Carbon Footprint of AI Data Centers: A Life Cycle Approach#
Alexandre d’ORGEVAL1, 2, 3*, Edi ASSOUMOU2, Valentina SESSA2, Ilknur COLAK3, Stuart SHEEHAN3, Quentin AVENAS1
1 IAC Partners, Paris, France.
2 Mines Paris - PSL, Centre de Mathématiques Appliquées, Sophia Antipolis, France.
3 Schneider Electric, France
*Corresponding author. Email: [email protected]
ABSTRACT
Data centers are energy-intensive infrastructures
that generate, manage, and store information for our
interconnected society. Models based on Artificial
Intelligence (AI) such as ChatGPT are increasingly
accessible, leading to significant energy consumption
and associated carbon emissions.
Assessing the carbon footprint of AI data centers is
essential for evaluating their environmental impact and,
consequently, promoting responsible AI development
and encouraging sustainable practices. In this work, we
evaluate an AI data center's carbon footprint using a life
cycle assessment approach. Unlike existing literature, we
analyze the entire data center architecture rather than
solely focusing on the servers’ footprint. Additionally, we
assess the impact of varying the electricity mix and
extending the lifetime of servers, providing potentials for
emission reductions.
Keywords: Data center, Life Cycle Assessment, Carbon
emission, Sustainable AI
1. INTRODUCTION
Data centers are critical infrastructures supporting
the exponential growth in data generation, particularly
in Artificial Intelligence (AI) and High-Performance
Computing (HPC). AI data centers are designed to handle
high computation demands and feature advanced
hardware like GPUs or TPUs, with high rack densities.
These facilities are essential for a wide range of
applications, from data storage and processing to the
complex computations required by AI models training
and scientific research. Global data creation is projected
to rise from 1.2 trillion gigabytes in 2010 to 175 trillion
gigabytes by 2025 [1], highlighting the need for robust
data center operations. Additionally, reports by the IEA
showed that the global energy consumption for data
# This is a paper for the 16th International Conference on Applied Energy (ICAE2024), Sep. 1-5, 2024, Niigata, Japan.
centers could more than double from 460 TWh in 2022
to 1000 TWh by 2026, with countries like Denmark
potentially experiencing increases up to 15% of their
total electricity use [2,3].
The importance of assessing the environmental
impacts of data centers is underscored by their
significant energy consumption and carbon emissions.
Evaluating these impacts is critical not only for reducing
carbon footprints but also for achieving sustainability
goals set by major internet giants and data center
operators. Companies like Google and Microsoft have
pledged to match 100% of their hourly electricity
consumption with zero-carbon energy purchases [4,5],
and Amazon aims to be carbon neutral by 2040 [6].
Additionally, to get on track with the Net Zero Scenario
defined by the IEA, emissions of data centers must be cut
in half by 2030 [7]. These commitments reflect a broader
industry trend towards sustainable practices,
emphasizing the urgent need for comprehensive
environmental assessments to guide these efforts and
promote green technologies and practices.
The literature provides comprehensive
methodologies for assessing the environmental impact
of data center architectures, emphasizing both
operational and embedded emissions.
Embedded emissions include the environmental
footprint of manufacturing data center hardware. In Life
Cycle Assessment (LCA) methodologies they are
commonly used to evaluate these impacts. For example,
the ACT framework proposed in [8] based on the work
done in [9] for the case of processors provides a detailed
model for estimating the embodied carbon footprint of
processors and other key server components based on
workload characteristics, hardware specifications, and
semiconductor fab characteristics. This model has been
the basis for calculating the embedded emissions for
CPUs, GPUs, DRAM and storage in various studies
[10,11]. However, current assessments often overlook
other hardware components such as cooling systems,
Energy Proceedings Vol 55, 2025 ISSN 2004-2965 2
which can be significant contributors to energy
consumption and emissions.
The operational carbon footprint focuses on the
energy consumed during the use phase. Tools like
Carbontracker [12] enable real-time monitoring of
energy consumption and carbon emissions for training
Deep Learning models. Studies highlight the importance
of considering the carbon intensity of the energy source,
with renewable energy sources significantly reducing
operational emissions. For instance, [13] emphasizes
detailed reporting of energy consumption and suggests
strategies for reducing emissions, such as optimizing
server utilization and improving cooling efficiency.
However, these studies often focus on the carbon
footprint of servers, excluding other significant
contributors like cooling and power systems.
The literature advocates for a holistic approach,
integrating both operational and embedded emissions.
Studies such as [14] and [15] argue that achieving
sustainability requires considering the entire lifecycle of
data center components, including emissions from
manufacturing, transportation, usage, and disposal.
Innovative strategies such as carbon-intensity-aware job
scheduling are also explored to reduce the overall carbon
footprint [16,17].
In this paper, we focus on LCA for AI data centers.
Evaluating their environmental impact is crucial to
promoting responsible AI development and encouraging
sustainable practices. As an example, we mention the
examples of two Large Language Models (LLMs): GPT-3
and BLOOM. These AI models require vast computational
resources, leading to substantial energy consumption
and associated carbon emissions. In [18], it is shown that
the carbon footprint of LLMs is heavily influenced by the
energy source’s carbon intensity. For instance, training
GPT-3 resulted in emissions of approximately 552 tons of
CO2eq, mainly due to the high carbon intensity of the
energy grid used. In contrast, BLOOM’s training
emissions were significantly lower at 30 tons, benefiting
from the lower carbon intensity of the French energy
grid. These comparisons illustrate the potential for
significant emission reductions by selecting energy- efficient infrastructures and cleaner energy sources.
In this paper, we propose using LCA to take a
comprehensive approach to analyzing the carbon
footprint of AI activities. The detailed nature of LCA
facilitates a holistic understanding of AI-related carbon
footprint assessments. In our study, we consider the
broader implications of carbon footprint exercises,
examining the impact from the perspective of entire data
center architecture rather than solely focusing on the
servers' footprint.
2. DEFINITION & SCOPE
The assessment was done based on a reference
design, published by Schneider Electric, dedicated to AI
applications [19]. The architecture is a 3.6MW data
center, comprised of 2 IT rooms – one AI cluster, and one
retrofitted room with an AI cluster installed with IT room,
and equipped with Nvidia’s H100 GPU. The methodology
for this study is structured according to the phases of an
LCA ensuring a comprehensive evaluation of the carbon
emissions associated with an AI data center.
The system boundaries are defined as follows: the
assessment encompasses the entire lifecycle of the data
center, including manufacturing, operational, and end- of-life phases. The components considered within the
boundaries include IT equipment (servers, storage,
networking), cooling systems, power infrastructure, and
building infrastructure. Components that comprise a
data center are complex and usually the bill of material
are not publicly shared, making LCA analysis a tedious
process for researchers. However, companies have
adopted various strategies to assess the carbon
footprints of their products, by using methodologies to
assess the environmental footprint of their components.
Two main methodologies exist: 1) developed by the MIT
(PAIA method) the Product Carbon Footprint (PCF) [20]
which is used by companies such as HP, Apple or Dell, 2)
developed by the PEP Ecopassport institution, the
Product Environmental Profiles (PEP) [21] are used by
companies such as Schneider, Legrand or ABB. In this
work, the analysis integrates detailed emissions data for
major components based on the PEP and PCF sheets
available. For components with no PEP or PCF
evaluation, proxies based on technological
representativeness are utilized, such as using similar
components from a competitor e.g., PDU from APC [22]
replaced by this product from Legrand [23]. Additionally,
the servers’ values were built from data collected from
the literature, as no PEP or PCF sheets for servers
integrating GPUs have been found. The study assumes a
20-year lifespan for the data center with fixed
replacement rate values for components, as provided by
manufacturers.
For the geographical scope, our results were
computed for operation in France. However some
product sheets used a European mix for the use phase,
and thus were adjusted to match France’s electricity mix.
We also computed the values for two other regions,
using the mix of Europe and Germany.
3. LIFE CYCLE INVENTORY ANALYSIS
3
The inventory phase involves the collection and
quantification of data on all material and energy inputs
and outputs throughout the lifecycle of the data center
components (manufacturing, distribution, installation,
use, End-of-Life (EoL)). These values were collected from
the collected PEP or PCF sheet, at the exception of the
servers. For the servers, the values where built from the
CPU and GPU results found in [8,9], and supported by
vendor specific values for storage components, and
DRAM values extracted from [8].
Manufacturing emissions
The reference design has two type of servers, one
focused towards AI, based on NVIDIA’s DGX pod
configuration with the H100 GPU, and another more
adapted to regular IT loads. For the AI optimized servers,
the CO2 footprint for manufacturing is computed as in
[8], that is:
=
( +
+
) ⋅
(1)
with Adie the die area,
the carbon emission per
unit area related to fab location and lithography,
emissions from chemicals and gases per unit area,
emissions from raw materials, and Yield the fab yield.
For regular IT servers, which are assumed to be air- cooled and have no GPUs, the data is normalized to MW
based on available data from Lenovo, HP, and Dell [24– 27].
Operational emissions
To compute the servers’ operational footprint, it is
assumed that two states can be taken by a component:
it is either at its TDP, or at its idle point, which gives the
following formula based on [28]:
=
∗ ∑
∗
+ (1 − ) ∗
=1
(2)
with n the number of components, TDP the Thermal
Design Power, ηi the utilization rate of the component
when active – assumed at 60% at 100% load, Pidle the
power consumed at idle, and CF the emission factor of
the country’s electricity mix. The energy mix considered
in the first case is that of France. For the PEP sheets, use
phases were adjusted to match France’s emissions
factor.
Furthermore, certain equipment was not
considered in this initial assessment. This includes
pumps, chemical dosing unit for cooling, storage tank,
air/waste separator, cables, for cooling which were
excluded due to data unavailability. For the servers,
switches and connectors were excluded also, because of
a lack of available data. Future iterations of this analysis
will aim to incorporate these components to provide a
more holistic view of the embodied emissions associated
with data centers.
4. LIFE CYCLE ASSESSMENT & INTERPRETATION
The focus of the current analysis was limited to the
CO2 footprint. The main reason behind these
assumptions is that for GPUs, no emission factors other
than CO2 were found at this stage in the literature. In
contrast, for CPUs, studies such as the LCA done by Dell
on a server or this study by the German Environmental
Agency provides data for up to 5 additional impacts. PEP
sheets data for cooling and power components can
provide up to 8 additional impact categories.
Furthermore, ongoing work at Boavizta aims to expand
the assessment to include other emission factors in
future analysis [29] potentially enabling multi-criteria
assessments for entire AI data centers. Finally, the XRAF
chiller from Schneider was replaced with that of BCW
family because of data consistency.
Figure 1 shows the overall adjusted results,
considering a 20-year data center lifetime, with
emissions detailed by component category. The lifecycle
phases are dominated by the use phase (29%) and
manufacturing phase (70%).
To better understand what settings could impact
the total carbon footprint, two cases are analyzed:
varying the electricity mix and increasing the lifetime of
the servers.
4.1. Electricity mix variation
The first use case examines the impact of varying the
electricity mix on the use phase emissions of data
4
centers. The emission factors were adjusted to match
those of France, Germany, and the average of the
European Union, based on 2023 data from the Electricity
Map website [30]. The results are illustrated in Figure 3.
As anticipated, a lower energy carbon intensity leads to
a lower overall carbon footprint. A data center located in
France could potentially achieve a 3.7-fold reduction – or
76%, in carbon emissions compared to one in Germany,
primarily due to France's electricity mix, which relies
heavily on nuclear energy. Given the substantial
investments by Internet giants in Power Purchase
Agreements (PPAs) and Guarantees of Origin (GOs) –
with Amazon and Meta being the top purchasers in 2023,
accounting for 26% of all PPAs – and the increasing
regulatory constraints on data centers in Europe, a viable
strategy from a CO2 viewpoint might be to establish data
centers in low-carbon regions such as France or the
Nordic countries. While the electricity mix can influence
the decision-making process, it is not the sole or decisive
factor when selecting data center locations. Other critical
factors such as the reliability of electricity supply, land
acquisition costs, political stability, and regulatory
environments also play significant roles in these
decisions.
4.2. Extending the lifetime of components
The second use case aims to assess the impact of
extending the lifetime of servers, with results illustrated
in Figure 2. Here, the indicative lifetime of 5 years is
extended by 50% to 7.5 years. Increasing the lifespan of
components reduces the frequency of replacements,
thereby decreasing embedded emissions. However, this
comes with a trade-off: future generations of servers are
likely to be more energy-efficient, potentially lowering
the carbon footprint of the use phase. Consequently,
hardware upgrades might therefore be more
advantageous in regions with higher carbon intensity
energy sources. However, this does not take into account
potential additional carbon intensity of new processors.
Extending the lifespan of servers results in significant
emissions savings for data centers, with the benefits
varying by location due to differences in electricity mix.
For a data center in France, extending the server lifespan
can save up to 19% of total emissions over a 20-year
period. For an average European data center, the savings
amount to 8%, while a data center in Germany sees a 5%
reduction in total emissions.
By comparing the reduction in manufacturing and
end-of-life (EoL) emissions to the use phase, it becomes
evident that for a data center in Germany, extending the
server lifetime is beneficial only if the next-generation
GPU (assuming the same carbon footprint for the
embedded emissions) is less than 6.1% more energy- efficient (respectively less 9.7% for a data center using
the average European mix, and less than 45.7% for one
in France). However, this analysis does not account for
potential technological adaptations required or the
effects on other environmental impacts (not yet
computed for GPUs).
Figure 1: Carbon footprint per lifecycle state (left) and system category (right).
Figure 3 Carbon footprint for different electricity mix scenario Figure 2 Carbon footprint of data centers for different
countries when increasing server lifetime.
5
5. CONCLUSIONS
This comprehensive Life Cycle Assessment (LCA) of
an AI data center, based on a Schneider Electric’s
reference design, is the first LCA done on an AI data
center architecture. The study also highlights the critical
influence of the electricity mix on carbon emissions,
showing a potential 3.7-fold reduction through
deployment in France compared to Germany due to
France's reliance on nuclear energy. Moreover,
extending server lifespans from 5 to 7 years can save up
to 19% of emissions in France, 8% in Europe, and 5% in
Germany over the entire lifecycle, yet this must be
weighed against potential efficiency gains of newer
hardware that could offset this lifetime prolongation.
Future analyses should include all relevant components
and expand beyond CO2 emissions towards other
environmental impacts.
REFERENCE
[1] Daigle BR. Data Centers Around the World: A Quick Look
2021.
[2] Data centres & networks. IEA n.d.
https://www.iea.org/energy-system/buildings/data- centres-and-data-transmission-networks.
[3] Electricity 2024 - Analysis and forecast to 2026 2024.
[4] Our Commitment to Sustainability. Google Sustainability
n.d. https://sustainability.google/commitments/.
[5] Azure Sustainability—Sustainable Technologies |
Microsoft Azure n.d. https://azure.microsoft.com/en- us/explore/global-infrastructure/sustainability.
[6] Cloud computing durable | Amazon Web Services.
Amazon Web Services, Inc n.d.
https://aws.amazon.com/fr/sustainability/.
[7] Net Zero Emissions by 2050 Scenario (NZE) – Global
Energy and Climate Model – Analysis. IEA n.d.
https://www.iea.org/reports/global-energy-and- climate-model/net-zero-emissions-by-2050-scenario- nze.
[8] Gupta U, Elgamal M, Hills G, Wei G-Y, Lee H-HS, Brooks
D, et al. ACT: designing sustainable computer systems
with an architectural carbon modeling tool. Proceedings
of the 49th Annual International Symposium on
Computer Architecture, New York, NY, USA: Association
for Computing Machinery; 2022, p. 784–99.
[9] DTCO including Sustainability: Power-Performance- Area-Cost-Environmental score (PPACE) Analysis for
Logic Technologies | IEEE Conference Publication | IEEE
Xplore n.d.
[10] Faiz A, Kaneda S, Wang R, Osi R, Sharma P, Chen F, et al.
LLMCarbon: Modeling the end-to-end Carbon Footprint
of Large Language Models 2024.
[11] Li B, Roy RB, Wang D, Samsi S, Gadepally V, Tiwari D.
Toward Sustainable HPC: Carbon Footprint Estimation
and Environmental Implications of HPC Systems.
Proceedings of the International Conference for High
Performance Computing, Networking, Storage and
Analysis, 2023, p. 1–15.
[12] Anthony LFW, Kanding B, Selvan R. Carbontracker:
Tracking and Predicting the Carbon Footprint of Training
Deep Learning Models 2020.
[13] Henderson P, Hu J, Romoff J, Brunskill E, Jurafsky D,
Pineau J. Towards the Systematic Reporting of the
Energy and Carbon Footprints of Machine Learning
2022.
[14] Wu C-J, Raghavendra R, Gupta U, Acun B, Ardalani N,
Maeng K, et al. Sustainable AI: Environmental
Implications, Challenges and Opportunities 2022.
[15] Strubell E, Ganesh A, McCallum A. Energy and Policy
Considerations for Deep Learning in NLP 2019.
[16] Zhang G, Zhang S, Zhang W, Shen Z, Wang L. Distributed
Energy Management for Multiple Data Centers With
Renewable Resources and Energy Storages. IEEE
Transactions on Cloud Computing 2022;10:2469–80.
[17] Zhao D, Zhou J. An energy and carbon-aware algorithm
for renewable energy usage maximization in distributed
cloud data centers. Journal of Parallel and Distributed
Computing.
[18] Luccioni AS, Viguier S, Ligozat A-L. Estimating the Carbon
Footprint of BLOOM, a 176B Parameter Language Model
2022.
[19] EcoStruxure Data Center Reference Designs n.d.
https://www.se.com/ww/en/work/solutions/for- business/data-centers-and-networks/reference- designs/.
[20] PAIA - Information & Communication Technology.
Quantis n.d. https://quantis.com/who-we-guide/our- impact/sustainability-initiatives/paia/.
[21] Create a PEP n.d. http://www.pep- ecopassport.org/create-a-pep/.
[22] APC NetShelter Rack PDU Advanced, Metered, 3Phase,
22.1kW 400V 32A or 17.3kW 415V 30A, 48 Outlets,
IEC309 - APDU10350ME | APC Francophone Africa n.d.
https://www.apc.com/africa/fr/product/APDU10350M E/apc-netshelter-rack-pdu-advanced-metered-3phase- 22-1kw-400v-32a-or-17-3kw-415v-30a-48-outlets- iec309.
[23] Technology S. PRO2TM HDOT® 11.0kW - 22.0kW (48-54)
outlets Rack PDU. Server Technology n.d.
https://www.servertech.com/power-distribution-unit- pdu/switched-pops-pdu/0u-vertical-2N34.
[24] pcf-thinksystem-sr250.pdf n.d.
[25] lca-poweredge-r6515-r7515-r6525-r7525.pdf n.d.
[26] Full_LCA_Dell_R740.pdf n.d.
[27] VGL0LLMZ.pdf n.d.
[28] Ji S, Yang Z, Chen X, Hu J, Shi Y, Jones AK, et al. Towards
Data-center Level Carbon Modeling and Optimization
for Deep Learning Inference 2024.
[29] Boavizta. Missions | Boavizta n.d.
https://boavizta.org/missions.
[30] Live 24/7 CO₂ emissions of electricity consumption n.d.
http://electricitymap.tmrow.co.
51作业君版权所有
