AI Safety Tracker

Environmental impact of AI

7 developers 9 published footprints 4 operators

What 7 frontier AI developers have published about the energy, emissions, sites and power sources behind their models, against Epoch AI's estimates of their training compute.

What are the environmental impacts of AI?

Models at or above 1e25 FLOP with a published training-emissions figure Joined by this site: Epoch AI's compute estimates against this site's index of published footprints. A lifecycle figure covering months of use is not counted as a training figure. Environmental Disclosure Index This site Licence MIT Updated quarterly Retrieved , 5 days ago Data 7 datasets Our own reading of what frontier developers and the operators of their data centres have published about training energy, emissions, power sources and energy per prompt, with permits and press reporting marked separately from company statements. Most cells record an absence. Company figures are self-reported and mostly unaudited, and market-based emissions reflect contracts rather than the grid at the hour of use. Cite AI Safety Tracker, "Environmental Disclosure Index". https://aisafetytracker.org/environment/. Retrieved 23 October 2026. Notable AI Models Epoch AI Licence CC-BY-4.0 Updated weekly Retrieved , 5 days ago Inclusion is editorial ("notability criteria"), not exhaustive. Compute figures are frequently estimates; check the per-row confidence and notes columns before treating any value as measured. Cite Epoch AI, "Notable AI Models". https://epoch.ai/data/ai-models. Retrieved 23 October 2026 via AI Safety Tracker, https://aisafetytracker.org/environment/.

1 of 24

Llama 3.1-405B. No figure is published for any model by OpenAI, Anthropic or xAI, or for any Gemini model.

Google data-centre electricity that was carbon-free in the hour it was used, 2025 Environmental Disclosure Index This site Licence MIT Updated quarterly Retrieved , 5 days ago Data 7 datasets Our own reading of what frontier developers and the operators of their data centres have published about training energy, emissions, power sources and energy per prompt, with permits and press reporting marked separately from company statements. Most cells record an absence. Company figures are self-reported and mostly unaudited, and market-based emissions reflect contracts rather than the grid at the hour of use. Cite AI Safety Tracker, "Environmental Disclosure Index". https://aisafetytracker.org/environment/. Retrieved 23 October 2026.

65%

Against 100% matched over the year by renewable purchases. No other operator publishes the hourly figure.

Google's figure for the median text prompt in Gemini Apps Environmental Disclosure Index This site Licence MIT Updated quarterly Retrieved , 5 days ago Data 7 datasets Our own reading of what frontier developers and the operators of their data centres have published about training energy, emissions, power sources and energy per prompt, with permits and press reporting marked separately from company statements. Most cells record an absence. Company figures are self-reported and mostly unaudited, and market-based emissions reflect contracts rather than the grid at the hour of use. Cite AI Safety Tracker, "Environmental Disclosure Index". https://aisafetytracker.org/environment/. Retrieved 23 October 2026.

0.24Wh

May 2025, with a published method. OpenAI states 0.34 Wh for an average query, without one.

Google's electricity emissions at the grid's average intensity, 2025 Location-based and market-based scope 2 emissions, from Google's 2026 Environmental Report. Environmental Disclosure Index This site Licence MIT Updated quarterly Retrieved , 5 days ago Data 7 datasets Our own reading of what frontier developers and the operators of their data centres have published about training energy, emissions, power sources and energy per prompt, with permits and press reporting marked separately from company statements. Most cells record an absence. Company figures are self-reported and mostly unaudited, and market-based emissions reflect contracts rather than the grid at the hour of use. Cite AI Safety Tracker, "Environmental Disclosure Index". https://aisafetytracker.org/environment/. Retrieved 23 October 2026.

15.1MtCO2e

2.8 MtCO2e once its renewable contracts are subtracted. Data centres drew 42.4 TWh, up from 30.6 TWh the year before.

Largest reported training run Compute is Epoch AI's estimate, not a disclosure. Notable AI Models Epoch AI Licence CC-BY-4.0 Updated weekly Retrieved , today Inclusion is editorial ("notability criteria"), not exhaustive. Compute figures are frequently estimates; check the per-row confidence and notes columns before treating any value as measured. Cite Epoch AI, "Notable AI Models". https://epoch.ai/data/ai-models. Retrieved 28 October 2026 via AI Safety Tracker, https://aisafetytracker.org/environment/.

9.4e26FLOP

GPT-6 Astra, OpenAI, 2026-09. No energy use, site or emissions figure is published for it.

What we know

What each developer has published about the energy behind its models, and what permits and press reporting add. 12 of 28 entries record nothing published.

  • OpenAI

    largest run 9.4e26 FLOP, GPT-6 Astra
    Training energy or emissions Not disclosed
    No energy or emissions figure for any model.
    Where it trains and serves Published
    Stargate sites built with Oracle and SoftBank. The first, at Abilene, Texas, runs on Oracle Cloud Infrastructure. OpenAI
    What powers those sites Reported by others
    Permits filed for on-site natural-gas turbines at the Abilene campus. OpenAI publishes no energy figure for any site. DCD
    Energy per prompt Published
    0.34 Wh for an average ChatGPT query, June 2025, stated without a method. Sam Altman

  • xAI

    largest run 5.0e26 FLOP, Grok 4
    Training energy or emissions Not disclosed
    No energy or emissions figure for any model.
    Where it trains and serves Reported by others
    Grok 3 was trained on Colossus, a converted factory in Memphis, Tennessee. R&D World
    What powers those sites Reported by others
    Colossus 1: 15 on-site gas turbines, 247 MW, permitted in July 2025. Around 35 ran before the permit was granted. Colossus 2: dozens of 35 MW gas turbines across the state line in Southaven, Mississippi, permitted in March 2026. DCDWDAM
    Energy per prompt Not disclosed
    No per-prompt figure.

  • Meta

    largest run 5.2e25 FLOP, Llama 4 Behemoth (preview)
    Training energy or emissions Published
    GPU hours and emissions in every Llama model card since Llama 2. Llama 3.1 405B: 8,930 tCO2e location-based. Llama 3.1 model card
    Where it trains and serves Partly published
    Meta's own GPU clusters. Which site trained a given model is not stated. Llama 3.1 model card
    What powers those sites Published
    100% of electricity matched with renewable energy since 2020, so market-based training emissions are reported as zero. Llama 3.1 model card
    Energy per prompt Not disclosed
    No per-prompt figure.

  • Google DeepMind

    largest run 5.0e25 FLOP, Gemini 1.0 Ultra
    Training energy or emissions Not disclosed
    No energy or emissions figure for any Gemini model. The Gemma 3 model card no longer carries the pre-training figure it once gave.
    Where it trains and serves Partly published
    Google's own data centres. Which sites train or serve a given model is not stated. Google
    What powers those sites Published
    65% carbon-free energy hour by hour across Google's data centres in 2025, and 100% matched with renewable purchases over the year. Google
    Energy per prompt Published
    0.24 Wh, 0.03 gCO2e and 0.26 mL of water for the median Gemini Apps text prompt, May 2025, with a published method. Google

  • Anthropic

    largest run 3.4e25 FLOP, Claude 3.7 Sonnet
    Training energy or emissions Not disclosed
    No energy or emissions figure for any model.
    Where it trains and serves Published
    Amazon's Trainium clusters including Project Rainier, Google TPUs, and from May 2026 all of SpaceX's Colossus 1 in Memphis, over 300 MW. AmazonGoogleSpaceX
    What powers those sites Reported by others
    Colossus 1 holds a permit for 15 on-site gas turbines. Amazon and Google state 100% annual renewable matching for their own consumption. Anthropic publishes no energy figure. DCDAmazonGoogle
    Energy per prompt Not disclosed
    No per-prompt figure.

  • DeepSeek

    largest run 9.7e24 FLOP, DeepSeek-V4-Pro
    Training energy or emissions Partly published
    GPU hours only: 2.788 million H800 GPU hours for DeepSeek-V3. No energy or emissions figure. DeepSeek-V3 report
    Where it trains and serves Not disclosed
    Not stated.
    What powers those sites Not disclosed
    Not stated.
    Energy per prompt Not disclosed
    No per-prompt figure.

  • Mistral AI

    largest run 7.7e23 FLOP, Mixtral 8x7B
    Training energy or emissions Published
    Lifecycle analysis of Mistral Large 2: 20.4 ktCO2e and 281,000 m3 of water for training and 18 months of use, to January 2025. Mistral AI
    Where it trains and serves Not disclosed
    Not stated in the lifecycle analysis.
    What powers those sites Not disclosed
    Not stated in the lifecycle analysis.
    Energy per prompt Published
    1.14 gCO2e and 45 mL of water for a 400-token Le Chat response. Mistral AI

"Published" means the developer's own model card, report or statement. "Reported by others" means a permit, a regulator or the press, with the developer silent. A company's clean-energy claims cover its whole operation, never one model. Largest runs are Epoch AI's compute estimates for models published since 2020. Our own reading of each developer's publications. Every claim links to where it was made. Environmental Disclosure Index This site Licence MIT Updated quarterly Retrieved , 5 days ago Data 7 datasets Our own reading of what frontier developers and the operators of their data centres have published about training energy, emissions, power sources and energy per prompt, with permits and press reporting marked separately from company statements. Most cells record an absence. Company figures are self-reported and mostly unaudited, and market-based emissions reflect contracts rather than the grid at the hour of use. Cite AI Safety Tracker, "Environmental Disclosure Index". https://aisafetytracker.org/environment/. Retrieved 23 October 2026. Notable AI Models Epoch AI Licence CC-BY-4.0 Updated weekly Retrieved , 5 days ago Inclusion is editorial ("notability criteria"), not exhaustive. Compute figures are frequently estimates; check the per-row confidence and notes columns before treating any value as measured. Cite Epoch AI, "Notable AI Models". https://epoch.ai/data/ai-models. Retrieved 23 October 2026 via AI Safety Tracker, https://aisafetytracker.org/environment/.

How big the training runs are

Every run since 2020 above 1.0e21 FLOP Floating-point operation. Training compute is reported as the total number of operations used in the final training run, on a logarithmic scale, so each axis step is ten times the one below it. Glossary entry with a compute estimate, by developer, on a logarithmic scale. The largest per developer is named.

1e211e221e231e241e251e261e271e25, EU thresholdOpenAIGPT-6 Astra (OpenAI, 2026-09): 9.4e26 FLOP, confidence Likely. No training emissions figure published.GPT-4.5 (OpenAI, 2025-02): 3.8e26 FLOP, confidence Likely. No training emissions figure published.GPT-5 (OpenAI, 2025-08): 6.6e25 FLOP, confidence Speculative. No training emissions figure published.GPT-4 (Mar 2023) (OpenAI, 2023-03): 2.1e25 FLOP, confidence Likely. No training emissions figure published.GPT-4 (Jun 2023) (OpenAI, 2023-06): 2.1e25 FLOP, confidence Likely. No training emissions figure published.gpt-oss-120b (OpenAI, 2025-08): 4.9e24 FLOP, confidence Confident. No training emissions figure published.GPT-3.5 (davinci-002) (OpenAI, 2022-03): 2.6e24 FLOP, confidence Speculative. No training emissions figure published.gpt-oss-20b (OpenAI, 2025-08): 5.5e23 FLOP, confidence Confident. No training emissions figure published.DALL·E 2 (OpenAI, 2022-04): 3.4e23 FLOP, confidence Speculative. No training emissions figure published.InstructGPT 175B (OpenAI, 2022-01): 3.2e23 FLOP, confidence Confident. No training emissions figure published.GPT-3 175B (davinci) (OpenAI, 2020-05): 3.1e23 FLOP, confidence Confident. Training emissions published.Codex (OpenAI, 2021-07): 7.3e22 FLOP, confidence Likely. No training emissions figure published.DALL-E (OpenAI, 2021-01): 4.7e22 FLOP, confidence Likely. No training emissions figure published.Statement Curriculum Learning (OpenAI, 2022-03): 1.8e22 FLOP, confidence Likely. No training emissions figure published.CLIP (ViT L/14@336px) (OpenAI, 2021-01): 1.0e22 FLOP, confidence Confident. No training emissions figure published.ADM (OpenAI, 2021-05): 6.2e21 FLOP, confidence Confident. No training emissions figure published.Whisper (OpenAI, 2022-09): 4.2e21 FLOP, confidence Likely. No training emissions figure published.GPT-6 AstraxAIGrok 4 (xAI, 2025-07): 5.0e26 FLOP, confidence Speculative. No training emissions figure published.Grok 3 (xAI, 2025-02): 3.5e26 FLOP, confidence Likely. No training emissions figure published.Grok-2 (xAI, 2024-08): 3.0e25 FLOP, confidence Confident. No training emissions figure published.Grok-1 (xAI, 2023-11): 2.9e24 FLOP, confidence Likely. No training emissions figure published.Grok 4MetaLlama 4 Behemoth (preview) (Meta, 2025-04): 5.2e25 FLOP, confidence Likely. No training emissions figure published.Llama 3.1-405B (Meta, 2024-07): 3.8e25 FLOP, confidence Confident. Training emissions published.Llama 3-70B (Meta, 2024-04): 7.9e24 FLOP, confidence Confident. Training emissions published.Llama 3.3 70B (Meta, 2024-12): 6.9e24 FLOP, confidence Confident. No training emissions figure published.Llama 4 Scout (Meta, 2025-04): 4.1e24 FLOP, confidence Likely. Training emissions published.Llama 4 Maverick (Meta, 2025-04): 2.2e24 FLOP, confidence Likely. Training emissions published.Movie Gen Video (Meta, 2024-10): 1.6e24 FLOP, confidence Confident. No training emissions figure published.Llama 2-70B (Meta, 2023-07): 8.1e23 FLOP, confidence Confident. Training emissions published.Llama 3.2 11B (Meta, 2024-09): 5.8e23 FLOP, confidence Confident. No training emissions figure published.LLaMA-65B (Meta, 2023-02): 5.5e23 FLOP, confidence Confident. No training emissions figure published.OPT-175B (Meta, 2022-05): 4.3e23 FLOP, confidence Confident. No training emissions figure published.Galactica (Meta, 2022-11): 3.2e23 FLOP, confidence Likely. No training emissions figure published.Llama Guard (Meta, 2023-12): 1.6e23 FLOP, confidence Confident. No training emissions figure published.Llama 2-7B (Meta, 2023-07): 8.4e22 FLOP, confidence Confident. No training emissions figure published.NLLB (Meta, 2022-07): 1.8e22 FLOP, confidence Confident. No training emissions figure published.Segment Anything Model (Meta, 2023-04): 7.8e21 FLOP, confidence Confident. No training emissions figure published.Make-A-Scene (Meta, 2022-03): 6.4e21 FLOP, confidence Likely. No training emissions figure published.Llama 4 Behemoth (preview)Google DeepMindGemini 1.0 Ultra (Google DeepMind, 2023-12): 5.0e25 FLOP, confidence Speculative. No training emissions figure published.PaLM 2 (Google DeepMind, 2023-05): 7.3e24 FLOP, confidence Likely. No training emissions figure published.Minerva (540B) (Google DeepMind, 2022-06): 2.7e24 FLOP, confidence Confident. No training emissions figure published.Flan-PaLM 540B (Google DeepMind, 2022-10): 2.5e24 FLOP, confidence Confident. No training emissions figure published.U-PaLM (540B) (Google DeepMind, 2022-10): 2.5e24 FLOP, confidence Confident. No training emissions figure published.Gopher (280B) (Google DeepMind, 2021-12): 6.3e23 FLOP, confidence Confident. No training emissions figure published.Chinchilla (Google DeepMind, 2022-03): 5.8e23 FLOP, confidence Confident. No training emissions figure published.FunSearch (Google DeepMind, 2023-12): 3.9e23 FLOP, confidence Speculative. No training emissions figure published.GLaM (Google DeepMind, 2021-12): 3.6e23 FLOP, confidence Confident. No training emissions figure published.LaMDA (Google DeepMind, 2022-02): 3.6e23 FLOP, confidence Confident. No training emissions figure published.AlphaCode (Google DeepMind, 2022-02): 2.4e23 FLOP, confidence Confident. No training emissions figure published.Flamingo (Google DeepMind, 2022-04): 2.2e23 FLOP, confidence Confident. No training emissions figure published.ViT-22B (Google DeepMind, 2023-02): 1.9e23 FLOP, confidence Confident. No training emissions figure published.PaLI (Google DeepMind, 2022-09): 1.7e23 FLOP, confidence Likely. No training emissions figure published.Switch (Google DeepMind, 2021-01): 8.2e22 FLOP, confidence Confident. No training emissions figure published.GShard (dense) (Google DeepMind, 2020-06): 4.8e22 FLOP, confidence Confident. No training emissions figure published.BASIC-L (Google DeepMind, 2021-11): 4.1e22 FLOP, confidence Likely. No training emissions figure published.Student of Games (Google DeepMind, 2021-12): 3.7e22 FLOP, confidence Speculative. No training emissions figure published.GOAT (Google DeepMind, 2021-07): 2.4e22 FLOP, confidence Speculative. No training emissions figure published.GraphCast (Google DeepMind, 2023-11): 2.1e22 FLOP, confidence Speculative. No training emissions figure published.RETRO-7B (Google DeepMind, 2022-02): 1.7e22 FLOP, confidence Confident. No training emissions figure published.FGN (Google DeepMind, 2025-06): 9.6e21 FLOP, confidence Confident. No training emissions figure published.ContextNet + Noisy Student (Google DeepMind, 2020-01): 8.2e21 FLOP, confidence Confident. No training emissions figure published.SigLIP 400M (Google DeepMind, 2023-03): 4.9e21 FLOP, confidence Confident. No training emissions figure published.Gato (Google DeepMind, 2022-05): 4.0e21 FLOP, confidence Confident. No training emissions figure published.AlphaFold 2 (Google DeepMind, 2020-11): 3.0e21 FLOP, confidence Likely. No training emissions figure published.Gemini 1.0 UltraAnthropicClaude 3.7 Sonnet (Anthropic, 2025-02): 3.4e25 FLOP, confidence Likely. No training emissions figure published.Claude 3.5 Sonnet (Anthropic, 2024-06): 2.7e25 FLOP, confidence Speculative. No training emissions figure published.Claude 2 (Anthropic, 2023-07): 3.9e24 FLOP, confidence Speculative. No training emissions figure published.Claude 3.7 SonnetDeepSeekDeepSeek-V4-Pro (DeepSeek, 2026-04): 9.7e24 FLOP, confidence Likely. No training emissions figure published.DeepSeek V4.1 Flash (DeepSeek, 2026-09): 4.3e24 FLOP, confidence Likely. No training emissions figure published.DeepSeek-R1 (May 2025) (DeepSeek, 2025-05): 4.0e24 FLOP, confidence Confident. No training emissions figure published.DeepSeek-R1 (DeepSeek, 2025-01): 3.5e24 FLOP, confidence Confident. No training emissions figure published.DeepSeek-V3 (DeepSeek, 2024-12): 3.3e24 FLOP, confidence Confident. No training emissions figure published.DeepSeek-V3 (Mar 2025) (DeepSeek, 2025-03): 3.3e24 FLOP, confidence Confident. No training emissions figure published.DeepSeek-V4-Flash (DeepSeek, 2026-04): 2.5e24 FLOP, confidence Likely. No training emissions figure published.DeepSeek-V2.5 (DeepSeek, 2024-09): 1.8e24 FLOP, confidence Confident. No training emissions figure published.DeepSeek-Coder-V2 236B (DeepSeek, 2024-06): 1.3e24 FLOP, confidence Confident. No training emissions figure published.DeepSeek-V4-ProMistral AIMixtral 8x7B (Mistral AI, 2023-12): 7.7e23 FLOP, confidence Speculative. No training emissions figure published.Mixtral 8x7BTraining compute, FLOP
Largest run per developer Other runs Ringed: emissions figure published Compute is Epoch AI's estimate. Whether emissions are published is this site's index. Notable AI Models Epoch AI Licence CC-BY-4.0 Updated weekly Retrieved , today Inclusion is editorial ("notability criteria"), not exhaustive. Compute figures are frequently estimates; check the per-row confidence and notes columns before treating any value as measured. Cite Epoch AI, "Notable AI Models". https://epoch.ai/data/ai-models. Retrieved 28 October 2026 via AI Safety Tracker, https://aisafetytracker.org/environment/. Environmental Disclosure Index This site Licence MIT Updated quarterly Retrieved , today Data 7 datasets Our own reading of what frontier developers and the operators of their data centres have published about training energy, emissions, power sources and energy per prompt, with permits and press reporting marked separately from company statements. Most cells record an absence. Company figures are self-reported and mostly unaudited, and market-based emissions reflect contracts rather than the grid at the hour of use. Cite AI Safety Tracker, "Environmental Disclosure Index". https://aisafetytracker.org/environment/. Retrieved 28 October 2026.
Use this figure

An SVG image of this figure, with its sources, date and address printed in it, linked to this section. SVG file. All embeddable figures. 

<a href="https://aisafetytracker.org/environment/#training-runs"><img src="https://aisafetytracker.org/embed/training-runs.svg" alt="Strip chart of frontier AI developers&#39; training runs by compute, with the runs that have a published training-emissions figure ringed, from AI Safety Tracker" width="1048" height="538" loading="lazy" style="max-width:100%;height:auto"></a>
Model Developer Compute, FLOP Published Training emissions
GPT-6 Astra OpenAI 9.4e26 2026-09 Not published
Grok 4 xAI 5.0e26 2025-07 Not published
GPT-4.5 OpenAI 3.8e26 2025-02 Not published
Grok 3 xAI 3.5e26 2025-02 Not published
GPT-5 OpenAI 6.6e25 2025-08 Not published
Llama 4 Behemoth (preview) Meta AI 5.2e25 2025-04 Not published
Gemini 1.0 Ultra Google DeepMind 5.0e25 2023-12 Not published
Composer 2.5 Cursor 3.9e25 2026-05 Not published
Llama 3.1-405B Meta AI 3.8e25 2024-07 Published
Claude 3.7 Sonnet Anthropic 3.4e25 2025-02 Not published

Compute is the nearest measured quantity to training energy, and not a measure of emissions. The energy a run needs is its compute divided by the efficiency of the chips it ran on, which improves with each hardware generation, and the emissions are that energy times the carbon intensity of the grid behind the data centre, which no developer on this chart states for its largest run. Compute figures are mostly Epoch AI's estimates, faintest where developers publish least, and cover the final training run only.

Training footprints that have been published

Every published figure for the energy or emissions of training a model, with who measured it and on what basis. Emissions per 1e24 FLOP are computed here where both figures exist.

Model Published by Energy, MWh tCO2e Basis Market-based tCO2e per 1e24 FLOP
GPT-3 175B third-party estimate 1,287 552 location-based - 1,758
BLOOM 176B developer study 433 24.7 location-based - 68
Llama 2 70B developer model card 688* 291 location-based - 360
DeepSeek-V3 developer technical report - - 2.79M GPU hours only - -
Llama 3 70B developer model card 4,480* 1,900 location-based - 242
Llama 3.1 405B developer model card 21,588* 8,930 location-based 0 235
Mistral Large 2 developer lifecycle analysis - 20,400 lifecycle - -
Llama 4 Maverick developer model card 1,666* 645 location-based 0 287
Llama 4 Scout developer model card 3,500* 1,354 location-based 0 332
What each figure covers 9
  • GPT-3 175B Final training run, estimated by Google researchers from published details. OpenAI has published no figure.
  • BLOOM 176B Final training run, dynamic power only, on the Jean Zay supercomputer in France. 50.5 tCO2e including idle power and hardware manufacture.
  • Llama 2 70B Pre-training. The card states the emissions were fully offset.
  • DeepSeek-V3 GPU hours only. No power, energy, location or emissions figure.
  • Llama 3 70B Pre-training. The card states the emissions were fully offset.
  • Llama 3.1 405B Training. The largest training run with a published emissions figure.
  • Mistral Large 2 Training and 18 months of use to January 2025 together, on a lifecycle basis, so not comparable with a training-only figure.
  • Llama 4 Maverick Pre-training.
  • Llama 4 Scout Pre-training. No figure is published for Llama 4 Behemoth, the largest Llama 4 run.

Location-based emissions Emissions from electricity counted at the average carbon intensity of the grid a data centre drew from, whatever contracts the operator holds. Glossary entry use the average carbon intensity of the grid the data centre drew from; Market-based emissions Emissions from electricity after subtracting renewable energy the operator bought by contract. A company matching all its annual use with renewable purchases can report close to zero while its data centres draw fossil power at some hours. Glossary entry subtract renewable energy bought by contract, which is how Llama 3.1 405B is reported at 8,930 tCO2e on one basis and zero on the other. Energy marked * is derived here from the model card's GPU hours and stated power per chip. Per unit of compute, the published figures differ 26-fold, from the French grid under BLOOM to the older chips and grid under GPT-3, whose figure is an outside estimate. Compute for each model is Epoch AI's estimate. Emissions per 1e24 FLOP computed by this site from the published emissions figure and Epoch AI's compute estimate for the same model. Environmental Disclosure Index This site Licence MIT Updated quarterly Retrieved , 5 days ago Data 7 datasets Our own reading of what frontier developers and the operators of their data centres have published about training energy, emissions, power sources and energy per prompt, with permits and press reporting marked separately from company statements. Most cells record an absence. Company figures are self-reported and mostly unaudited, and market-based emissions reflect contracts rather than the grid at the hour of use. Cite AI Safety Tracker, "Environmental Disclosure Index". https://aisafetytracker.org/environment/. Retrieved 23 October 2026. Notable AI Models Epoch AI Licence CC-BY-4.0 Updated weekly Retrieved , 5 days ago Inclusion is editorial ("notability criteria"), not exhaustive. Compute figures are frequently estimates; check the per-row confidence and notes columns before treating any value as measured. Cite Epoch AI, "Notable AI Models". https://epoch.ai/data/ai-models. Retrieved 23 October 2026 via AI Safety Tracker, https://aisafetytracker.org/environment/.

Energy per prompt

Every per-prompt figure a developer has published. Each measures a different statistic.

  • 0.24 Wh Google, median text prompt, Gemini Apps, May 2025. 0.03 gCO2e. 0.26 mL of water. Method published. Source
  • 0.34 Wh OpenAI, average query, ChatGPT, June 2025. Method not published. Source
  • 1.14 gCO2e Mistral AI, 400-token response, Le Chat, January 2025. 45 mL of water. Method published. Source

A median prompt and an average query are different statistics: a few very long requests raise an average and leave a median where it was. None of the three covers image or video generation, long reasoning or agent runs, and Google's emissions figure uses its market-based fleet average. Anthropic, Meta, xAI and DeepSeek publish no per-prompt figure. Environmental Disclosure Index This site Licence MIT Updated quarterly Retrieved , 5 days ago Data 7 datasets Our own reading of what frontier developers and the operators of their data centres have published about training energy, emissions, power sources and energy per prompt, with permits and press reporting marked separately from company statements. Most cells record an absence. Company figures are self-reported and mostly unaudited, and market-based emissions reflect contracts rather than the grid at the hour of use. Cite AI Safety Tracker, "Environmental Disclosure Index". https://aisafetytracker.org/environment/. Retrieved 23 October 2026.

Renewable claims against the grid

What the operators of frontier data centres report for their own electricity. Renewable matching Buying as much renewable electricity over a year as a company uses, through contracts or certificates. Says nothing about whether the power used in a given hour, on a given grid, was clean. Glossary entry counts renewable energy bought over a year; Hourly carbon-free energy The share of electricity that was carbon-free in the same hour and on the same grid in which it was used. Google is the only data-centre operator that publishes it. Glossary entry counts what was clean in the hour and on the grid where it was used.

Operator Period Data centres, TWh Matched over the year Carbon-free by the hour Note Water and cooling
Google 2025 42.4 100% 65% Data-centre electricity rose 38% on 2024. The hourly carbon-free share was 13% across Google's Asia Pacific grid regions and 68% in the United States. Consumed 41 billion litres of water across its data centres and offices in 2025, up 34% on 2024. States that cooling with water needs less energy than cooling with air. Source
Microsoft FY2025 - 100% not published Matched 100% of annual electricity with renewable energy in fiscal 2025. No hourly or per-region figure is published. Every new datacentre design since August 2024 cools chips in a closed loop that evaporates no water once filled, at a stated nominal increase in electricity use. The first such sites are expected online in late 2027. The existing fleet mixes air and water cooling. Source
Amazon 2025 - 100% not published Matched 100% of electricity with renewable energy for a third year. Runs AWS, where Anthropic trains and serves on Trainium chips. not recorded
Meta since 2020 - 100% not published Stated in its Llama model cards: net zero operations and 100% renewable matching since 2020. not recorded
Google, carbon-free share of data-centre electricity by grid region, hour by hour, 2025 Google CFE counts Google's clean-energy contracts in that grid region; grid CFE is the region's own carbon-free share. Both from Google's 2026 Environmental Report. Environmental Disclosure Index This site Licence MIT Updated quarterly Retrieved , 5 days ago Data 7 datasets Our own reading of what frontier developers and the operators of their data centres have published about training energy, emissions, power sources and energy per prompt, with permits and press reporting marked separately from company statements. Most cells record an absence. Company figures are self-reported and mostly unaudited, and market-based emissions reflect contracts rather than the grid at the hour of use. Cite AI Safety Tracker, "Environmental Disclosure Index". https://aisafetytracker.org/environment/. Retrieved 23 October 2026. lowest first
Grid region Google Whole grid
Singapore, Energy Market Authority 5% 5%
Taiwan, Taiwan Power Company 15% 15%
Japan, TEPCO 23% 18%
US, South Carolina 31% 25%
US, Southern Company 42% 33%
US, PJM 57% 40%
US, Tennessee Valley Authority 58% 47%
Ireland, EirGrid 60% 50%
US, Duke Energy Carolinas 65% 57%
US, NV Energy 65% 32%
Show the remaining 12
Germany, Germany 70% 64%
United Kingdom, National Grid ESO 75% 71%
Belgium, Elia 79% 72%
Netherlands, TenneT 82% 55%
US, Bonneville Power Administration 83% 84%
US, ERCOT, Texas 83% 46%
US, Southwest Power Pool 84% 47%
US, Salt River Project, Arizona 86% 56%
US, MISO 88% 36%
Chile, Sistema Interconectado Central 90% 62%
Denmark, Energinet 92% 89%
Finland, Fingrid 98% 95%

Every operator here matches all of its annual electricity with renewable purchases. Only Google publishes how much was carbon-free in the hour of use: 65% across its data centres, and 5% on the Energy Market Authority grid in Singapore. The remainder came from fossil-fuelled generation on the grid at that hour. No operator publishes which grid region serves which model, so the regional figures describe where Google computes, not where any one model runs. Environmental Disclosure Index This site Licence MIT Updated quarterly Retrieved , 5 days ago Data 7 datasets Our own reading of what frontier developers and the operators of their data centres have published about training energy, emissions, power sources and energy per prompt, with permits and press reporting marked separately from company statements. Most cells record an absence. Company figures are self-reported and mostly unaudited, and market-based emissions reflect contracts rather than the grid at the hour of use. Cite AI Safety Tracker, "Environmental Disclosure Index". https://aisafetytracker.org/environment/. Retrieved 23 October 2026.

Datacentre electricity to 2035

Electricity used by all datacentres worldwide: the International Energy Agency's estimate for 2024 and 2025, and its Base Case for 2030 and 2035. The 2035 range spans its four cases. The AI portion is marked where the agency states one.

04008001,2001,600202420252026202720282029203020312032203320342035Electricity used by datacentres, TWh a year1,700, top of range700, foot of range2024: 415 TWh, estimated by the IEA2025: 485 TWh, estimated by the IEA2030: about 945 TWh, IEA Base Case2035: about 1,200 TWh, IEA Base Case4154859451,200, Base Case2024: about 62 TWh used by accelerated servers, 15% of the total2030: about 465 TWh at AI-focused datacentres, IEA Base Case62, accelerated servers465, AI-focused datacentres
All datacentres, estimated All datacentres, IEA Base Case Range across the four cases AI portion, defined differently each year Every figure is the IEA's, read from Energy and AI (2025) and Key Questions on Energy and AI (2026). The dashed line joins the years the IEA states. Nothing is this site's forecast. Environmental Disclosure Index This site Licence MIT Updated quarterly Retrieved , 4 days ago Data 7 datasets Our own reading of what frontier developers and the operators of their data centres have published about training energy, emissions, power sources and energy per prompt, with permits and press reporting marked separately from company statements. Most cells record an absence. Company figures are self-reported and mostly unaudited, and market-based emissions reflect contracts rather than the grid at the hour of use. Cite AI Safety Tracker, "Environmental Disclosure Index". https://aisafetytracker.org/environment/. Retrieved 24 October 2026.

The projections are the IEA's scenarios, and the dashed line joins the years it states without showing a path it publishes between them. The 2024 AI mark is the share of electricity used by accelerated servers, which the IEA uses as a proxy and calls imperfect. The 2030 mark is its AI-focused datacentres in the Base Case, and the 2025 level of that group is not stated in terawatt hours. The two marks are defined differently and are not joined. The 2035 range runs from 700 to 1,700 TWh across the IEA's four cases, and its text does not say which case sets which end. The April 2026 report restates the 2030 total as around 950 TWh.

Datacentre power by country

AI datacentre capacity by country in Epoch AI's hub, and every datacentre's electricity by region in the IEA's estimate for 2024. The two measure different things.

AI datacentres, IT power by country Epoch AI's AI Data Centers hub: power operating on the retrieval date and planned at the last stage of each site's timeline, summed by the country Epoch records. AI Data Centers Epoch AI Licence CC-BY-4.0 Updated several times a week Retrieved , today The largest AI sites Epoch has found, mostly in the United States. Epoch estimated in September 2026 that they hold 44% of the world's deployed AI compute, with a 90% interval of 23% to 81%. Power, compute and cost are Epoch's estimates from satellite imagery, permits and company statements; four in five IT power estimates fall within a factor of 1.4 of the true value. The owner is the owner of the chips, not necessarily of the building. Capacity after the date of retrieval is Epoch's projection. Coordinates are read from the hub's map page, which publishes them; the downloads do not. Cite Epoch AI, "AI Data Centers". https://epoch.ai/data/ai-data-centers. Retrieved 28 October 2026 via AI Safety Tracker, https://aisafetytracker.org/environment/. United States 33.1 GW of 38.2 GW planned
Country Planned Operating now
United States 77 sites 33.1 GW 12 GW
China 5 sites 2.1 GW 731 MW
Malaysia 3 sites 1.1 GW 1.1 GW
United Arab Emirates 1 site 1 GW 0 MW
Portugal 1 site 233 MW 33 MW
Finland 1 site 145 MW 75 MW
Norway 2 sites 134 MW 42 MW
United Kingdom 1 site 88 MW 88 MW
Indonesia 1 site 72 MW 72 MW
Sweden 1 site 48 MW 0 MW
Show the remaining 2
Australia 1 site 41 MW 41 MW
Iceland 1 site 10.2 MW 5.1 MW
All datacentres, electricity by region, 2024 The shares are the IEA's, from Energy and AI (2025). The terawatt hours are those shares of the IEA's 2024 world total, multiplied here: the IEA states no regional figure in terawatt hours. Environmental Disclosure Index This site Licence MIT Updated quarterly Retrieved , 4 days ago Data 7 datasets Our own reading of what frontier developers and the operators of their data centres have published about training energy, emissions, power sources and energy per prompt, with permits and press reporting marked separately from company statements. Most cells record an absence. Company figures are self-reported and mostly unaudited, and market-based emissions reflect contracts rather than the grid at the hour of use. Cite AI Safety Tracker, "Environmental Disclosure Index". https://aisafetytracker.org/environment/. Retrieved 24 October 2026. 415 TWh world total
Region Share of world TWh, computed
United States 45% 187
China 25% 104
Europe 15% 62
Rest of the world 15% 62

Power is the capacity a datacentre can draw and electricity is the energy it used, so the tables cannot be converted into each other: neither source states how fully the capacity is used. The hub holds 95 sites and is not a census of datacentres. Europe is the IEA's region and not the European Union. The rest of the world is the remainder after the three regions the IEA names, and includes Japan, for which it gives no share. Lawrence Berkeley National Laboratory states 192 TWh for the United States in 2024, 46% of the IEA's world total. Environmental Disclosure Index This site Licence MIT Updated quarterly Retrieved , 5 days ago Data 7 datasets Our own reading of what frontier developers and the operators of their data centres have published about training energy, emissions, power sources and energy per prompt, with permits and press reporting marked separately from company statements. Most cells record an absence. Company figures are self-reported and mostly unaudited, and market-based emissions reflect contracts rather than the grid at the hour of use. Cite AI Safety Tracker, "Environmental Disclosure Index". https://aisafetytracker.org/environment/. Retrieved 23 October 2026. AI Data Centers Epoch AI Licence CC-BY-4.0 Updated several times a week Retrieved , 5 days ago The largest AI sites Epoch has found, mostly in the United States. Epoch estimated in September 2026 that they hold 44% of the world's deployed AI compute, with a 90% interval of 23% to 81%. Power, compute and cost are Epoch's estimates from satellite imagery, permits and company statements; four in five IT power estimates fall within a factor of 1.4 of the true value. The owner is the owner of the chips, not necessarily of the building. Capacity after the date of retrieval is Epoch's projection. Coordinates are read from the hub's map page, which publishes them; the downloads do not. Cite Epoch AI, "AI Data Centers". https://epoch.ai/data/ai-data-centers. Retrieved 23 October 2026 via AI Safety Tracker, https://aisafetytracker.org/environment/.

What this page cannot show

Questions a reader brings that no published source answers.

  • Energy used by the largest runs

    No developer publishes it for any model above 1e26 FLOP.

  • Which grid serves which model

    Inference moves between regions with demand. No operator publishes where a given model is served from, hour by hour.

  • Water per model

    Published by Google and Mistral AI per prompt, and by Mistral AI for one model's lifecycle.

  • Chip manufacture

    Of the footprints listed here, only BLOOM's study and Mistral AI's lifecycle analysis count the emissions of making the hardware.

  • Chinese developers

    DeepSeek publishes GPU hours and nothing on power, sites or emissions.

  • Tokens behind the energy

    How much the models are used, as each company states it.

    Usage

Related

  • Datacentres Every tracked AI datacentre, its power and its operator.
  • Capability Training compute for every notable model.
  • Token volumes Every aggregate token count a company has stated.