Frequently asked questions
Short answers drawn from the figures the data pages publish, each with its source. The longer answers are in the explainers.
- Which AI model or service has the smallest environmental footprint?
- How much energy does one AI prompt use?
- How much electricity do datacentres use?
- How much law applies to AI?
- Do AI developers publish safety testing when a model is released?
- Does this site make forecasts?
- Where does the data come from, and may it be reused?
- How are errors reported and corrected?
Which AI model or service has the smallest environmental footprint?
No published measure ranks AI models or services by environmental footprint. The developers that publish a figure per prompt each measure something different, and the published figures for electricity, clean energy and water describe whole companies, not models. 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 datacentre operators recorded here, Google, Microsoft, Amazon and Meta, match their electricity over a year with renewable energy bought by contract. 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 does not show whether clean power was on the grid in the hour it was used. Google alone publishes that 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 : 65% across its datacentres in 2025, from 98% on Finland's grid to 5% in Singapore. From each operator's own 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.
Google consumed 41 billion litres of water across its datacentres and offices in 2025, up 34% on the year before, and states that cooling with water needs less energy than cooling with air. Microsoft has designed every new datacentre since August 2024 to cool chips in a closed loop that evaporates no water once filled, with a stated nominal increase in electricity use. The first of those sites are expected online in late 2027; its existing datacentres use both air and water cooling. From Google's 2026 Environmental Report and Microsoft's December 2024 post on its datacentre design. 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.
Some developers also compute at sites with gas generation on site, as reported by others: permits have been filed for natural-gas turbines at OpenAI's Stargate campus in Abilene, Texas, and xAI's Colossus 1 in Memphis has permitted gas turbines. Anthropic's models run in Amazon's and Google's datacentres and, from May 2026, on Colossus 1, so one developer's models can draw on several operators and grids. Power sources marked reported are from permits and press reporting, not from the developer. 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.
Every entry, with a link to each statement it rests on, is on the environment page.
How much energy does one AI prompt use?
Three developers publish a figure for a single prompt or response: Google, OpenAI and Mistral AI. Each measures a different statistic in a different way, so the figures cannot be ranked against each other.
Google puts the median text prompt in its Gemini apps at 0.24 Wh, 0.03 gCO2e and 0.26 mL of water (May 2025), by a published method that counts idle machines and datacentre overhead. OpenAI's chief executive put an average ChatGPT query at 0.34 Wh (June 2025), without stating which models or queries are averaged. Mistral AI puts one 400-token response from Mistral Large 2 at 1.14 gCO2e and 45 mL of water, from a lifecycle analysis that gives no 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.
A few very long requests raise an average and leave a median where it was. None of the three figures 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 electricity do datacentres use?
The International Energy Agency puts the electricity used by every datacentre in the world at around 415 TWh in 2024, about 1.5% of global demand, and 485 TWh in 2025. It does not count AI separately. Recorded by this site from the International Energy Agency's Energy and AI (2025) and Key Questions on Energy and AI (2026), 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.
Its nearest measure of AI is the electricity drawn by accelerated servers, those built around graphics processors or other AI chips: 15% of the 2024 total, which the IEA describes as an imperfect proxy. The IEA's projections to 2035 are drawn on the environment page, and the power of individual AI datacentres is on the datacentres 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.
How much law applies to AI?
This site's law index holds 86 instruments across 11 jurisdictions. 48 of them are binding law, and 10 of those name AI in their own title. The remainder apply to AI without naming it. AI Law and Policy Index This site Our own reading of primary instruments, with a link to each. Not legal advice, not a compliance checklist and not exhaustive. Instruments that never mention AI are included where they apply to it, which is most of the Australian list. Statuses and commencement dates go stale; every record carries its review date. Cite AI Safety Tracker, "AI Law and Policy Index". https://aisafetytracker.org/policy/. Retrieved 23 October 2026.
Each instrument is tagged by how binding it is and links to its primary source on the policy page, by jurisdiction, and on the timeline, by year of adoption. The index holds more instruments for Australia than for any other jurisdiction, so the totals are not a comparison between jurisdictions.
Do AI developers publish safety testing when a model is released?
Of the 69 releases by frontier developers indexed since January 2025, 48 had a system card on the day of release, 4 had one later and 17 have none. Release Safety Disclosure Index This site Our own reading of the system card, model card or safety report each developer published for each release Epoch AI lists from 2025, joined to Epoch's list. Every cell is what a developer says about its own model under its own framework: whether a card came out with the release, whether it reports dangerous-capability results, which government institutes it names as pre-release testers, and whether it says a risk threshold was reached. None of it is verified. A release Epoch adds after the last review is shown as awaiting review. Cite AI Safety Tracker, "Release Safety Disclosure Index". https://aisafetytracker.org/alignment/#releases. Retrieved 18 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 18 October 2026 via AI Safety Tracker, https://aisafetytracker.org/faq/.
A System card A document a developer publishes for a model, describing its evaluations, safeguards and limitations. Also called a model card or safety report. Dated here by when it first appeared. Glossary entry is the developer's own account of how a model was tested before release. Each release, its card and what the card reports are listed on the alignment page.
Does this site make forecasts?
No. Projections on the site are the IEA's own, labelled with the scenario they belong to, and the straight-line extrapolations on the progress page are labelled as fits, with their window and both directions of bias.
The scope and method are set out on the about and methodology pages.
Where does the data come from, and may it be reused?
Every figure carries a marker naming its source and licence, and every source is listed on the sources page with what it cannot show. The indexes this site codes from primary documents are MIT licensed and downloadable as JSON and, where the records are flat, as CSV.
Sources, licences and known limitations are on the sources page, and the downloads under downloads. Figures from other publishers remain theirs and are reproduced under their licences.
How are errors reported and corrected?
By email to the address on the corrections page. Corrections that change what a figure says are listed there with the date.
The address and the list of corrections are on the corrections page.