—
—
Running AI: data centres
AI is trained and run in data centres, which also serve the rest of the internet; no official statistics split out AI. Two countries measure data-centre electricity at the meter, Ireland and the Netherlands; everything else is an estimate, and each figure here carries its year and who made it.
Where do data centres take the most electricity?
Data centres’ share of a country’s electricity, by year. Solid lines are measured at the meter; the dotted line is a modelled estimate.
Ireland: data centres’ share of electricity, by quarter
Ireland: data centres against everyone else
Electricity used each year, in gigawatt-hours.
How much energy does one AI answer take?
What companies say about their own services, and what independent or peer-reviewed work finds. Watt-hours per answer where given; the figures count different things, so read the notes.
| Who | What was measured | Energy | CO₂ | Water | Notes |
|---|
What the companies report about themselves
Electricity use from the companies’ own environmental reports, in terawatt-hours a year. These are the companies’ figures, not independent measurements, and each counts something slightly different.
| Company | What is counted | Latest | Efficiency (PUE) | Water |
|---|
Estimates for the world and the United States
Water
The climate side
Training AI
Training a frontier model means running tens of thousands of chips for months.
—
Power drawn by AI training runs
Estimated power of the whole training cluster (chips, servers, networking and cooling), in megawatts, on a log scale: each gridline is ten times the one below. Orange: frontier models, the largest of their time. Shape is Epoch’s confidence in the estimate: circle “confident” (within 3 times), diamond “likely” (within 10 times), triangle “speculative” (within 30 times).
The largest training runs
By estimated power draw. Energy is power times training time, where the training time is known.
| Model | Published | Power | Chips | Training time | Energy | Confidence |
|---|
AI chips: work per watt
Peak 16-bit performance per watt of rated power for AI chips by release date, in trillion operations per second per watt.
What this means
- Efficiency is not keeping up. Epoch AI finds that better chips (about 12 times more efficient over ten years), lower-precision number formats and longer training runs cut the power needed for a given amount of training by roughly half each year, but the compute used for frontier training grows much faster, so total power keeps rising.
- Training is only part of it. Running AI models for users (inference) also uses power, and is not counted in this section. Data-centre electricity, which includes it, is above.
- The climate side. Emissions depend on where the power comes from: the same training run emits far less on a clean grid. Few developers publish the energy or emissions of their training runs, so these are estimates from the hardware used.
What this drives
- Electricity in Europe →How clean the grid is.
Related
- Carbon footprints →Emissions per person by country.
How we know this
- Sources: Measured data-centre electricity: Central Statistics Office Ireland (table MEC02, read daily through its open API) and Statistics Netherlands (CBS, yearly file, copied by hand). Estimates: the IEA’s Energy and AI (2025) and Key Questions on Energy and AI (April 2026), Lawrence Berkeley National Laboratory’s 2024 report and 2025 update (June 2026), NVE Norway, and the European Commission’s first data-centre efficiency report (September 2026). Training: Epoch AI, an independent research institute: its Notable AI Models database and its Machine Learning Hardware database.
- Update cadence: Ireland’s figures are read every day and change about once a year (released each July); Epoch adds major models within about two weeks of release and we check every day; everything else is quoted from reports and rechecked at each review, last on 7 November 2026.
- Methods: growth rates are our least-squares fit of the logarithm against time for frontier models with an estimate (power since 2010, compute since 2018). Chip efficiency is our calculation: peak 16-bit (FP16/BF16) tensor performance divided by rated power (TDP). Energy is estimated power times reported training time.
- Uncertainty: the world and US figures are modelled estimates with wide forecast ranges; the training figures are estimates from the hardware and training time, and Epoch rates each model “Confident” (within 3 times), “Likely” (within 10 times) or “Speculative” (within 30 times).
- Company figures: what Google, Microsoft, Meta and Amazon publish about themselves, kept apart from measured statistics; they count different things (Google and Meta: data centres; Microsoft: the whole company by fiscal year; Amazon: no electricity figure). Energy-per-answer figures from companies cover only what each company chose to include.
- Limitations: no statistic separates AI from the rest of what data centres do; few developers publish their training details, so recent models are often missing or estimated.
- Primary data: CSO MEC02 · CBS Netherlands · IEA Key Questions on Energy and AI (2026) · IEA Energy and AI (2025) · LBNL 2025 update · LBNL 2024 report · Epoch AI models · Epoch AI hardware · Epoch AI on power trends · How this page is built