How much energy does AI use?
No published figure gives the total energy used by AI. The nearest are the International Energy Agency's estimate that datacentres of every kind used around 415 TWh of electricity in 2024, 15% of it by the accelerated servers it uses as a proxy for AI, and Epoch AI's estimate that the 95 large AI datacentres it tracks have 14.2 GW of IT power operating, the power that reaches their computers. The IEA figure recorded by this site from Energy and AI (2025), CC BY 4.0. The hub total summed by this site over Epoch AI's per-site estimates. 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 28 October 2026. AI Data Centers Epoch AI 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/explainers/how-much-energy-does-ai-use/.
The answers below use only figures published on the datacentres and environment pages, where each is set out with its source. What that electricity emits, and the water it consumes, are answered in What are the environmental impacts of AI?
What is the difference between power and energy?
Power is the rate at which electricity is drawn, measured in watts: a megawatt (MW) is a million watts and a gigawatt (GW) a thousand megawatts. Energy is power drawn over time, measured in watt-hours: a site drawing one gigawatt for one hour uses one gigawatt-hour, and a terawatt-hour (TWh) is a thousand gigawatt-hours. The size of a datacentre is stated as power; what datacentres use over a year is stated as energy.
A datacentre's IT power The electrical power available to a datacentre's computing equipment, before cooling, power conversion and other overheads. Recorded per site as the power operating now and the power planned at the last stage of the site's timeline. Glossary entry is what reaches the computers themselves. Its Facility power The whole electrical capacity of a datacentre: IT power plus cooling, lighting and power conversion. Glossary entry adds the cooling, lighting and power conversion around them, so it is always the larger figure.
How much power do AI datacentres draw?
Epoch AI follows the largest AI datacentres, mostly in the United States, through satellite imagery, permits and company statements. Across the 95 sites in its hub it estimates 14.2 GW of IT power operating, with compute equal to about 14.8 million H100 equivalent A unit of compute capacity: the number of Nvidia H100 chips with the same total performance as the chips a site actually holds. Written H100e. Glossary entry . The largest operating site is Anthropic-Amazon New Carlisle, at 1 GW. Summed by this site over Epoch AI's per-site estimates. IT power excludes cooling and other overheads. AI Data Centers Epoch AI 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/explainers/how-much-energy-does-ai-use/.
From permits and stated plans, Epoch projects 38.2 GW of IT power at the same sites once the last of them is complete, in 2030. The largest planned site is Microsoft Fairwater Wisconsin, at 2.3 GW. Planned is Epoch's projection. AI Data Centers Epoch AI 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/explainers/how-much-energy-does-ai-use/.
The hub is not a census. Epoch estimated in September 2026 that its sites hold 44% of the world's deployed AI compute, and gives 23% to 81% as the range it is 90% confident contains the true share. Four in five of its power estimates fall within a factor of 1.4 of the true value, by Epoch's own account. AI Data Centers Epoch AI 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/explainers/how-much-energy-does-ai-use/.
How much electricity do datacentres use?
Every datacentre in the world, AI or otherwise, used around 415 TWh in 2024, about 1.5% of global electricity demand, by the IEA's estimate. Accelerated servers, the machines built around graphics processors and other AI chips, used 15% of it. 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.
In the United States, Lawrence Berkeley National Laboratory puts datacentre electricity at 192 TWh in 2024, 4.7% of the country's electricity, and gives a share for AI only for 2030, as a projection. Google, the one operator of frontier AI datacentres that publishes the figure, reports 42.4 TWh for its datacentres in 2025, up 38% on the year before, for all of its work and not broken down by 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 energy does it take to train an AI model?
Seven training runs have an energy figure that is published or can be derived from what is published, and the largest is Llama 3.1 405B, at 21,588 MWh. For five of them, all Meta models, this site derives the energy from the GPU hours and power per chip stated in the model card. The figure for GPT-3 175B is an outside estimate, and the figure for BLOOM 176B is the developer's own. 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.
- Llama 3.1 405B (Meta): 21,588 MWh, derived from the model card.
- Llama 3 70B (Meta): 4,480 MWh, derived from the model card.
- Llama 4 Scout (Meta): 3,500 MWh, derived from the model card.
- Llama 4 Maverick (Meta): 1,666 MWh, derived from the model card.
- GPT-3 175B (OpenAI): 1,287 MWh, an outside estimate.
- Llama 2 70B (Meta): 688 MWh, derived from the model card.
- BLOOM 176B (BigScience): 433 MWh.
None of the four runs Epoch AI records above 1e26 FLOP has a published energy figure, including the largest, GPT-6 Astra at 9.4e26 FLOP, published by OpenAI in September 2026. A run's energy is its computation divided by the efficiency of the chips it ran on, so a compute estimate alone does not give it. Compute is Epoch AI's estimate. Whether energy 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 28 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 28 October 2026 via AI Safety Tracker, https://aisafetytracker.org/explainers/how-much-energy-does-ai-use/.
How much energy does one AI prompt use?
Google states 0.24 Wh for the median text prompt in Gemini Apps as of May 2025, with a published method. OpenAI states 0.34 Wh for an average query in ChatGPT, without one. A median is the middle prompt and an average is the total divided by the count, so the two figures measure different things. Mistral AI gives a per-response figure for emissions and water, without one for energy. 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 AI datacentre capacity is planned in Australia?
Epoch AI's hub holds one Australian site. This site's AI Datacentre Index adds 13 more from developers' own statements, counting a site only where a primary source says it is built for AI or names its AI tenant: two operating, two under construction, two approved and seven proposed. Together they state 4.9 GW of IT capacity at campuses built for AI, of which 4.6 GW is at proposed sites not yet approved. The largest is Western Downs Digital Park, at 1.4 GW, developed by Zerra DC. Summed by this site over IT capacity as each developer states it, at campuses built for AI. Sites that state only facility power, and general-purpose buildings with an AI tenant, are left out of the sum. AI Datacentre Index This site AI datacentre sites that Epoch AI's hub does not hold, recorded from the statements of developers, named tenants and planning authorities. Australia only so far. A site is included only where a primary source says it is built for AI compute or names the AI deployment inside it. Capacity is as the developer states it, IT or facility power as labelled, and is not an estimate; most of it is proposed and not yet approved. Coordinates are this site's placement of each stated address or area. Cite AI Safety Tracker, "AI Datacentre Index". https://aisafetytracker.org/datacentres/#added. Retrieved 18 October 2026.
These are statements of intent, and most of the capacity is not yet approved, so they are kept apart from Epoch's estimates of what is built. Eight of the 13 sites state how they will be supplied with power. AI Datacentre Index This site AI datacentre sites that Epoch AI's hub does not hold, recorded from the statements of developers, named tenants and planning authorities. Australia only so far. A site is included only where a primary source says it is built for AI compute or names the AI deployment inside it. Capacity is as the developer states it, IT or facility power as labelled, and is not an estimate; most of it is proposed and not yet approved. Coordinates are this site's placement of each stated address or area. Cite AI Safety Tracker, "AI Datacentre Index". https://aisafetytracker.org/datacentres/#added. Retrieved 18 October 2026.
What is not known?
- The electricity used by AI worldwide, which no publisher states.
- Whether a given datacentre trains models or serves them, and what powers and cools it, which Epoch's hub records for no site.
- The energy of any training run above 1e26 FLOP.
- The energy of image or video generation, long reasoning or agent runs, which no per-prompt figure covers.
Every site, with its power now and planned, is on the datacentres page. AI Data Centers Epoch AI 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/explainers/how-much-energy-does-ai-use/. 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 28 October 2026.