The Carbon Cost of Conversation, Sustainability in the Age of Language Models

Md. Mainul Islam; Mohammad Shakhawat Hossen; Naznin Akter; Prasun Goswami; Sayed Mahbub Hasan Amiri; Sayed Majhab Hasan Amiri

REVIEW 4 major objections 5 minor 91 references

The Carbon Cost of Conversation, Sustainability in the Age of Language Models

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper tries to establish that large language models carry an undercounted environmental cost across training, cooling, hardware, and disposal, and that current voluntary reporting hides rather than mitigates it.

desk verdict A broad survey on LLM environmental costs with a correct thesis and an untrustworthy quantitative core; the numbers fail elementary sanity checks. read the letter →

arxiv 2507.20018 v2 pith:QIK3YFOF submitted 2025-07-26 cs.CY cs.AIcs.CL

Sayed Mahbub Hasan Amiri , Prasun Goswami , Md. Mainul Islam , Mohammad Shakhawat Hossen , Sayed Majhab Hasan Amiri , Naznin Akter This is my paper

classification cs.CYcs.AIcs.CL

keywords EnvironmentalSustainabilityNaturalLanguageProcessing(NLP)CarbonFootprintEthicalAIResourceDepletionPolicyGovernanceGreenComputingEquityinTechnology

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper argues that large language models carry a serious, largely unaccounted environmental burden—training electricity, cooling water, hardware manufacturing, and disposal—and that current industry reporting and regulation do not capture it. To make the case it collects published carbon and water estimates for models like GPT-3, PaLM, BLOOM, BERT, and GPT-4, and uses Mistral 7B as a counterexample of a far cheaper model. A sympathetic reader would care because the paper's central numbers, if right, put a single frontier training run on the scale of hundreds of cars' annual emissions, making AI's climate cost a current policy problem rather than a speculative one.

What carries the argument

The argument runs on lifecycle accounting applied to specific models: estimate training electricity as hardware count multiplied by runtime and device power draw, convert to CO2 using a grid carbon intensity, then add water for cooling and emissions from manufacturing and disposal. The paper compares those figures against everyday emissions such as cars, flights, and homes, and against what companies disclose, so the mechanism is as much about reporting gaps as about energy math. Tables 1, 5, and 6 carry the quantitative load, with corporate sustainability scores and model-by-model footprints doing the comparative work.

What would settle it

Ask OpenAI or Microsoft to publish GPT-4's actual training energy use; if the true figure is below roughly 1,000 MWh, the paper's 3,500–4,200 MWh estimate fails. Alternatively, a back-of-envelope check on Mistral 7B—8,000 H100 GPUs at about 700 W each for 14 days implies over 1.8 GWh, nearly twenty times the 98 MWh the paper reports—would show the estimates are not internally consistent.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that the environmental cost of large language models is already comparable to familiar high-emission activities and grows with model scale, while remaining almost invisible in corporate reporting. The paper's headline quantitative claim is that GPT-4, estimated at 1.7 trillion parameters trained on 25,000 A100 GPUs for 90–100 days, likely consumed 3,500–4,200 MWh of electricity and emitted 1,500–1,800 metric tons of CO2—equivalent to the annual emissions of about 300 gasoline-powered cars—plus roughly 1.4 million liters of water for cooling. It contrasts this with Mistral 7B, which it reports at 98 MWh and 42 metric tons of CO2, less than 3% of GPT-4's footprint. The paper further claims that across the industry, redundant training runs, quarterly retraining, cooling overhead, and unregulated inference multiply this cost, and that the burden falls disproportionately on the Global South through resource extraction and e-waste.

Load-bearing premise

The headline numbers depend on guessed hardware counts and training durations for models whose owners do not disclose them; if GPT-4 was not trained on 25,000 A100s for 90–100 days, or Mistral 7B on 8,000 H100s for 14 days, the emission estimates do not hold.

Editorial extensions

If this is right

A single frontier training run can emit as much CO2 as about 300 cars driven for a year, so the cumulative footprint of a handful of large models is a material emissions source in its own right.
Smaller, openly trained models can deliver comparable performance at under a tenth of the energy, which means near-term efficiency gains are available without waiting for new regulation.
If emission reporting standards were applied, the paper's numbers imply that mandatory Scope 3 disclosure would substantially revise corporate climate claims.
A carbon tax at the levels the paper discusses would add only modestly to frontier training costs, so stronger policy instruments would be needed to change behavior.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The paper's own Mistral 7B numbers fail a basic sanity check: 8,000 H100 GPUs running 14 days would draw roughly 1.5–2 GWh, not 98 MWh, so at least one of those figures is wrong; all its quantitative estimates should be read as order-of-magnitude.
That same contradiction suggests a cheap, testable audit: check any published training-energy figure against fleet size, runtime, and device wattage, and inconsistent claims stand out immediately.
If efficiency lowers the cost per query, total energy use can still grow as cheaper AI invites more use; the paper's policy proposals would be stronger if they targeted total compute, not just per-model efficiency.
Tracing the paper's equity claims to specific data-center regions and supply chains would let third parties verify which communities actually bear the water and e-waste burden.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

The headline: this paper is a survey/opinion piece that gets the general direction right—LLM training and deployment have real environmental costs—but its quantitative backbone is not trustworthy. The authors compile numbers from Strubell, Patterson, Luccioni, Li, etc., and the compilation is riddled with arithmetic errors and misattributions.

What it does well: it covers a wide range of topics—water use, e-waste, corporate pledges, policy gaps, equity—and the reference list is extensive. For someone new to the area, it could serve as a broad orientation. The framing of greenwashing and the Global South burden is not novel, but it is worth restating if done accurately.

The soft spots are serious. The Mistral 7B case study is the starkest example: Section 8.A.ii says the model was trained on 8,000 H100s for 14 days and consumed 98 MWh. At the H100's nominal 700 W, that hardware fleet would draw roughly 1,880 MWh; even at half power, you get ~940 MWh. The 98 MWh figure is off by a factor of 10–19, and the emissions number (42 tCO2) inherits that error. The GPT-4 estimate fails the same sanity check: 25,000 A100s for 90–100 days at 400 W gives ~22,800 MWh, not the claimed 3,500–4,200. BERT-base is listed at 79 MWh and 33 tCO2, orders of magnitude above the original Strubell numbers. These are not cosmetic typos; the paper's central claim that 'training a single LLM emits CO2 equivalent to hundreds of cars' rests on figures that collapse under inspection.

There are also citation problems. Several references are vague (e.g., 'AI Now Institute (2023)' has no author or title), and some citations are clearly wrong—attributing the 1,287 MWh/552 tCO2 GPT-3 figure to Strubell 2019, which predates GPT-3. The paper contains no new measurements, no model, no reproducible pipeline. Its value as a review is undermined because readers cannot trust the numbers, and the original sources are often misquoted.

Recommendation: I would not send this to a serious referee as a quantitative contribution, and I would not cite its numbers in anything I write. It could be a very rough draft for a perspective piece after the authors redo every calculation from primary sources and fix the internal contradictions. As it stands, desk reject is the right call.

Would I bring it to reading group? Maybe, as a negative example of why source-checking matters. But not for its content.

Referee Report

4 major / 5 minor

Summary. The paper argues that large language models (LLMs) impose a severe and underappreciated environmental burden, quantifying the carbon, water, and e-waste costs of models such as GPT-4, GPT-3, BERT-base, PaLM, BLOOM, and Mistral 7B. It contrasts corporate sustainability claims with actual practices, identifies regulatory gaps and greenwashing, and proposes technical, policy, and cultural reforms. The central quantitative claim, stated in the abstract and Section 8, is that training a single frontier LLM can emit CO2 equivalent to hundreds of cars driven annually, with GPT-4 specifically estimated at 3,500-4,200 MWh and 1,500-1,800 metric tons of CO2.

Significance. The topic is important and the paper draws attention to real issues in AI sustainability, including Scope 3 reporting gaps, water usage in data centers, and the equity dimensions of e-waste. If its quantitative claims were reliable, the paper could serve as a useful wake-up call. However, the quantitative foundation is unsound: the paper's own numbers are internally inconsistent, physically implausible, and often at odds with the cited sources. It provides no derivations, no uncertainty ranges, and no reproducible methodology. The qualitative discussions are broadly consistent with the existing literature, but the paper does not make a new scientific contribution beyond summarizing known concerns. The central claims, as stated, cannot be trusted in their current form.

major comments (4)

[Section 8.A.ii, Table 5] The Mistral 7B numbers are internally contradictory. The paper states that Mistral 7B was trained on 8,000 NVIDIA H100 GPUs for 14 days and consumed 98 MWh. At the H100's nominal 700 W power envelope, this configuration would draw approximately 8,000 x 0.7 kW x 14 x 24 h = 1,882 MWh, a factor of 19 above the reported value; even at a very conservative average of 350 W per GPU, the total is about 941 MWh. The paper then uses the 98 MWh figure to claim that Mistral uses 'less than 3% of GPT-4's footprint.' This efficiency comparison collapses if the hardware or energy figure is corrected, so this is not a minor typo.
[Section 8.A.i, Table 5] The GPT-4 estimate fails the same sanity check. With 25,000 A100 GPUs at 400 W for 95 days, the raw energy is about 22,800 MWh, five to six times the reported 3,500-4,200 MWh, and no utilization model or efficiency factor is provided. Additionally, the paper states that GPT-4's inference 'demands 50 MWh monthly, akin to powering 4,000 U.S. homes.' At a typical 30 kWh/day per home, 4,000 homes would consume about 3,600 MWh per month, not 50 MWh. Because the abstract and conclusion rest on the claimed scale of GPT-4's footprint, these inconsistencies are load-bearing.
[Table 1, Section 3.A, Section 6.A] The BERT-base figures contradict the cited source. The paper reports that BERT-base consumes 79 MWh and emits 33 metric tons of CO2, citing Strubell et al. [66], but that paper reports BERT-base training energy on the order of 1.5 MWh and emissions well below 33 tCO2. This erroneous baseline is then reused in Section 6.A to claim that TinyBERT reduces training energy from 79 MWh to 6 MWh, so the error propagates into the efficiency-solution argument.
[Sections 8.A.i, 8.A.ii] No derivation or methodology is provided for any of the energy, emission, or water numbers in Table 5. The hardware counts and training durations are attributed to sources that do not appear to contain them: the published Mistral 7B technical report does not specify 8,000 H100 GPUs for 14 days, and the cited 'AI Now Institute' report is not a standard source for GPT-4's undisclosed training configuration. Every quantitative statement in the paper is inherited from other work or presented as an unsupported assumption, with no error bars, sensitivity analysis, or independent verification. Given the internal arithmetic failures, the entire numerical core of the paper is unreliable.

minor comments (5)

[Section 4.A, Table 5] The word 'litters' is repeatedly used instead of 'liters' or 'litres' (e.g., '700,000 litters' and '4.9 million litters').
[Section 7.C] The sentence 'Ethical NLP demands cantering these communities in sustainability efforts' contains a typo: 'cantering' should be 'centering.'
[Section 2.B] The claim that 'doubling a model's size typically quadruples its training time and energy use' is not generally supported by the scaling-law literature cited (Kaplan et al. [45]); energy scaling depends on hardware, batch size, and training steps.
[Section 8.A.ii] The statement that Mistral 7B 'reduced energy use by 70% compared to similarly sized models like LLaMA-7B' is not supported by the cited Mistral AI technical report, which does not report such a comparison.
[References] Several references appear mismatched or incomplete: for example, Strubell et al. [66] is cited for GPT-3 training energy, but the paper predates GPT-3 and addresses BERT/Transformer models. The authors should verify every reference against the specific claim it supports.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: the paper compiles externally sourced estimates without fitting its inputs or renaming its own assumptions as predictions.

full rationale

The paper's quantitative backbone (GPT-3 1,287 MWh / 552 tCO2; PaLM 2,500 MWh / 1,100 tCO2; GPT-4 3,500-4,200 MWh / 1,500-1,800 tCO2; Mistral 7B 98 MWh / 42 tCO2) is presented as inherited from cited third-party analyses (Strubell et al. 2019; Patterson et al. 2021; Luccioni et al. 2022, 2023; Mistral AI 2023), not derived from assumptions introduced in this paper. No parameter is fitted to a subset of data and then renamed a prediction; no uniqueness theorem or ansatz from the authors' prior work is invoked; and the paper's own text does not define any quantity in terms of the quantity it purports to estimate. The reference list contains no self-citation chain bearing on the main claim. The serious internal inconsistency between Mistral's stated 8,000 H100 GPUs for 14 days and the reported 98 MWh, and the similar mismatch for GPT-4's 25,000 A100 GPUs versus 3,500-4,200 MWh, undermines the reliability of the compiled numbers, but inconsistency of sourced estimates is a correctness and verification problem, not a circularity of derivation. Accordingly, no circular step can be exhibited with the required quote-and-reduction specificity, and the honest finding is no significant circularity.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The paper introduces no new parameters or entities, but it relies on several untested assumptions. The listed free parameters are the unverified hardware/time assumptions used to estimate GPT-4 and Mistral 7B footprints, plus one misquoted BERT-base value. The axioms include the reliability of corporate reports, the use of cited scaling laws, and the faithful representation of third-party estimates.

free parameters (4)

GPT-4 assumed training hardware = 25,000 A100 GPUs for 90-100 days
Used in Section 8.A.i to estimate GPT-4 training energy of 3,500-4,200 MWh and 1,500-1,800 tCO2. No source provided; OpenAI has not disclosed training details.
Mistral 7B assumed training hardware = 8,000 H100 GPUs for 14 days
Used in Section 8.A.ii to estimate 98 MWh and 42 tCO2. This hardware/time combination is physically inconsistent with the stated energy; likely a misquote of Mistral's actual setup.
BERT-base training energy and emissions = 79 MWh, 33 tCO2
Table 1 cites Strubell et al. (2019) but the values are roughly 50x higher than the actual figures in that paper (about 1.5 MWh, 0.65 tCO2 for BERT-base).
AWS AI energy consumption = 12.3 TWh annually
Section 8.B.ii states independent analysts estimate AWS AI workloads use 12.3 TWh annually, but no specific source is cited; likely a fabricated or speculative figure.

assumptions (4)

domain assumption Grid carbon intensity factors used to convert electricity to CO2 are representative and constant across model training locations
The paper reports emissions multiples (e.g., GPT-3 552 tCO2 from 1,287 MWh) without accounting for the grid mix; it even notes Iowa coal-heavy for GPT-4 but does not adjust the overall calculations.
domain assumption Corporate sustainability disclosures (Google, Microsoft, Amazon) are accurate enough for comparison
The paper relies on company reports for renewable percentages, PUE, and offset reliance while simultaneously accusing the same companies of greenwashing (Sections 5.A, 5.B, Tables 3 and 6).
domain assumption Scaling laws (doubling model size quadruples training time/energy) apply to all models cited
Section 2.A states computation scales cubically with model size, a simplification of Kaplan et al. scaling laws; used to support the 'chasing scale' argument.
domain assumption All cited third-party studies are faithfully represented
The paper repeatedly attributes numbers to sources that do not contain them (e.g., BERT-base, GPT-4 energy estimates), so the reliability of the whole review depends on an assumption that is contradicted by the text itself.

how reviews work

0 comments

Cite this review

Pith. "Pith review of The Carbon Cost of Conversation, Sustainability in the Age of Language Models." pith.science (2026). https://pith.science/paper/QIK3YFOF

@misc{pith2026250720018,
  author       = {Pith},
  title        = {Pith review of: The Carbon Cost of Conversation, Sustainability in the Age of Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QIK3YFOF}},
  note         = {Machine review of arXiv:2507.20018}
}

read the original abstract

Large language models (LLMs) like GPT-3 and BERT have revolutionized natural language processing (NLP), yet their environmental costs remain dangerously overlooked. This article critiques the sustainability of LLMs, quantifying their carbon footprint, water usage, and contribution to e-waste through case studies of models such as GPT-4 and energy-efficient alternatives like Mistral 7B. Training a single LLM can emit carbon dioxide equivalent to hundreds of cars driven annually, while data centre cooling exacerbates water scarcity in vulnerable regions. Systemic challenges corporate greenwashing, redundant model development, and regulatory voids perpetuate harm, disproportionately burdening marginalized communities in the Global South. However, pathways exist for sustainable NLP: technical innovations (e.g., model pruning, quantum computing), policy reforms (carbon taxes, mandatory emissions reporting), and cultural shifts prioritizing necessity over novelty. By analysing industry leaders (Google, Microsoft) and laggards (Amazon), this work underscores the urgency of ethical accountability and global cooperation. Without immediate action, AIs ecological toll risks outpacing its societal benefits. The article concludes with a call to align technological progress with planetary boundaries, advocating for equitable, transparent, and regenerative AI systems that prioritize both human and environmental well-being.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

91 extracted references · 52 canonical work pages

[1]

Ayana, G. et al. (2024). Decolonizing global AI governance: assessment of the state of decolonized AI governance in Sub-Saharan Africa. R. Soc. Open Sci. 11: 231994. https://doi.org/10.1098/rsos.231994

work page doi:10.1098/rsos.231994 2024
[2]

AI Now Institute. (2023). The environmental costs of AI: A case study of GPT-4

2023
[3]

Amazon SEC Filing. (2023). Definitive Proxy Statement. Retrieved from https://www.sec.gov

2023
[4]

Amazon Sustainability Report. (2022). Amazon.com. Retrieved from https://sustainability.aboutamazon.com

2022
[5]

Andrae, A. S. G. (2020). New perspectives on internet electricity use in

2020
[6]

Andrae, A. S. G. (2023). Hyperscale data centers and AI’s energy paradox. Joule, 7(5), 1–15

2023
[7]

W., Althaf, S., & Cruz Rios, F

Babbitt, C. W., Althaf, S., & Cruz Rios, F. (2021). The role of consumerism in electronic waste. Nature Electronics, 4(3), 152–160

2021
[8]

Belkhir, L., & Elmeligi, A. (2018). Assessing ICT global emissions footprint: Trends to 2040 & recommendations. Journal of Cleaner Production, 177, 448–

2018

Show all 91 references

[9]

M., Gebru, T., McMillan-Major, A., & Shmitchell, S

Bender, E. M., Gebru, T., McMillan-Major, A., & Shmitchell, S. (2021). On the dangers of stochastic parrots: Can language models be too big? Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency, 610–

2021
[10]

Biderman, S. et al. (2023). Pythia: A suite for analyzing large language models across training and scaling. arXiv preprint arXiv:2304.01373

2023 arXiv
[11]

Boyd, S. B. et al. (2022). Life-cycle assessment of semiconductor devices. Environmental Science & Technology, 56(8), 4623–

2022
[12]

Brock, A., & Sovacool, B. K. (2023). Corporate carbon neutrality: A critical review of offsetting practices. Energy Research & Social Science, 98, 103027. https://doi.org/10.1016/j.erss.2023.103027

2023
[13]

B., Mann, B., Ryder, N., Subbiah, M., Kaplan, J., Dhariwal, P.,

Brown, T. B., Mann, B., Ryder, N., Subbiah, M., Kaplan, J., Dhariwal, P., ... & Amodei, D. (2020). Language models are few-shot learners. Advances in Neural Information Processing Systems, 33, 1877–

2020
[14]

F., & Dooley, K

Carton, W., Lund, J. F., & Dooley, K. (2023). Undoing equivalence: Rethinking carbon accounting for just carbon removal. Frontiers in Climate, 5, 1–

2023
[15]

CDP. (2023). Global supply chain report 2023. CDP Worldwide. Retrieved from https://www.cdp.net

2023
[16]

& Fiedel, N

Chowdhery, A., Narang, S., Devlin, J., Bosma, M., Mishra, G., Roberts, A., ... & Fiedel, N. (2022). PaLM: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311. https://doi.org/10.48550/arXiv.2204.02311 Education Research Team [Education and Computer Science] 19

2022 doi arXiv
[17]

Cihon, P., Schuett, J., & Baum, S. D. (2021). Corporate governance of AI in the public interest. AI and Ethics, 1(4), 461–468. https://doi.org/10.1007/s43681-021- 00067-y

2021 doi
[18]

Climate Change AI. (2024). AI for Climate Hackathon 2024. Retrieved from https://www.climatechange.ai

2024
[19]

Davies, M., Wild, A., Orchard, G., Sandamirskaya, Y., Guerra, G. A. F., Joshi, P., ... & Stewart, T. C. (2024). Loihi 2: A neuromorphic research processor. IEEE Transactions on Computers, 73(3), 1–14. https://doi.org/10.1109/TC.2023.3324360

2024
[20]

https://doi.org/10.3389/fclim.2023.1137000

2023
[21]

M., Tong, S., Lepikhin, D., Xu, Y.,

Du, N., Huang, Y., Dai, A. M., Tong, S., Lepikhin, D., Xu, Y., ... & Zhou, Y. (2022). GLaM: Efficient scaling of language models with mixture-of- experts. Proceedings of the 39th International Conference on Machine Learning, 5547–5569. https://doi.org/10.48550/arXiv.2112.06905

2022 doi arXiv
[22]

Energy Justice Network. (2023). Microsoft’s dirty cloud. Retrieved from https://www.energyjustice.net

2023
[23]

EPA. (2023). Greenhouse gas equivalencies calculator. United States Environmental Protection Agency. Retrieved from https://www.epa.gov/energy/greenhouse-gas- equivalencies-calculator

2023
[24]

European Commission. (2023). Carbon Border Adjustment Mechanism. EUR- Lex. https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:52023PC0289

2023
[25]

Evans, R., & Gao, J. (2023). DeepMind AI reduces Google data center cooling bill by 40%. DeepMind Blog. Retrieved from https://deepmind.google

2023
[26]

W., Lee, K., & Toutanova, K

Devlin, J., Chang, M. W., Lee, K., & Toutanova, K. (2018). BERT: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805. https://doi.org/10.48550/arXiv.1810.04805

2018 doi arXiv
[27]

FTC. (2023). FTC cracks down on tech company for deceptive environmental claims. Federal Trade Commission. Retrieved from https://www.ftc.gov

2023
[28]

Furber, S., Rhodes, O., Temple, S., & Plana, L. A. (2024). SpiNNaker2: A 10 million core neuromorphic system for embodied AI. Frontiers in Neuroscience, 18, 1–12. https://doi.org/10.3389/fnins.2024.1234567

2024
[29]

Gavin, M., Andrae, A., & Galbraith, E. (2023). Climate TRACE: Tracking real-time atmospheric carbon emissions. Environmental Science & Technology, 57(8), 4623–

2023
[30]

Google Sustainability. (2023). AI for climate action. Retrieved from https://sustainability.google

2023
[31]

Green Software Foundation (GSF). (2023). Software Carbon Intensity Specification. Greensoftware.foundation. Retrieved from https://greensoftware.foundation

2023
[32]

P., Kuehr, R., & Bel, G

Forti, V., Baldé, C. P., Kuehr, R., & Bel, G. (2020). The Global E-Waste Monitor

2020
[33]

Harris, T. (2021). The environmental cost of AI. Nature Machine Intelligence, 3(6), 423–425. https://doi.org/10.1038/s42256-021-00343-w

2021 doi
[34]

Harris, T., Rolnick, D., & Donti, P. L. (2023). Prioritizing applications for climate action. Nature Machine Intelligence, 5(6), 567–570. https://doi.org/10.1038/s42256- 023-00671-z

2023 doi
[35]

Hasani, R., Lechner, M., Amini, A., Rus, D., & Grosu, R. (2023). Liquid structural plasticity for sustainable neural networks. Nature Machine Intelligence, 5(6), 567–

2023
[36]

Hugging Face. (2023). Carbon Leaderboard. Huggingface.co. Retrieved from https://huggingface.co

2023
[37]

https://doi.org/10.1021/acs.est.3c01234

doi
[38]

IBM Research. (2023). Quantum-enhanced natural language processing. IBM Journal of Research and Development, 67(5/6), 1–

2023
[39]

International Transport Forum. (2023). Decarbonising freight transport. OECD Publishing. https://doi.org/10.1787/b8d7dac2-en

2023 doi
[41]

& Liu, Q

Jiao, X., Yin, Y., Shang, L., Jiang, X., Chen, X., Li, L., ... & Liu, Q. (2020). TinyBERT: Distilling BERT for natural language understanding. Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, 4163–

2020
[42]

Jones, N. (2018). How to stop data centres from gobbling up the world’s electricity. Nature, 561(7722), 163–166. https://doi.org/10.1038/d41586-018- 06610-y

2018 doi
[43]

P., Young, C., Patil, N., Patterson, D., Agrawal, G., Bajwa, R.,

Jouppi, N. P., Young, C., Patil, N., Patterson, D., Agrawal, G., Bajwa, R., ... & Boyle, R. (2017). In-datacenter performance analysis of a tensor processing unit. Proceedings of the 44th Annual International Symposium on Computer Architecture, 1–12. https://doi.org/10.1145/30...

2017
[44]

B., Avent, B., Bellet, A., Bennis, M., Bhagoji, A

Kairouz, P., McMahan, H. B., Avent, B., Bellet, A., Bennis, M., Bhagoji, A. N., ... & Zhao, S. (2021). Advances and open problems in federated learning. Foundations and Trends® in Machine Learning, 14(1–2), 1–

2021
[45]

B., Chess, B., Child, R.,

Kaplan, J., McCandlish, S., Henighan, T., Brown, T. B., Chess, B., Child, R., ... & Amodei, D. (2020). Scaling laws for neural language models. arXiv preprint arXiv:2001.08361. https://doi.org/10.48550/arXiv.2001.08361

2020 doi arXiv
[46]

Hugging Face. (2024). EcoBench: Benchmarking energy-efficient AI. Huggingface.co. Retrieved from https://huggingface.co/ecobench

2024
[47]

H., & Kang, J

Lee, J., Yoon, W., Kim, S., Kim, D., Kim, S., So, C. H., & Kang, J. (2020). BioBERT: a pre-trained biomedical language representation model for biomedical Education Research Team [Education and Computer Science] 21 text mining. Bioinformatics, 36(4), 1234–

2020
[48]

https://doi.org/10.1147/JRD.2023.3323982

2023
[49]

Lottick, K., Susai, S., & Doshi-Velez, F. (2023). Carbon-aware computing for machine learning. Proceedings of the 40th International Conference on Machine Learning, 1–10. https://doi.org/10.48550/arXiv.2302.04036

2023 doi arXiv
[50]

Q., Sablayrolles, A., Mensch, A., Bamford, C., Chaplot, D

Jiang, A. Q., Sablayrolles, A., Mensch, A., Bamford, C., Chaplot, D. S., Casas, D., ... & Sayed, W. E. (2023). Mistral 7B. arXiv preprint arXiv:2310.06825. https://doi.org/10.48550/arXiv.2310.06825

2023 doi arXiv
[51]

S., Viguier, S., & Ligozat, A

Luccioni, A. S., Viguier, S., & Ligozat, A. L. (2023). Estimating the carbon footprint of frontier AI models. Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency, 1–

2023
[52]

Microsoft Research. (2023). Carbon Aware SDK. GitHub Repository. Retrieved from https://github.com/microsoft

2023
[53]

Microsoft Sustainability Report. (2023). Microsoft.com. Retrieved from https://www.microsoft.com/sustainability

2023
[54]

Mistral AI. (2023). Mistral 7B: Efficient and open language models. Mistral AI Technical Report. https://doi.org/10.5281/zenodo.8313433

2023 doi
[55]

C., Mocanu, E., Stone, P., Nguyen, P

Mocanu, D. C., Mocanu, E., Stone, P., Nguyen, P. H., Gibescu, M., & Liotta, A. (2023). Dynamic sparse training for sustainable AI. IEEE Transactions on Neural Networks, 34(8), 1–15. https://doi.org/10.1109/TNNLS.2023.3248052

2023
[56]

NIST AI Risk Management Framework. (2023). National Institute of Standards and Technology. https://doi.org/10.6028/NIST.AI.100-1

2023 doi
[57]

M., Rothchild, D.,

Patterson, D., Gonzalez, J., Le, Q., Liang, C., Munguia, L. M., Rothchild, D., ... & Dean, J. (2021). Carbon emissions and large neural network training. arXiv preprint arXiv:2104.10350. https://doi.org/10.48550/arXiv.2104.10350

2021 doi arXiv
[58]

Lacoste, A., Luccioni, A., Schmidt, V., & Dandres, T. (2019). Quantifying the carbon emissions of machine learning. arXiv preprint arXiv:1910.09700. https://doi.org/10.48550/arXiv.1910.09700

2019 doi arXiv
[59]

Sanh, V., Wolf, T., & Rush, A. (2020). Movement pruning: Adaptive sparsity by fine-tuning. Advances in Neural Information Processing Systems, 33, 20378– 20389. https://doi.org/10.48550/arXiv.2005.07683

2020 doi arXiv
[60]

Schmidt, V., Goyal, A., Joshi, A., & Feldmann, A. (2023). Carbontracker: AI for sustainable machine learning. Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency, 1–

2023
[61]

A., & Ren, S

Li, P., Yang, J., Islam, M. A., & Ren, S. (2023). Making AI less “thirsty”: Uncovering and addressing the secret water footprint of AI models. arXiv preprint arXiv:2304.03271. https://doi.org/10.48550/arXiv.2304.03271

2023 doi arXiv
[62]

& Lintner, W

Shehabi, A., Smith, S., Sartor, D., Brown, R., Herrlin, M., Koomey, J., ... & Lintner, W. (2016). United States data center energy usage report. Lawrence Berkeley National Laboratory. https://doi.org/10.2172/1372902

2016 doi
[63]

S., Viguier, S., & Ligozat, A

Luccioni, A. S., Viguier, S., & Ligozat, A. L. (2022). Estimating the carbon footprint of BLOOM, a 176B parameter language model. Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency, 1877–

2022
[64]

Smith, S., Patwary, M., Norick, B., LeGresley, P., Rajbhandari, S., Casper, J., ... & He, Y. (2022). Using deepspeed and megatron to train megatron-turing NLG 530B, a large-scale generative language model. arXiv preprint arXiv:2201.11990. https://doi.org/10.48550/arXiv.2201.11990

2022 doi arXiv
[65]

Sovacool, B. K. (2021). When subterranean slavery supports sustainability transitions? Power, patriarchy, and child labor in artisanal Congolese cobalt mining. The Extractive Industries and Society, 8(1), 271–

2021
[67]

Sze, V., Chen, Y., Yang, T., & Emer, J. (2023). Efficient processing of deep neural networks: A tutorial and survey. Proceedings of the IEEE, 111(5), 1–

2023
[68]

UK Government. (2023). Streamlined Energy and Carbon Reporting. Gov.uk. Retrieved from https://www.gov.uk

2023
[69]

UN Water. (2022). The United Nations World Water Development Report 2022: Groundwater: Making the invisible visible. UNESCO. https://unesdoc.unesco.org/ark:/48223/pf0000380721

2022
[70]

USGS. (2023). Mineral Commodity Summaries 2023. U.S. Geological Survey. https://doi.org/10.3133/mcs2023

2023 doi
[71]

West, T. A. P., Börner, J., & Sills, E. O. (2023). Overstated carbon emission reductions from voluntary REDD+ projects in the Brazilian Amazon. Proceedings of the National Academy of Sciences, 120(21), e2209613120. https://doi.org/10.1073/pnas.2209613120

2023 doi
[72]

Wu, Y., Liu, J., Du, Y., & Ren, S. (2022). Quantization meets federated learning. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 1–10. https://doi.org/10.1109/CVPR52688.2022.00008

2022
[73]

Preskill, J. (2023). Quantum computing and the entanglement frontier. Annual Review of Condensed Matter Physics, 14(1), 1–25. https://doi.org/10.1146/annurev- conmatphys-031720-103511

2023 doi
[76]

https://doi.org/10.1145/3593013.3594098
[77]

A., & Etzioni, O

Schwartz, R., Dodge, J., Smith, N. A., & Etzioni, O. (2020). Green AI. Communications of the ACM, 63(12), 54–63. https://doi.org/10.1145/3381831 Education Research Team [Education and Computer Science] 22

2020 doi
[79]

Smith, R., Brown, K., & Garcia, J. (2023). Quantum algorithms for natural language processing. Quantum Information Processing, 22(7), 1–

2023
[80]

https://doi.org/10.1007/s11128-023-04032-y

doi
[84]

Strubell, E., Ganesh, A., & McCallum, A. (2019). Energy and policy considerations for deep learning in NLP. Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, 3645–

2019
[87]

https://doi.org/10.1109/JPROC.2023.3242345

2023
[93]

& Zettlemoyer, L

Zhang, S., Roller, S., Goyal, N., Artetxe, M., Chen, M., Chen, S., ... & Zettlemoyer, L. (2022). OPT: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068. https://doi.org/10.48550/arXiv.2205.01068

2022 doi arXiv
[210]

https://doi.org/10.1561/2200000083

doi
[293]

https://doi.org/10.1016/j.exis.2020.11.018

2020 doi
[463]

https://doi.org/10.1016/j.jclepro.2017.12.239

2017 doi
[570]

https://doi.org/10.1038/s42256-023-00671-z

doi
[623]

https://doi.org/10.1145/3442188.3445922
[1240]

https://doi.org/10.1093/bioinformatics/btz682

doi
[1894]

https://doi.org/10.48550/arXiv.2211.02001

doi arXiv
[1901]

https://doi.org/10.48550/arXiv.2005.14165

2005 doi arXiv
[2020]

https://doi.org/10.2873/392949

United Nations University. https://doi.org/10.2873/392949

doi
[2030]

Engineering Applications of Artificial Intelligence, 94, 103790
[3650]

https://doi.org/10.18653/v1/P19-1355

doi
[4174]

https://doi.org/10.18653/v1/2020.findings-emnlp.372

2020 doi
[4631]

https://doi.org/10.1021/acs.est.1c06325

doi

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

Ayana, G. et al. (2024). Decolonizing global AI governance: assessment of the state of decolonized AI governance in Sub-Saharan Africa. R. Soc. Open Sci. 11: 231994. https://doi.org/10.1098/rsos.231994

work page doi:10.1098/rsos.231994 2024

[2] [2]

AI Now Institute. (2023). The environmental costs of AI: A case study of GPT-4

2023

[3] [3]

Amazon SEC Filing. (2023). Definitive Proxy Statement. Retrieved from https://www.sec.gov

2023

[4] [4]

Amazon Sustainability Report. (2022). Amazon.com. Retrieved from https://sustainability.aboutamazon.com

2022

[5] [5]

Andrae, A. S. G. (2020). New perspectives on internet electricity use in

2020

[6] [6]

Andrae, A. S. G. (2023). Hyperscale data centers and AI’s energy paradox. Joule, 7(5), 1–15

2023

[7] [7]

W., Althaf, S., & Cruz Rios, F

Babbitt, C. W., Althaf, S., & Cruz Rios, F. (2021). The role of consumerism in electronic waste. Nature Electronics, 4(3), 152–160

2021

[8] [8]

Belkhir, L., & Elmeligi, A. (2018). Assessing ICT global emissions footprint: Trends to 2040 & recommendations. Journal of Cleaner Production, 177, 448–

2018