multiple smaller models to converge on final architecture. These mul-
tiple training runs, that we do not account for, can account for at least half of the total
footprint [43]. Future work can also address the impact of inferences in the sector, as
they weight more heavily over the life cycle of models. It will however require access
to user information.
6.4. Rebound Effect is Prevalent in AI
Our findings evidence rebound effect consistently with previous observations: Wu
et al. [63] identify a rebound effect at Facebook. Patterson et al. [48] show that a
constant share of Google’s energy consumption is attributable to AI when the total
consumption increased. Efforts from de Vries [15] and Desislavov et al. [16] suggest
that our findings should also apply to inferences.
6.5. Greener Energy Cannot Void Carbon Impact
Figure 10 shows that reducing carbon intensity of electricity used seems insufficient
to curb the exponential growth of the carbon footprint of training ML models. Further-
more, the electricity consumption of data centers destabilizes local electricity grids
[46], potentially causing the prolongation of fossil fuel power plants [10, 1]. On-site
renewable energy production to match data centers consumption can also be a source
of electricity grid instability [22]. Matching the carbon footprint through carbon off-
setting also has limited potential [25, 35].
6.6. Impacts Go Beyond Carbon Footprint
This study examined the carbon footprint and metallic resource depletion of train-
ing AI models, finding that both metrics have increased over time due impacts in dif-
ferent phases in the life cycle of hardware. Importantly, AI environmental impacts
extend beyond these metrics, including water usage [44], ecosystem destruction [12],
and pollution from hardware mining and disposal. Additionally, AI poses significant
social consequences and ethical challenges [5, 30], highlighting the need for a more
21
comprehensive assessment that incorporates qualitative analyses and a broader range
of impacts.
7. Conclusion
We have shown that ML and in particular language model training had increasing
environmental impacts between 2013 and 2025. While hardware upgrades, algorithmic
optimizations, and carbon-aware training have been widely adopted, we find that these
strategies fail to mitigate the growing environmental impacts of ML and NLP systems,
highlighting the prevalence of rebound effect in the sector.
This trend poses a critical challenge for the NLP community: Growth of the field’s
environmental impacts conflicts with climate change urgency. While AI could support
sustainability [51], its current trajectory risks undermining these benefits by driving
unsustainable resource consumption [14].
Impact reduction must be combined with a broader consideration of the role of AI
in a sustainable society.
Acknowledgments
This work has received funding from the French "Agence Nationale pour la Recherche"
under grant agreement InExtenso - ANR-23-IAS1-0004. Clément Morand was sup-
ported by a doctoral grant from ENS Rennes. The authors thank TechPowerUp for
allowing us to us their data in this study. Clément Morand would like to thank Aina
Rasoldier for his help in gathering the data from TechPowerUp, Loïc Lannelongue for
discussions during the experiments framing, Adrien Berthelot for insightfull discussion
and bibliographical advice, and Aurélie Bugeau for her comments on earlier versions
of this manuscript.
CRediT author statement
Clément Morand: Conceptualization, Data curation, Methodology, Software, Writ-
ing - Original Draft Aurélie Névéol: Conceptualization, Writing- Review & Editing,
22
Supervison Anne-Laure Ligozat: Conceptualization, Writing- Review & Editing, Su-
pervision, Methodology.
Appendix A. Additional Information on Graphics Cards Dataset Curation
NVIDIA Workstation Graphics Cards. In order to gather information on graphics cards
characteristics, we curated a data set with information on 174 NVIDIA workstation
graphics cards models released between 2013 and 2024 included. The main infor-
mation gathered for each model includes: Release date, die area, technological node,
memory type, memory size, Thermal Design Power (TDP) and compute power (Sin-
gle, Double, Tensor and Half floating precision). First part of the data set was retrieved
from the TechPowerUp GPU database.9Another data set of 84 graphics cards models
based on a Wikipedia page listing NVIDIA graphics cards10 was retrieved. We merged
the two data sets to cross-validate the specifications of the cards. This validated in-
formation on 83 out of the 173 models (47% of the models) in the TechPowerUp data
set. One model (Tesla P6) was included in our Wikipedia data set and not in our
TechPowerUp data set, and has been added to the final data set. In cases of divergent
information, NVIDIA’s published datasheets are taken as reference. These datasheets
have also been used to validate information on the compute power of the most popular
graphics cards (the P100, V100, A100 and H100 card families). The final dataset will
be released with the paper.
non-NVIDIA-Workstation Graphics Cards.. To be able to assess the environmental im-
pacts of all ML models, we also gather information on other graphics card models used
for training. We relied on different sources: the Epoch AI data set on machine learning
training hardware [17] provided details for NVIDIA non-workstation cards (11 card
models), the Google Cloud Platform documentation and publications by Google on
9The https://www.techpowerup.com/gpu-specs/ website was accessed on December 12,
2023 for cards released from 2013 to 2023 and on March 17, 2025 for cards released in 2024. TechPowerUp
granted us authorization to use and share this data as part of our research project.
10The https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing
_unit website was accessed on March 6, 2025, its content is available under a CC BY-SA 4.0 license.
23
their hardware, and website of the manufacturers, press releases and benchmarks for
the other cards (Cerebras CS-2, Huawei Ascend 910 and AMD Instinct MI250X). Our
final data set is available with the accompanying code.
Appendix B. Details for Exact Replication of Environmental Assessments
Appendix B.1. Processing Ambiguous Card Names
8 ML models are excluded from analysis because various hardware is used through-
out the training process and adequate attribution is not possible. Other models docu-
ment hardware ambiguously (e.g., "A100" could refer to several cards with different
features). Then, the one we expect to be most frequently used, is chosen as the refer-
ence value, and other options are used to compute a value interval.
Appendix B.2. Training Duration Estimates
Ensuring Consistency and Excluding Anomalies:. In order to validate estimates from
GPU-h2base and correct the underestimation, we compare GPU-h1and GPU-h2
base for each of the 119 models where both values are available, taking GPU-h1as
a ground truth. This allows us to obtain an estimate for the ratio of hardware perfor-
mance: ratio =GPU-h2base
GPU-h1. This ratio should in theory be lower than one as a ratio
greater than one would in theory mean that the model has used the GPUs at more than
their peak performance. We define anomalies as models where the obtained ratio is
greater than one or smaller than 10%. Analyzing the obtained ratios revealed anomalies
for 19 models (in 16% of cases). Anomalies especially occur with cases of fine-tuned
models where GPU-h2base includes training the base models while GPU-h1only
accounts for the fine-tuning process. Anomalies also occur when the number of FLOP
in the EpochAI database was estimated based on compute power values different from
the ones in our graphics cards database due to inconsistencies between the Epoch AI
database and manufacturer data. Both types of anomalies should not pose problems as
we use GPU-h1as our final estimation for these models.
24
Linear Model for Correcting Underestimations:. After checking for consistency be-
tween the two estimation methods and excluding benign anomalies, we develop a linear
model to correct the underestimation from GPU-h2base compared to GPU-h1. We
build a linear model to predict GPU-h1using GPU-h2base. This model is computed
on 100 observations excluding the anomalies. To ensure a linear relation between both
variables, we estimate log(GPU-h1)∼log(GPU-h2base)as GPU-h1and GPU-h2
base both are exponentially distributed. The linear regression analysis revealed a sta-
tistically significant model (F(1,98) = 6525, p <2×10−16), with an adjusted R² of 0.98,
meaning that 98 percent of the variance in the observations is explained by our model.
The model equation is log(GPU-h1)=1.3+1.0 log(GPU-h2base)with a standard
error of 0.12 for the intercept and 0.01 for the regression coefficient. This indicates that
an increase of 1 for the log(GPU-h2base)value leads to an average increase of 1.00
units in log(GPU-h1). This positive relationship between log(GPU-h2base)and
log(GPU-h1)was found to be statistically significant (t(98) = 80.78, p <2×10−16),
affirming the predictive power of log(GPU-h2base)on log(GPU-h1). In addition
to the regression analysis, a scatterplot with the fitted regression line was examined
to ensure model assumptions were met. Homoscedasticity was confirmed (studen-
tized Breusch-Pagan test = .08, p = 0.77) and the residuals appeared to be independent
(Durbin-Watson D = 1.81, p = .17).
Appendix B.3. Details on Modeled Server Characteristics
We made the hypothesis that servers equipped with NVIDIA workstations cards
contain 4 graphics cards, 2 CPUs and 512 GB memory. We also made the hypothesis
that servers equipped with NVIDIA non-workstations cards contain 2 graphics cards,
2 CPUs and 192 GB memory. These configurations were chosen based on a review
of configurations in the different generations of NVIDIA DGX servers as well as on
configurations in different computing facilities, including computing facilities available
in the Université Paris-Saclay.
For non-NVIDIA hardware, we searched documentation (e.g., Google Cloud Plat-
form documentation, publications by Google) to obtain information on the number of
chips and processors per server. For instance, for TPUv3, a server with two CPUs
25
manages every four TPU chips. Missing information on the memory quantity in each
server, the value of 448 GB memory is chosen based on the value for the TPU v5p
for the TPU v2, v3 and v4. This value seems consistent with the value chosen for the
NVIDIA servers.
We use values consistent with hyper-scale data centers for the hardware lifespan,
average utilization over its life cycle and infrastructure energy consumption. We use a
lifespan of 3 years for the hardware [47], and, using information from Meta, a hardware
utilization of 50% [63].
Based on the data from Masanet et al. [38], we model an increase in the efficiency
of data centers between 2010 and 2018, accompanied with the shift from traditional
data centers to hyperscale data centers. Assuming a uniform distribution between Tra-
ditional, Cloud and Hyperscale data centers, we use an average PUE of 1.75 for data
centers in 2010. (This average value of 1.75 is close to the average PUE value for Cloud
(non-hyperscale) data centers in 2010) Assuming that all AI workloads take place in
hyperscale data centers starting in 2018, we use an average PUE of 1.2 from 2018 on-
wards. For dates between 2010 and 2018, we use a linear interpolation to estimate the
used average PUE.
Appendix B.4. Software Used in the Analysis
Data manipulation and statistical analyses have been performed using emacs Org
mode 9.1.9, Python 3.8.10 using the pandas version 2.0.3 library and R version 3.6.3
(2020-02-29) with the ggplot2_3.4.3, dplyr_1.1.3, lmtest_0.9-40, stringr_1.5.0 and
zoo_1.8-12 libraries. Platform: x86_64-pc-linux-gnu (64-bit) Running under: Ubuntu
20.04.6 LTS
Consistent with open science practices, details on the sources, processes and method-
ological choices are available at [40].
26
Appendix C. Complementary Figures
100
101
102
103
104
2013
2014
2015
2016
2017
2018
2019
2020
2021
2022
2023
2024
2025
Release Date
compute efficiency
(GFLOPs/W)
Figure C.12: Evolution of the compute efficiency of NVIDIA workstation graphics cards (2013-2025).
0
50
100
150
200
250
2013
2014
2015
2016
2017
2018
2019
2020
2021
2022
2023
2024
2025
Release Date
Production GWP
(kgCO2 eq)
(a) GWP
0.00
0.01
0.02
0.03
2013
2014
2015
2016
2017
2018
2019
2020
2021
2022
2023
2024
2025
Release Date
Production ADP
(kgSb eq)
(b) ADPe
Figure C.13: Evolution of the production impacts of NVIDIA workstation graphics cards.
References
[1] Ambrose, J., 2024. Ireland’s datacentres overtake electricity use of all urban
homes combined. URL: https://www.theguardian.com/world/
article/2024/jul/23/ireland-datacentres-overtake-ele
ctricity-use-of-all-homes-combined-figures-show. the
Guardian, https://www.theguardian.com/world/article/2024
/jul/23/ireland-datacentres-overtake-electricity-use
-of-all-homes-combined-figures-show.
27
[2] Anthony, L.F.W., Kanding, B., Selvan, R., 2020. Carbontracker: Track-
ing and predicting the carbon footprint of training deep learning models.
arXiv:2007.03051.
[3] Avelar, V., Azevedo, D., French, A., Power, E.N., 2012. PUE: a comprehensive
examination of the metric. White paper 49.
[4] Bannour, N., Ghannay, S., Névéol, A., Ligozat, A.L., 2021. Evaluating the carbon
footprint of NLP methods: a survey and analysis of existing tools, in: Moosavi,
N.S., Gurevych, I., Fan, A., Wolf, T., Hou, Y., Marasovi´
c, A., Ravi, S. (Eds.),
Proceedings of the Second Workshop on Simple and Efficient Natural Language
Processing, Association for Computational Linguistics, Virtual. pp. 11–21. URL:
https://aclanthology.org/2021.sustainlp-1.2, doi:10.186
53/v1/2021.sustainlp-1.2.
[5] Bender, E.M., Gebru, T., McMillan-Major, A., Shmitchell, S., 2021. On the dan-
gers of stochastic parrots: Can language models be too big?, in: Proceedings of
the 2021 ACM Conference on Fairness, Accountability, and Transparency, As-
sociation for Computing Machinery, New York, NY, USA. p. 610–623. URL:
https://doi.org/10.1145/3442188.3445922, doi:10.1145/34
42188.3445922.
[6] Berthelot, A., Caron, E., Jay, M., Lefèvre, L., 2024. Estimating the environmental
impact of Generative-AI services using an LCA-based methodology, in: Procedia
CIRP, Turin, Italy. pp. 1–10. URL: https://inria.hal.science/ha
l-04346102.
[7] Boavizta, 2021. Numérique et environnement : Comment évaluer l’empreinte de
la fabrication d’un serveur, au-delà des émissions de gaz à effet de serre˜
? URL:
https://www.boavizta.org/blog/empreinte-de-la-fabrica
tion-d-un-serveur.https://www.boavizta.org/blog/empr
einte-de-la-fabrication-d-un-serveur, Accessed on May 15,
2023.
28
[8] Bol, D., Pirson, T., Dekimpe, R., 2021. Moore’s law and ICT innovation in
the anthropocene, in: 2021 Design, Automation & Test in Europe Conference &
Exhibition (DATE), pp. 19–24. doi:10.23919/DATE51398.2021.947411
0.
[9] Bouza, L., Bugeau, A., Lannelongue, L., 2023. How to estimate carbon foot-
print when training deep learning models? a guide and review. Environmental
Research Communications 5, 115014. URL: https://dx.doi.org/10.
1088/2515-7620/acf81b, doi:10.1088/2515-7620/acf81b.
[10] Chu, A., 2024. Us slows plans to retire coal-fired plants as power demand from
ai surges. URL: https://www.ft.com/content/ddaac44b-e245-4
c8a-bf68-c773cc8f4e63. financial Times.
[11] Clément, L.P.P.V., Jacquemotte, Q.E., Hilty, L.M., 2020. Sources of variation
in life cycle assessments of smartphones and tablet computers. Environmental
Impact Assessment Review 84, 106416.
[12] Comber, R., Eriksson, E., 2023. Computing as Ecocide, in: Ninth Computing
within Limits 2023, LIMITS. Https://limits.pubpub.org/pub/a8h46wqy.
[13] Coroama, V.C., Mattern, F., 2019. Digital rebound - why digitalization will not
redeem us our environmental sins, in: Wolff, A. (Ed.), Proceedings of the 6th
International Conference on ICT for Sustainability, ICT4S 2019, Lappeenranta,
Finland, June 10-14, 2019, CEUR-WS.org. URL: https://ceur-ws.org/
Vol-2382/ICT4S2019\_paper\_31.pdf.
[14] Dauvergne, P., 2021. The globalization of artificial intelligence: con-
sequences for the politics of environmentalism. Globalizations 18, 285–
299. URL: https://doi.org/10.1080/14747731.2
020.1785670, doi:10.1080/14747731.2020.1785670,
arXiv:https://doi.org/10.1080/14747731.2020.1785670.
[15] de Vries, A., 2023. The growing energy footprint of artificial intelligence. Joule
7, 2191–2194. URL: https://www.sciencedirect.com/science/
29
article/pii/S2542435123003653, doi:https://doi.org/10.1
016/j.joule.2023.09.004.
[16] Desislavov, R., Martínez-Plumed, F., Hernández-Orallo, J., 2023. Trends in AI
inference energy consumption: Beyond the performance-vs-parameter laws of
deep learning. Sustainable Computing: Informatics and Systems 38, 100857.
URL: https://www.sciencedirect.com/science/article/pi
i/S2210537923000124, doi:https://doi.org/10.1016/j.susc
om.2023.100857.
[17] Epoch AI, 2024a. Data on machine learning hardware. URL: https://epoc
h.ai/data/machine-learning-hardware. accessed: 2025-03-05.
[18] Epoch AI, 2024b. Data on notable AI models. URL: https://epoch.ai/d
ata/notable-ai-models. accessed: 2025-02-28.
[19] Ethayarajh, K., Jurafsky, D., 2020. Utility is in the eye of the user: A critique of
NLP leaderboards, in: Webber, B., Cohn, T., He, Y., Liu, Y. (Eds.), Proceedings
of the 2020 Conference on Empirical Methods in Natural Language Processing,
EMNLP 2020, Online, November 16-20, 2020, Association for Computational
Linguistics. pp. 4846–4853. URL: https://doi.org/10.18653/v1/20
20.emnlp-main.393, doi:10.18653/v1/2020.emnlp-main.393.
[20] Ficher, M., Bauer, T., Ligozat, A.L., 2025. A comprehensive review of the end-
of-life modeling in LCAs of digital equipment. The International Journal of Life
Cycle Assessment 30, 20 – 42. URL: https://doi.org/10.1007/s113
67-024-02367-x, doi:10.1007/s11367-024-02367-x.
[21] Forster, P., Storelvmo, T., Armour, K., Collins, W., Dufresne, J.L., Frame,
D., Lunt, D., Mauritsen, T., Palmer, M., Watanabe, M., Wild, M., Zhang, H.,
2023. The earth’s energy budget, climate feedbacks and climate sensitivity,
in: Masson-Delmotte, V., Zhai, P., Pirani, A., Connors, S., Péan, C., Berger,
S., Caud, N., Chen, Y., Goldfarb, L., Gomis, M., Huang, M., Leitzell, K.,
Lonnoy, E., Matthews, J., Maycock, T., Waterfield, T., Yelekçi, O., Yu, R.,
30
Zhou, B. (Eds.), Climate Change 2021 – The Physical Science Basis: Work-
ing Group I Contribution to the Sixth Assessment Report of the Intergovern-
mental Panel on Climate Change, Cambridge University Press. p. 923–1054.
doi:10.1017/9781009157896.009.
[22] Gnibga, W.E., Blavette, A., Orgerie, A.C., 2024. Renewable Energy in Data
Centers: the Dilemma of Electrical Grid Dependency and Autonomy Costs. IEEE
Transactions on Sustainable Computing 9, 315–328. URL: https://hal.sc
ience/hal-04189173, doi:10.1109/TSUSC.2023.3307790.
[23] Google, 2024. environmental report 2024. URL: https://www.gstatic.
com/gumdrop/sustainability/google-2024-environmental
-report.pdf. accessed March 18 2025.
[24] Gossart, C., 2015. Rebound effects and ICT: A review of the literature, in: Hilty,
L.M., Aebischer, B. (Eds.), ICT Innovations for Sustainability, Springer Interna-
tional Publishing, Cham. pp. 435–448. doi:10.1007/978-3-319-09228
-7_26.
[25] Greenfield, P., 2023. Revealed: more than 90% of rainforest carbon offsets by
biggest certifier are worthless, analysis shows. URL: https://www.thegua
rdian.com/environment/2023/jan/18/revealed-forest-car
bon-offsets-biggest-provider-worthless-verra-aoe. the
Guardian, https://www.theguardian.com/environment/2023/j
an/18/revealed-forest-carbon-offsets-biggest-provide
r-worthless-verra-aoe.
[26] Groeneveld, D., Beltagy, I., Walsh, E., Bhagia, A., Kinney, R., Tafjord, O., Jha,
A., Ivison, H., Magnusson, I., Wang, Y., Arora, S., Atkinson, D., Authur, R.,
Chandu, K., Cohan, A., Dumas, J., Elazar, Y., Gu, Y., Hessel, J., Khot, T., Merrill,
W., Morrison, J., Muennighoff, N., Naik, A., Nam, C., Peters, M., Pyatkin, V.,
Ravichander, A., Schwenk, D., Shah, S., Smith, W., Strubell, E., Subramani, N.,
Wortsman, M., Dasigi, P., Lambert, N., Richardson, K., Zettlemoyer, L., Dodge,
J., Lo, K., Soldaini, L., Smith, N., Hajishirzi, H., 2024. OLMo: Accelerating
31
the science of language models, in: Ku, L.W., Martins, A., Srikumar, V. (Eds.),
Proceedings of the 62nd Annual Meeting of the Association for Computational
Linguistics (Volume 1: Long Papers), Association for Computational Linguistics,
Bangkok, Thailand. pp. 15789–15809. URL: https://aclanthology.o
rg/2024.acl-long.841/, doi:10.18653/v1/2024.acl-long.84
1.
[27] Gupta, U., Kim, Y.G., Lee, S., Tse, J., Lee, H.H.S., Wei, G.Y., Brooks, D., Wu,
C.J., 2022. Chasing carbon: The elusive environmental footprint of computing.
IEEE Micro 42, 37–47. doi:10.1109/MM.2022.3163226.
[28] International Energy Agency (IEA), 2024. Electricity 2024. URL: https:
//www.iea.org/reports/electricity-2024. licence: CC BY 4.0.
[29] Jay, M., Ostapenco, V., Lefevre, L., Trystram, D., Orgerie, A.C., Fichel, B., 2023.
An experimental comparison of software-based power meters: focus on cpu and
gpu, in: 2023 IEEE/ACM 23rd International Symposium on Cluster, Cloud and
Internet Computing (CCGrid), pp. 106–118. doi:10.1109/CCGrid57682.
2023.00020.
[30] Jiang, H.H., Brown, L., Cheng, J., Khan, M., Gupta, A., Workman, D., Hanna,
A., Flowers, J., Gebru, T., 2023. AI art and its impact on artists, in: Proceedings
of the 2023 AAAI/ACM Conference on AI, Ethics, and Society, Association for
Computing Machinery, New York, NY, USA. p. 363–374. URL: https://do
i.org/10.1145/3600211.3604681, doi:10.1145/3600211.3604
681.
[31] Lannelongue, L., Grealey, J., Inouye, M., 2021. Green algorithms: Quantifying
the carbon footprint of computation. Advanced Science 8, 2100707. URL:
https://onlinelibrary.wiley.com/doi/abs/10.1002/advs
.202100707, doi:https://doi.org/10.1002/advs.202100707,
arXiv:https://onlinelibrary.wiley.com/doi/pdf/10.1002/advs.202100707.
[32] Lei, N., Masanet, E., 2022. Climate and technology-specific PUE and WUE
32
estimations for u.s. data centers using a hybrid statistical and thermodynamics-
based approach. Resources, Conservation and Recycling 182, 106323. URL:
https://www.sciencedirect.com/science/article/pii/S0
921344922001719, doi:https://doi.org/10.1016/j.resconre
c.2022.106323.
[33] Li, P., Yang, J., Islam, M.A., Ren, S., 2025. Making AI less "thirsty": Uncovering
and addressing the secret water footprint of AI models. URL: https://arxi
v.org/abs/2304.03271,arXiv:2304.03271.
[34] Ligozat, A.L., Lefevre, J., Bugeau, A., Combaz, J., 2022. Unraveling the hidden
environmental impacts of AI solutions for environment life cycle assessment of
AI solutions. Sustainability 14. URL: https://www.mdpi.com/2071-1
050/14/9/5172, doi:10.3390/su14095172.
[35] Lohmann, L., 2009. Toward a different debate in environmental accounting: The
cases of carbon and cost–benefit. Accounting, Organizations and Society 34,
499–534. URL: https://www.sciencedirect.com/science/arti
cle/pii/S0361368208000287, doi:https://doi.org/10.1016/
j.aos.2008.03.002.
[36] Luccioni, A.S., Strubell, E., Crawford, K., 2025. From efficiency gains to rebound
effects: The problem of jevons’ paradox in ai’s polarized environmental debate,
in: Proceedings of the 2025 ACM Conference on Fairness, Accountability, and
Transparency, Association for Computing Machinery, New York, NY, USA. p.
76–88. URL: https://doi.org/10.1145/3715275.3732007, doi:10
.1145/3715275.3732007.
[37] Luccioni, A.S., Viguier, S., Ligozat, A.L., 2023. Estimating the carbon footprint
of BLOOM, a 176b parameter language model. Journal of Machine Learning
Research 24, 1–15. URL: http://jmlr.org/papers/v24/23-0069.
html.
[38] Masanet, E., Shehabi, A., Lei, N., Smith, S., Koomey, J., 2020. Re-
calibrating global data center energy-use estimates. Science 367, 984–
33
986. URL: https://www.science.org/doi/abs/10.1
126/science.aba3758, doi:10.1126/science.aba3758,
arXiv:https://www.science.org/doi/pdf/10.1126/science.aba3758.
[39] Meta, 2024. 2024 sustainability report. URL: https://sustainability
.atmeta.com/wp-content/uploads/2024/08/Meta-2024-Sus
tainability-Report.pdf. accessed March 18 2025.
[40] Morand, C., 2025a. Code used for the paper "the environmental impacts of ma-
chine learning training keep rising evidencing rebound effect". URL: https:
//doi.org/10.5281/zenodo.15310869, doi:10.5281/zenodo.1
5310869.
[41] Morand, C., 2025b. Dataset on nvidia workstation graphics cards from 2013
to 2024 and their estimated environmental impact [data set]. URL: https:
//doi.org/10.5281/zenodo.15310940, doi:10.5281/zenodo.1
5310940.
[42] Morand, C., Ligozat, A.L., Névéol, A., 2024. MLCA: a tool for Machine
Learning Life Cycle Assessment, in: 2024 10th International Conference on
ICT for Sustainability (ICT4S), IEEE, Stockholm, Sweden. pp. 227–238. URL:
https://hal.science/hal-04643414, doi:10.1109/ICT4S64576
.2024.00031.
[43] Morrison, J., Na, C., Fernandez, J., Dettmers, T., Strubell, E., Dodge, J., 2025.
Holistically evaluating the environmental impact of creating language models,
in: The Thirteenth International Conference on Learning Representations. URL:
https://openreview.net/forum?id=04qx93Viwj.
[44] Mytton, D., 2021. Data center water consumption. npj Clean Water 4. URL:
https://doi.org/10.1038/s41545-021-00101-w, doi:10.1038/
s41545-021-00101-w.
[45] van Oers, L., Guinée, J.B., Heijungs, R., 2020. Abiotic resource depletion po-
tentials (ADPs) for elements revisited—updating ultimate reserve estimates and
34
introducing time series for production data. The International Journal of Life Cy-
cle Assessment 25, 294–308. doi:https://doi.org/10.1007/s11367
-019-01683-x.
[46] Ortar, N., Taylor, A.R.E., Velkova, J., Brodie, P., Johnson, A., Marquet, C., Pollio,
A., Cirolia, L., 2023. De Gruyter, Berlin, Boston. chapter 4 Powering ‘smart’
futures: data centres and the energy politics of digitalisation. pp. 125–168. URL:
https://doi.org/10.1515/9783110745641-005, doi:doi:10.1
515/9783110745641-005.
[47] Ostrouchov, G., Maxwell, D., Ashraf, R.A., Engelmann, C., Shankar, M., Rogers,
J.H., 2020. Gpu lifetimes on titan supercomputer: Survival analysis and reliabil-
ity, in: SC20: International Conference for High Performance Computing, Net-
working, Storage and Analysis, pp. 1–14. doi:10.1109/SC41405.2020.0
0045.
[48] Patterson, D., Gonzalez, J., Hölzle, U., Le, Q., Liang, C., Munguia, L.M.,
Rothchild, D., So, D.R., Texier, M., Dean, J., 2022. The carbon footprint
of machine learning training will plateau, then shrink. Computer 55, 18–28.
doi:10.1109/MC.2022.3148714.
[49] Pirson, T., Delhaye, T.P., Pip, A.G., Le Brun, G., Raskin, J.P., Bol, D., 2023. The
environmental footprint of IC production: Review, analysis, and lessons from
historical trends. IEEE Transactions on Semiconductor Manufacturing 36, 56–
67. doi:10.1109/TSM.2022.3228311.
[50] Reiter, E., 2025. We should evaluate real-world impact. Com-
putational Linguistics , 1–13URL: https://doi.org/
10.1162/coli.a.18, doi:10.1162/coli.a.18,
arXiv:https://direct.mit.edu/coli/article-pdf/doi/10.1162/coli.a.18/2537110/coli.a.18.pdf.
[51] Rolnick, D., Donti, P.L., Kaack, L.H., Kochanski, K., Lacoste, A., Sankaran, K.,
Ross, A.S., Milojevic-Dupont, N., Jaques, N., Waldman-Brown, A., Luccioni,
A.S., Maharaj, T., Sherwin, E.D., Mukkavilli, S.K., Kording, K.P., Gomes, C.P.,
35
Ng, A.Y., Hassabis, D., Platt, J.C., Creutzig, F., Chayes, J., Bengio, Y., 2022.
Tackling climate change with machine learning. ACM Comput. Surv. 55. URL:
https://doi.org/10.1145/3485128, doi:10.1145/3485128.
[52] Schneider, I., Xu, H., Benecke, S., Patterson, D., Huang, K., Ranganathan, P.,
Elsworth, C., 2025. Life-cycle emissions of ai hardware: A cradle-to-grave ap-
proach and generational trends. URL: https://arxiv.org/abs/2502.0
1671,arXiv:2502.01671.
[53] Schwartz, R., Dodge, J., Smith, N.A., Etzioni, O., 2020. Green AI. Commun.
ACM 63, 54–63. URL: https://doi.org/10.1145/3381831, doi:10
.1145/3381831.
[54] Scialom, T., Chakrabarty, T., Muresan, S., 2022. Fine-tuned language models are
continual learners, in: Goldberg, Y., Kozareva, Z., Zhang, Y. (Eds.), Proceedings
of the 2022 Conference on Empirical Methods in Natural Language Processing,
Association for Computational Linguistics, Abu Dhabi, United Arab Emirates.
pp. 6107–6122. URL: https://aclanthology.org/2022.emnlp-m
ain.410/, doi:10.18653/v1/2022.emnlp-main.410.
[55] Sevilla, J., Heim, L., Ho, A., Besiroglu, T., Hobbhahn, M., Villalobos, P., 2022.
Compute trends across three eras of machine learning, in: 2022 International Joint
Conference on Neural Networks (IJCNN), pp. 1–8. doi:10.1109/IJCNN550
64.2022.9891914.
[56] Shehabi, A., Smith, S., Hubbard, A., Newkirk, A., Lei, N., S., M.A.B., Holecek,
B., Koomey, J., Masanet, E., Sartor, D., 2024. 2024 United States Data Center
Energy Usage Report. Technical Report. Lawrence Berkeley National Laboratory,
Berkeley, California. LBNL-2001637.
[57] Strubell, E., Ganesh, A., McCallum, A., 2019. Energy and policy considera-
tions for deep learning in NLP, in: Korhonen, A., Traum, D.R., Màrquez, L.
(Eds.), Proceedings of the 57th Conference of the Association for Computational
Linguistics, ACL 2019, Florence, Italy, July 28- August 2, 2019, Volume 1:
36
Long Papers, Association for Computational Linguistics. pp. 3645–3650. URL:
https://doi.org/10.18653/v1/p19-1355, doi:10.18653/v1/p1
9-1355.
[58] Thompson, N., Greenewald, K., Lee, K., Manso, G.F., 2023. The Computational
Limits of Deep Learning, in: Ninth Computing within Limits 2023, LIMITS.
Https://limits.pubpub.org/pub/wm1lwjce.
[59] Treviso, M., Lee, J.U., Ji, T., van Aken, B., Cao, Q., Ciosici, M.R., Hassid, M.,
Heafield, K., Hooker, S., Raffel, C., Martins, P.H., Martins, A.F.T., Forde, J.Z.,
Milder, P., Simpson, E., Slonim, N., Dodge, J., Strubell, E., Balasubramanian, N.,
Derczynski, L., Gurevych, I., Schwartz, R., 2023. Efficient methods for natural
language processing: A survey. Transactions of the Association for Computa-
tional Linguistics 11, 826–860. URL: https://aclanthology.org/202
3.tacl-1.48/, doi:10.1162/tacl_a_00577.
[60] Varoquaux, G., Luccioni, S., Whittaker, M., 2025. Hype, sustainability, and the
price of the bigger-is-better paradigm in ai, in: Proceedings of the 2025 ACM
Conference on Fairness, Accountability, and Transparency, Association for Com-
puting Machinery, New York, NY, USA. p. 61–75. URL: https://doi.or
g/10.1145/3715275.3732006, doi:10.1145/3715275.3732006.
[61] Velkova, J., 2019. Data centres as impermanent infrastructures. Culture Machine
18. URL: http://culturemachine.net/vol-18-the-nature-o
f-data-centers/data-centers-as-impermanent/.
[62] Verdecchia, R., Sallou, J., Cruz, L., 2023. A systematic review of green
AI. WIREs Data Mining and Knowledge Discovery 13, e1507. URL:
https://wires.onlinelibrary.wiley.com/doi/abs/10.1
002/widm.1507, doi:https://doi.org/10.1002/widm.1507,
arXiv:https://wires.onlinelibrary.wiley.com/doi/pdf/10.1002/widm.1507.
[63] Wu, C., Raghavendra, R., Gupta, U., Acun, B., Ardalani, N., Maeng, K., Chang,
G., Behram, F.A., Huang, J., Bai, C., Gschwind, M., Gupta, A., Ott, M., Mel-
nikov, A., Candido, S., Brooks, D., Chauhan, G., Lee, B., Lee, H.S., Akyildiz,
37
B., Balandat, M., Spisak, J., Jain, R., Rabbat, M., Hazelwood, K.M., 2022. Sus-
tainable AI: environmental implications, challenges and opportunities, in: Mar-
culescu, D., Chi, Y., Wu, C. (Eds.), Proceedings of Machine Learning and Sys-
tems 2022, MLSys 2022, Santa Clara, CA, USA, August 29 - September 1, 2022,
mlsys.org. URL: https://proceedings.mlsys.org/paper/2022/h
ash/ed3d2c21991e3bef5e069713af9fa6ca-Abstract.html.
[64] Wu, C.J., Acun, B., Raghavendra, R., Hazelwood, K., 2024. Beyond efficiency:
Scaling AI sustainably. IEEE Micro 44, 37–46. doi:10.1109/MM.2024.34
09275.
38
