How we track the energy of AI
What the dashboard shows, where the numbers come from, how current they are, and what they can’t tell you.
The short answer
AI uses electricity twice: every time it is used, in data centres, and once when each model is trained, on tens of thousands of specialised chips running for weeks or months. Data centres as a whole used about 485 TWh in 2025 by the International Energy Agency’s estimate, around 1.5% of the world’s electricity, and the agency expects that to roughly double by 2030. No statistic separates AI from the rest of what data centres do. Where data centres are measured at the meter, in Ireland and the Netherlands, their share of the country’s electricity is far above the world average: 23% in Ireland in 2025, 4.6% in the Netherlands in 2024. One AI answer costs a few tenths of a watt-hour by every published figure; the growth comes from how much AI is used and trained, not from the cost of a single answer.
How we know
Measured. Ireland’s Central Statistics Office publishes, from grid meter data, how much electricity data centres and all other customers use each quarter since 2015; we read it every day through its open interface and add up full years. Statistics Netherlands publishes a yearly file of electricity delivered from the public grid to connections whose main business is a data centre, 2017 to 2024, which we copy by hand. Norway’s energy regulator reports the metered total for Norway in its yearly reports.
Estimated. The world figures are the IEA’s, from Energy and AI (April 2025, for 2024) and Key Questions on Energy and AI (April 2026, for 2025 and the outlook to 2030). The US figures are Lawrence Berkeley National Laboratory’s, from its 2024 report (which also has the water figures) and its 2025 update of June 2026. The European Commission’s first report on data-centre efficiency (September 2026) gives the number of data centres that reported for 2024 and their average efficiency; its EU electricity total is the IEA’s estimate. These are modelled numbers with ranges, and each one on the dashboard carries its year and its author.
What the companies say. Google, Microsoft, Meta and Amazon publish yearly environmental reports. We copy their electricity, efficiency and water figures and keep them in their own section, labelled as the companies’ figures: they are mostly checked by auditors but not independently measured, and each company draws the line differently (Google and Meta report their data centres, Microsoft its whole company by fiscal year, Amazon no electricity figure at all). The energy per AI answer is treated the same way: what Google, OpenAI and Mistral say about their own services sits next to what Epoch AI, researchers publishing in the journal Joule, the AI Energy Score benchmark and a peer-reviewed study found, each with its date and what it counted.
Training. Epoch AI, an independent research institute, keeps a database of notable AI models: models that advanced the state of the art, were highly cited, widely used or historically important. For each it records, where known, the training compute, the chips used, their number and the training time, and estimates the power the whole training cluster drew, including servers, networking and cooling. Epoch rates each estimate Confident (within 3 times), Likely (within 10 times) or Speculative (within 30 times); the dashboard shows that rating on every point. Growth rates are our least-squares fit of the logarithm of power (frontier models since 2010) or compute (since 2018) against time; frontier models are those in the top 5 by training compute when published. Epoch’s own analysis (published 2024, data updated November 2025) found 2.1 times a year for the top 10 and 2.2 times for the top 5, close to our fit, and in August 2025 Epoch and EPRI projected that the largest single training runs could reach 4 to 16 gigawatts by 2030. Chip efficiency is peak 16-bit performance divided by rated power, at one precision so chips compare like for like. Energy for a training run is estimated power times reported training time; where the time is not published, no energy figure is given.
How current each part is
- Ireland: read every day; the statistics office releases a new year each July (latest: 7 July 2026, data to the end of 2025).
- Netherlands: a yearly file, latest 15 December 2025 (data to 2024, provisional); copied by hand.
- IEA, LBNL, NVE, European Commission: yearly or one-off reports; every figure checked against the document on 7 November 2026. The IEA’s World Energy Outlook 2026 is due on 18 November 2026.
- Company reports: Google June 2026, Microsoft July 2026, Amazon July 2026, Meta September 2025 (no 2026 edition by 7 November 2026).
- Energy per answer: publications from 2024 to April 2026, checked 7 November 2026.
- Models and chips: Epoch adds major models within about two weeks of release; we check every day.
What the dashboard can’t tell you
How much of data-centre electricity is AI. No official statistic separates it; the IEA estimates that electricity for AI-focused data centres grew 50% in 2025, but that is a model, not a meter.
Whether the companies’ figures are comparable. Each counts a different boundary and some report by fiscal year, so the lines in their chart should be read one company at a time.
What a particular use of AI costs you. The published figures are medians or averages for short text answers on particular systems; a long reasoning answer, an image or a video can cost ten times more or beyond, and none of the figures includes the network or your device.
Most training figures are estimates from the hardware, not measurements: few developers publish the energy their training used, and the newest models often have no estimate yet.
Further reading
- CSO Ireland: Data Centres Metered Electricity Consumption 2025. cso.ie
- Statistics Netherlands: Elektriciteit geleverd aan datacenters, 2017–2024. cbs.nl
- IEA: Key Questions on Energy and AI (2026) and Energy and AI (2025). iea.org
- Lawrence Berkeley National Laboratory: United States Data Center Energy Usage Report: 2025 Update (2026). lbl.gov
- European Commission: Report on the energy efficiency of data centres in the EU, COM(2026) 500. eur-lex.europa.eu
- Google: Measuring the environmental impact of AI inference (2025). cloud.google.com
- Oviedo et al. (2026), Joule: energy use of AI inference, efficiency pathways, and test-time scaling. doi.org/10.1016/j.joule.2026.102430
- Luccioni, Jernite and Strubell (2024), ACM FAccT: Power Hungry Processing: Watts Driving the Cost of AI Deployment? doi.org/10.1145/3630106.3658542
- Epoch AI: The power required to train frontier AI models is doubling annually; How much power will frontier AI training demand in 2030? epoch.ai
Credits
Irish data from the Central Statistics Office and Dutch data from Statistics Netherlands (both CC BY 4.0); training data from Epoch AI (CC BY 4.0). World, US, EU and Norwegian figures are quoted from the International Energy Agency, Lawrence Berkeley National Laboratory, the European Commission and NVE; company figures from the companies’ own reports. We download or quote their public data and add our own calculations as described above.
← Back to the dashboard