What are the environmental impacts of AI?
The environmental impacts of AI that published figures measure are the electricity drawn by the datacentres that train and run the models, the carbon dioxide emitted in generating that electricity, and the water consumed in cooling those buildings, in producing their power and in making their chips. Most of those figures describe datacentres or whole companies, and few describe AI on its own. Of the 24 models trained with the most compute, those at or above the line at which the EU AI Act presumes a model carries systemic risk, one has a published figure for the emissions of its training. 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 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 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/explainers/environmental-impact-of-ai/.
The answers below use only figures published on the environment and datacentres pages, where each is set out record by record with its source.
How much electricity do datacentres use, and how much of it is AI?
The International Energy Agency (IEA) puts the electricity used by every datacentre in the world at around 415 TWh in 2024, about 1.5% of global electricity demand. It does not count AI separately. Its nearest measure is the electricity drawn by accelerated servers, the machines built around graphics processors or other AI chips, which it puts at 15% of the total and describes as an imperfect proxy for AI. A server is counted by the hardware inside it, whatever work it happens to run. Recorded by this site from the International Energy Agency's Energy and AI (2025), CC BY 4.0. Environmental Disclosure Index This site 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.
For the United States, Lawrence Berkeley National Laboratory (LBNL) puts datacentre electricity at 192 TWh in 2024, 4.7% of the country's electricity, excluding cryptocurrency mining. It states an AI share of that electricity only for 2030, as a projection. For 2024 it counts servers instead: of 30.9 million installed in US datacentres, 4.2 million were accelerated, a share of 13.6% that this site computes from the two figures. Recorded by this site from Lawrence Berkeley National Laboratory's United States Data Center Energy Usage Report: 2025 Update, CC BY 4.0. Environmental Disclosure Index This site 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.
Of the four operators of frontier AI datacentres recorded here, Google alone publishes how much electricity its datacentres drew: 42.4 TWh in 2025, up 38% from 30.6 TWh the year before. The figure covers all of Google's work in those buildings, AI included, and is not broken down by model. From Google's 2026 Environmental Report. Company-wide, not a figure for any model. Environmental Disclosure Index This site 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.
How much carbon does that electricity emit?
The emissions from a datacentre's electricity are the energy it uses multiplied by the carbon intensity of the grid that supplied it, which is the carbon dioxide released for each unit of electricity generated. The same work therefore emits more on a grid that burns coal or gas than on one that runs on hydro, wind, solar or nuclear power. Companies report the result in two ways. 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 intensity of the grid the datacentre 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 the renewable electricity a company has bought by contract. For 2025, Google reports 15.1 million tonnes of CO2 equivalent on the first basis and 2.8 million on the second. Location-based and market-based scope 2 emissions, from Google's 2026 Environmental Report. Environmental Disclosure Index This site 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.
All four operators, Google, Microsoft, Amazon and Meta, state that they match all of their annual electricity with renewable purchases. 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 is counted over a year, so renewable power bought at one time and place is set against electricity used at another. 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 only the clean electricity available in the same hour and on the same grid as the electricity used. Google is the only operator that publishes it: 65% across its datacentres in 2025, with 13% across its Asia Pacific grid regions and 68% in the United States. By grid region it runs from 98% on the Fingrid grid in Finland to 5% on the Energy Market Authority grid in Singapore. The remainder came from fossil-fuelled generation on the grid at that hour. Annual matching from each operator's own report; hourly carbon-free energy from Google's 2026 Environmental Report. Environmental Disclosure Index This site 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.
How much energy does it take to train an AI model?
Training a model is one very large computation, and its size is measured in 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 , the number of arithmetic operations it performs, written in powers of ten: 1e25 is a one followed by 25 zeros. Epoch AI estimates the compute of every notable training run. The largest it records is GPT-6 Astra, published by OpenAI in September 2026, at 9.4e26 FLOP, and no energy, site or emissions figure has been published for it. Compute is Epoch AI's estimate, not a disclosure. Notable AI Models Epoch AI 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/explainers/environmental-impact-of-ai/.
Compute is the nearest measured quantity to a run's energy. The two are linked by the efficiency of the chips the run used, which improves with each generation of hardware: the energy is the compute divided by that efficiency, and the emissions are the energy multiplied by the carbon intensity of the grid behind the datacentre. A compute estimate on its own therefore gives neither figure.
Of the 24 models Epoch records at or above 1e25 FLOP, the Systemic risk threshold The training-compute level in Article 51(2) of the EU AI Act at which a general-purpose model is presumed to carry systemic risk. It is a rebuttable presumption rather than a fixed boundary. Glossary entry above which the EU AI Act presumes a model carries systemic risk, one has a published figure for the emissions of training: Llama 3.1 405B. None of the four above 1e26 FLOP has one. Meta's model card for Llama 3.1 405B gives 8,930 tonnes of CO2 equivalent on a location-based basis and zero on a market-based basis, and the energy behind it, 21,588 MWh, is derived by this site from the card's GPU hours and stated power per chip. For OpenAI, Google DeepMind, Anthropic and xAI, the index records no published training energy or emissions figure for any model. Compute is Epoch AI's estimate. Whether a training figure is published is this site's index. Environmental Disclosure Index This site 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 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/explainers/environmental-impact-of-ai/.
Where a published emissions figure and a compute estimate exist for the same model, the emissions for the same amount of compute differ 26-fold, from BLOOM 176B, trained on the French grid, to GPT-3 175B, on older chips and a different grid, whose figure is an outside 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 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 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/explainers/environmental-impact-of-ai/.
How much energy does one AI prompt use?
Answering a prompt with a trained model is called inference, and three developers have published a figure for the energy or emissions of one. Google gives 0.24 Wh for the median text prompt in Gemini Apps as of May 2025, with 0.03 grams of CO2 equivalent and 0.26 mL of water, and has published its method, which counts idle machines and datacentre overhead. OpenAI gives 0.34 Wh for an average query in ChatGPT, in a blog post by its chief executive that does not say which models or which queries are averaged. Mistral AI gives 1.14 grams of CO2 equivalent and 45 mL of water for one 400-token response from Mistral Large 2, from a lifecycle analysis, without a separate energy figure. Environmental Disclosure Index This site 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.
The three figures measure different things. A median is the middle prompt and an average is the total divided by the count, so a few very long requests raise an average and leave a median where it was. None of them covers image or video generation, long reasoning or agent runs. Anthropic, Meta, xAI and DeepSeek publish no per-prompt figure. Environmental Disclosure Index This site 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.
How much water does AI use?
The IEA estimates the water consumed by all datacentres worldwide at around 560 billion litres in 2023. About two thirds of it is consumed through electricity generation and fuel supply, a quarter in cooling the buildings, and the rest in making the chips. The IEA gives no portion of it for AI. Recorded by this site from the International Energy Agency's Energy and AI (2025), CC BY 4.0. Environmental Disclosure Index This site 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.
Two developers publish water for a single response: Google, at 0.26 mL for its median text prompt, and Mistral AI, at 45 mL for one 400-token response counted over its lifecycle. The two come from different methods. Mistral AI also gives the water used over the lifecycle of one model, training and use together, and no other developer publishes water for a model. Environmental Disclosure Index This site 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.
What is not published?
This site's Environmental Disclosure Index records what each of seven frontier developers has published on four questions: the energy or emissions of training, where it trains and serves its models, what powers those sites, and the energy of one prompt. 12 of its 28 entries record nothing published. Environmental Disclosure Index This site 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.
- No developer publishes the energy used by any training run above 1e26 FLOP.
- Inference moves between regions with demand, and no operator publishes which grid serves a given model, hour by hour.
- Of the published training footprints, only BLOOM's study and Mistral AI's lifecycle analysis count the emissions of making the hardware.
- DeepSeek publishes the GPU hours of its training and nothing on power, sites or emissions.
Each developer's entries, with a link to every statement they rest on, are on the environment page. Environmental Disclosure Index This site 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 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/explainers/environmental-impact-of-ai/.