Article · 18 min read

How much energy does AI use?

A question to a chatbot uses less electricity than a TV does in ten seconds. Yet AI companies are building data centres that use as much electricity as big cities, and in Sweden data centres are queuing to connect thousands of megawatts. Here is how it all fits together.

Short answer

An ordinary text prompt uses about 0.3 watt-hours, about as much as a TV uses in ten seconds. But reasoning answers, images and video use tens to thousands of times more, and billions of prompts a day add up. The world’s data centres used 485 terawatt-hours of electricity in 2025, just under 2 percent of all electricity, and the IEA expects that to double by 2030, with AI as the biggest driver. More efficient chips and models are not keeping up with growing use, and where data centres are built decides how much they emit.

A TV for nine seconds, or a whole city

How much electricity does it take when you ask a chatbot a question? Google has measured it for its AI service Gemini. If you line up all text prompts by how much electricity they used, the one in the middle needed 0.24 watt-hours in May 2025. That is less electricity than a TV uses in nine seconds.

At the same time, AI companies are building data centres that use as much electricity as big cities. The largest one tracked by the research institute Epoch AI, Anthropic and Amazon’s site in New Carlisle, Indiana, can supply its computers with about 1,000 megawatts in November 2026. The city of San Diego draws about 800 megawatts on average. In Sweden, data centres were queuing in April 2026 to connect a total of 4,643 megawatts to the grid. If all of them were built and ran at full power around the clock, they would use about 40 terawatt-hours a year, roughly 30 percent of all the electricity Sweden used in 2025.

Both pictures are true. This article explains how, from a single question to all the world’s data centres. In short, an ordinary question uses little electricity, but what you ask for, how many people ask and where the electricity comes from all matter a great deal.

First, some units. A watt-hour is the energy a one-watt lamp uses in an hour. A thousand watt-hours make a kilowatt-hour, the unit on the electricity bill, and a terawatt-hour is a billion kilowatt-hours. Power is measured in watts. A 100-megawatt data centre running at full power for a whole year uses 876 gigawatt-hours.

A question uses less than people thought

For a long time the numbers were rough estimates. In 2023 the researcher Alex de Vries calculated that a question to ChatGPT might use about 3 watt-hours. In 2024 the American research institute EPRI wrote that an AI query uses ten times as much electricity as an ordinary Google search, which according to Google itself used 0.3 watt-hours in 2009. The ten-times claim was widely repeated.

Since 2025 there has been better evidence, and it points to much lower numbers. Besides Google’s measurement, OpenAI’s chief executive Sam Altman wrote in June 2025 that an average ChatGPT query uses about 0.34 watt-hours, but he did not explain how the number was arrived at. Researchers at Microsoft estimated in the journal Joule in 2026 that a query to the largest models uses about 0.31 watt-hours, and that the widely cited numbers are 4 to 20 times too high. The old estimates assumed, among other things, that questions and answers were very long.

0.3 watt-hours is not much. A phone takes about 19 watt-hours to charge, according to the US Environmental Protection Agency, and boiling a litre of water in a kettle takes 100 to 130 watt-hours. For the electricity it takes to boil a litre of water, you could ask several hundred questions.

But it depends on what you ask for. Bigger models use more. When MIT Technology Review had Meta’s open models measured in 2025, an answer from the smallest, Llama 3.1 8B, took about 114 joules, counting cooling and everything else the data centre needs. The largest, Llama 3.1 405B, needed about 6,700 joules, almost 60 times as much. A joule is a watt-second, so 3,600 joules make a watt-hour.

The biggest difference comes when the model reasons. A reasoning model first writes a long line of reasoning for itself before it answers, and every token costs electricity. We explain how a model writes token by token in How does a language model work? When Sasha Luccioni and Boris Gamazaychikov of the AI Energy Score project measured a range of open models in December 2025, those that reason used on average 30 times more energy than those that do not. The same model could use between 150 and 700 times more with reasoning switched on, because it then wrote between 300 and 800 times more tokens. The measurements were made one query at a time, which according to Google does not match how models run in large data centres. Microsoft’s researchers, who modelled large data centres, estimated that a query where the model works 15 times longer than usual uses about 3.9 watt-hours.

Images and video use even more. An image from the open model Stable Diffusion 3 Medium took about 0.6 watt-hours in MIT Technology Review’s measurement. A 2026 study estimated that an eight-second video uses between about 20 and 530 watt-hours depending on the model, from Google’s Veo 3 to OpenAI’s Sora 2 Pro. The most expensive one equals the electricity for over 2,000 ordinary text prompts, or for boiling four litres of water.

Figure 1A question in everyday terms

This figure is interactive and needs JavaScript.

Tap a row to compare it with everyday life. The scale is logarithmic: each step to the right is ten times more. The AI numbers are measurements and estimates from Google, OpenAI, Microsoft, MIT Technology Review, AI Energy Score and a 2026 study of video models. The everyday numbers come from the EPA, Forbrugerrådet Tænk, the Swedish Energy Agency and the Swedish Transport Administration.

It depends on what you count

The numbers differ so much not only because of the models. It also depends on what is counted. Google’s measurement was the first of its kind from a major AI company, and it shows how much the boundaries matter.

Of the 0.24 watt-hours, 0.14 went to the AI chips that compute the answer and 0.06 to the computers around the chips, with their processors and memory. 0.02 went to machines kept in reserve to handle peaks, and 0.02 to the data centre’s own needs, such as cooling and losses when power is converted. Counting the way Google says narrower measurements do, with only the chips and only in the most efficient data centres, gives 0.10 watt-hours.

Figure 2What gets counted

This figure is interactive and needs JavaScript.

Google’s measurement of text prompts to Gemini in May 2025. Drag the slider or press the button to count one part at a time. The narrow measurement counts only the AI chips, and only in the most efficient data centres.

Google’s number is not the whole picture either. It covers text prompts only, not images or video. It does not include the energy to train the model, the networks that carry the question, or your own phone or computer. Google itself has written that the data have not been checked by any independent party. And a median says nothing about the prompts that use the most. Google chose the median precisely because a few extreme prompts would otherwise pull up the average.

Water shows the same thing. According to the company, Google’s median prompt consumed 0.26 millilitres of water, about five drops. That number covers only the water in the data centre itself. Experts interviewed by The Verge point out that most of a data centre’s water use comes from the power plants that make its electricity. American researchers who also counted that water found that the model GPT-3 consumed half a litre for roughly 10 to 50 medium-length answers, though they assumed that each answer used 4 watt-hours. The French company Mistral AI counts the whole life cycle, including making the servers, and arrives at 45 millilitres for a 400-token answer. The numbers measure different things and cannot be compared directly.

Emissions are even harder to compare. Google gives 0.03 grams of carbon dioxide per prompt, based on the electricity the company buys through renewable energy contracts. Mistral gives 1.14 grams for an answer, based on the electricity mix where its data centres are. Both include the manufacturing of the hardware.

From one question to billions

A single question uses little, but there are a great many of them. In July 2025 ChatGPT’s users sent more than 2.5 billion messages a day, according to a study by researchers at OpenAI and Harvard. At Altman’s 0.34 watt-hours that comes to about 850 megawatt-hours a day, or just over 300 gigawatt-hours a year. That is roughly the electricity 160,000 Swedish apartments use, but only about 0.2 percent of Sweden’s electricity use.

Google is even bigger. In May 2026 the company’s services processed more than 3.2 quadrillion tokens a month, seven times more than a year earlier and over 300 times more than two years earlier. Microsoft’s researchers have calculated that a billion queries a day need about 0.7 gigawatt-hours, and 1.7 gigawatt-hours if one query in ten is a long one where the model reasons.

For an individual user the electricity is small. Someone who asks ChatGPT 20 questions a day for a year uses, by Altman’s number, about 2.5 kilowatt-hours. That is just under two five-minute showers according to the Swedish Energy Agency’s figures. But six videos from the most energy-hungry model in the 2026 study use as much as that whole year of questions.

Figure 3You and everyone else

This figure is interactive and needs JavaScript.

Do the sums yourself. A text prompt counts as 0.34 watt-hours, OpenAI’s number for ChatGPT. The videos use the averages for Veo 3 and Sora 2 Pro in the 2026 study. The comparisons use Statistics Sweden’s standard apartment, which uses 2,000 kilowatt-hours of electricity a year, and Sweden’s electricity use in 2025, 134 terawatt-hours. The training numbers are estimates by the IEA and Epoch AI.

Then there is training. Before a model can answer anything it has to be trained, for weeks or months on thousands or tens of thousands of chips. Training GPT-3 in 2020 took about 1,300 megawatt-hours, according to an estimate by researchers at Google and the University of California, Berkeley. For GPT-4 the International Energy Agency (IEA) estimates about 42 gigawatt-hours. For xAI’s Grok 4 from 2025, Epoch AI estimates 310 gigawatt-hours, about the same as a year of ChatGPT messages by the sum above. According to Epoch, the power the largest training runs need has grown about 2.2 times a year, and the largest now draw over 100 megawatts.

The final training run is also only part of the work. When researchers at Ai2 and the University of Washington accounted in 2026 for everything behind the Olmo 3 family of models, 82 percent of the computation went to experiments, failed attempts and tests before the final training. How the energy divides between training and use, known as inference, varies. At Google about three fifths of the energy for machine learning went to using the models and two fifths to training them, according to a study of the years 2019 to 2021. Epoch AI judged in 2025 that AI’s power demand was split roughly equally between training, experiments and use. For a single popular model, use still catches up quickly. GPT-4’s training equals about 50 days of ChatGPT messages.

More efficient, yet more

AI is quickly becoming more efficient. According to Google, the energy for a median prompt to Gemini fell 33-fold in a year, from May 2024 to May 2025, mostly thanks to more efficient models and better-used computers. Emissions per prompt fell 44-fold.

Yet electricity use keeps rising. Over the same year the number of tokens Google processed grew 50-fold, and the year after a further seven-fold. Google’s data centres used 42.4 terawatt-hours of electricity in 2025, 38 percent more than the year before and more than twice as much as in 2021. The company’s emissions rose 18 percent in a year. Microsoft’s rose 25 percent in a year according to its latest sustainability report, mainly because it built more data centres.

Figure 4More efficient, but more

This figure is interactive and needs JavaScript.

Google’s own numbers: the energy per median prompt to Gemini, the number of tokens per month and its data centres’ electricity use. Everything is shown relative to 2024 on a logarithmic scale, where each step is ten times. The curves measure different things and cannot be combined exactly.

The phenomenon has a name: the Jevons paradox. The British economist William Stanley Jevons wrote in 1865 that it is wholly a confusion of ideas to suppose that the economical use of fuel means using less. The very contrary is the truth. When James Watt made the steam engine more efficient, it became worthwhile to use it for far more, and coal consumption rose. When the Chinese company DeepSeek launched its R1 model in January 2025, which it said was as good as OpenAI’s o1 and which was reportedly built for a fraction of its rivals’ budgets, Microsoft’s chief executive Satya Nadella wrote that the Jevons paradox strikes again: as AI gets more efficient and accessible, its use will skyrocket.

AI is also being used in new ways that take more per task. Reasoning models write hundreds of times more tokens than ordinary ones, and AI agents working on their own make many calls in a row before they are done. So the energy per token can fall while the energy per task rises.

The world’s data centres

All this shows up in the statistics. The world’s data centres used about 415 terawatt-hours of electricity in 2024, roughly 1.5 percent of all electricity, according to the IEA. In 2025 that rose 17 percent to 485 terawatt-hours, about as much as Germany generates in a year. Data centres are used for much more than AI, such as cloud services, streaming video and search. But electricity for data centres built for AI rose 50 percent in a year, and according to the IEA the servers used for AI account for almost half of the increase to 2030.

The IEA expects data centres’ electricity use to double to about 950 terawatt-hours in 2030, roughly 3 percent of the world’s electricity, with the electricity for AI data centres tripling. After 2030 the uncertainty is large. In 2035 data centres could use anything from 700 to 1,700 terawatt-hours across the IEA’s scenarios, depending on how fast AI spreads and how much more efficient the technology becomes.

That is a lot, but not the most. To 2030, data centres account for about a tenth of the growth in the world’s electricity use. Electric motors in industry, air conditioning and electric cars each account for more. The problem is that data centres cluster in a few places.

In 2024 the United States accounted for 45 percent of the world’s data centre electricity, China for 25 and Europe for 15. In the United States, data centres are expected to account for almost half of the growth in electricity use to 2030, and by then to use more electricity than all production of aluminium, steel, cement and chemicals combined. In Ireland, 23 percent of electricity went to data centres in 2025, compared with 28 percent for all households. In Virginia, one of the world’s biggest data centre hubs, the share is about a quarter.

Figure 5The world’s data centres

This figure is interactive and needs JavaScript.

The numbers for 2024 and the scenarios for 2035 come from the IEA report Energy and AI (April 2025), the numbers for 2025 and 2030 from the IEA’s update in April 2026. The shares of electricity come from the IEA, the US Department of Energy, Ireland’s Central Statistics Office, EPRI and Svenska kraftnät.

Where the electricity comes from varies. In 2024 about 30 percent of data centres’ electricity came from coal, 27 percent from renewables, 26 percent from natural gas and 15 percent from nuclear, according to the IEA. The companies are betting on nuclear power. Microsoft has a 20-year agreement for the electricity from the Three Mile Island 1 reactor in Pennsylvania, which is to be restarted and is expected to run from 2027. Google, Amazon and Meta have signed agreements for new nuclear power, mainly small reactors that are to start delivering from around 2030. But to 2030, gas and coal together are expected to supply over 40 percent of the new electricity for data centres.

Sweden: clean power and a long queue

In Sweden, data centres use about 3 terawatt-hours a year today, just over 2 percent of the electricity, according to Svenska kraftnät, the national grid operator, and the Swedish Energy Agency. There are no official statistics, and the research body Energiforsk puts it at about 4. But interest is high. Svenska kraftnät writes that the new demand for electricity comes mainly from data centres, not from industry as before. In April 2026 data centres accounted for 4,643 megawatts in Svenska kraftnät’s queue.

Several large projects are under way. The Canadian firm Brookfield wants to invest up to 95 billion kronor in a 750-megawatt AI data centre in Strängnäs. That is too much for the local grid, according to Svenska kraftnät, which is studying how the site could get all the electricity it wants. In 2024 Microsoft promised to invest 33.7 billion kronor in its data centres in Sandviken, Gävle and Staffanstorp. EcoDataCenter is building a new data centre in Borlänge, where the French company Mistral AI is investing 1.2 billion euros in an AI data centre due to open in 2027. And in June 2026 Google started building a data centre in Horndal in Avesta municipality.

Sweden has electricity with very low emissions and district heating networks that can make use of the computers’ waste heat. In Stockholm, Stockholm Exergi heats 11,000 apartments today with surplus heat from data centres and others. Google’s data centre in Horndal is being built to deliver heat to its surroundings, and in Finland waste heat from Microsoft’s new data centres is to cover about 40 percent of the district heating in Fortum’s network in Espoo and Kirkkonummi.

But the electricity for data centres competes with other uses. Since 1 July 2023 data centres have paid the ordinary energy tax on their electricity, after an earlier tax break was abolished. Svenska kraftnät has warned that new large electricity users could push up prices, and the Social Democrats have demanded that the government draw up a plan for how data centres should be established. The government has started a dialogue with companies that build data centres.

Nobody knows how much it will be. In Svenska kraftnät’s scenarios for 2050, data centres use anything from 9 to 41 terawatt-hours a year, compared with 3 today. The Swedish Energy Agency judges that it could be just over 9 terawatt-hours as early as around 2030 if the announced plans go ahead.

Where a data centre stands also decides its emissions. The electricity used in Sweden in 2025 caused on average about 21 grams of carbon dioxide equivalent per kilowatt-hour, according to Electricity Maps. In Germany it was 335 grams and in Poland 677. The same question therefore causes over 30 times more emissions in Poland than in Sweden. But even clean electricity is a limited resource. The electricity that goes to data centres cannot also go to steelworks, electric cars or heat pumps.

Figure 6Same question, different power

This figure is interactive and needs JavaScript.

Choose what the electricity is for and tap a country. The emissions per kilowatt-hour are for 2025. The European countries come from Electricity Maps and cover the electricity used in the country, including imports. The United States, China, the EU and the world come from Ember and cover electricity generation. The methods differ somewhat. A text prompt counts as 0.34 watt-hours and training GPT-4 as 42 gigawatt-hours.

What you can do

For anyone using AI, the electricity for an ordinary question is small compared with most things in everyday life. A few rules of thumb:

  • Pick the right tool. A smaller model or an ordinary search is often enough for simple questions. Save the reasoning modes for tasks that need them.
  • Think before making video. A short AI video can use as much electricity as thousands of text prompts.
  • Keep it in proportion. A year of ChatGPT questions, 20 a day, uses about as much electricity as driving an electric car 15 kilometres. The car, the heating and the shower matter far more to a household’s energy use than chatting does.
  • Ask for the numbers. The big decisions are about where data centres are built, what electricity they get and how openly the companies report. That is where users, customers and voters make a difference.

The EU has started to set requirements. Since 2024, data centres with at least 500 kilowatts of IT power must report, among other things, their energy use, their water use and how much waste heat is reused. The EU AI Act requires providers of general-purpose AI models to document the model’s known or estimated energy consumption, but that information goes only to the authorities, when they ask for it. The European Commission is looking into a possible energy and emissions label for AI, and the first sustainability ratings for individual data centres are expected in 2027.

Our view

This section is SI-bolaget’s opinion. The rest of the article reports the research. Here we say what we ourselves think.

AI’s energy question is real, but it is not decided by your next question to a chatbot. It is decided by how much is built, where it is built and what electricity it gets. That is why we think:

  • Openness. AI companies should report how much electricity and water their services use, with shared methods and checked by someone outside. Today we have to trust the companies that report, while those that do not escape scrutiny.
  • The long view. Data centres are built for decades. Sweden needs to plan for them as it does for other heavy industry, with new fossil-free electricity and stronger grids, so that they do not crowd out everything else.
  • Humans at the centre. The electricity AI uses should do people good. The benefit has to be weighed against what it costs households and industry, and the waste heat should be put to use.

More efficient AI is good, but it is not enough. The story from Jevons’s steam engines to Google’s data centres shows that what becomes cheaper to use also gets used more.

Sources

  1. Elsworth et al., Measuring the environmental impact of delivering AI at Google Scale, Google, 2025, and Amin Vahdat and Jeff Dean, How much energy does Google’s AI use? We did the math, Google Cloud, 21 August 2025.
  2. Epoch AI, The largest AI data centers by power capacity, read 14 November 2026.
  3. Energiforsk and Nepp, Så får industrin tillräckligt med el till 2045, report 2026:1199, June 2026.
  4. Svenska kraftnät, Kraftbalansen på den svenska elmarknaden, rapport vår 2026, 15 May 2026.
  5. Alex de Vries, The growing energy footprint of artificial intelligence, Joule 7, 2023, and EPRI, Powering Intelligence: Analyzing Artificial Intelligence and Data Center Energy Consumption, 2024.
  6. Urs Hölzle, Powering a Google search, Official Google Blog, 11 January 2009.
  7. Sam Altman, The Gentle Singularity, 10 June 2025.
  8. Oviedo et al., Energy use of AI inference, efficiency pathways, and test-time scaling, Microsoft, Joule, 2026.
  9. Josh You, How much energy does ChatGPT use?, Epoch AI, February 2025.
  10. Everyday electricity: the EPA (phone charge), Forbrugerrådet Tænk (kettle), the Swedish Energy Agency (shower), the Swedish Transport Administration (new electric cars, 17 kilowatt-hours per 100 kilometres) and Statistics Sweden (standard apartment).
  11. James O’Donnell and Casey Crownhart, We did the math on AI’s energy footprint. Here’s the story you haven’t heard., MIT Technology Review, 20 May 2025.
  12. Sasha Luccioni and Boris Gamazaychikov, AI Energy Score v2: Refreshed Leaderboard, now with Reasoning, Hugging Face, 4 December 2025.
  13. Jegham, Gamazaychikov and Luccioni, Lights, Camera, Carbon: Architectural Scaling Laws for Video Generation Energy Consumption, 2026.
  14. Casey Crownhart, In a first, Google has released data on how much energy an AI prompt uses, MIT Technology Review, 21 August 2025, and Justine Calma in The Verge, 21 August 2025.
  15. Li, Yang, Islam and Ren, Making AI Less “Thirsty”: Uncovering and Addressing the Secret Water Footprint of AI Models, 2025.
  16. Mistral AI, Our contribution to a global environmental standard for AI, 22 July 2025.
  17. Chatterji et al., How People Use ChatGPT, NBER Working Paper 34255, September 2025.
  18. Sundar Pichai, Google I/O 2025: From research to reality, May 2025, and I/O 2026: Welcome to the agentic Gemini era, May 2026.
  19. Patterson et al., Carbon Emissions and Large Neural Network Training, 2021, and Patterson et al., The Carbon Footprint of Machine Learning Training Will Plateau, Then Shrink, 2022.
  20. IEA, Energy and AI, April 2025, with the chapters on demand and supply and the full report (on GPT-4).
  21. Epoch AI, What did it take to train Grok 4?, September 2025, and How much power will frontier AI training demand in 2030?, August 2025.
  22. Morrison, Smith and Strubell, The Hidden Cost of Thinking: Energy Use and Environmental Impact of LMs Beyond Pretraining, 2026.
  23. Google, 2026 Environmental Report, June 2026.
  24. Microsoft, Responsibly building the AI future, on the 2026 sustainability report, 9 July 2026.
  25. William Stanley Jevons, The Coal Question, 1865, chapter 7, and Satya Nadella as reported by Fortune, 27 January 2025.
  26. IEA, Key Questions on Energy and AI, April 2026, and Our World in Data, How much energy do data centers and artificial intelligence use?, July 2026.
  27. US Department of Energy, DOE Releases New Report Evaluating Increase in Electricity Demand from Data Centers, December 2024, and EPRI, Powering Intelligence 2026, with Virginia as reported by the American Public Power Association, February 2026.
  28. Central Statistics Office, Data Centres Metered Electricity Consumption 2025, 7 July 2026.
  29. Constellation, Constellation to Launch Crane Clean Energy Center, 20 September 2024, and Crane Clean Energy Center; the agreements of Google, Amazon and Meta.
  30. Svenska kraftnät, Långsiktig marknadsanalys 2026, June 2026, and the Swedish Energy Agency, PM Framtida elbehov på länsnivå, 2025.
  31. Brookfield via MFN, 4 June 2025, and SVT Sörmland, Lokala kraftnätet klarar inte enorma AI-satsningen i Strängnäs, June 2025, and Svenska kraftnät varnar: nya elslukare kan tvinga upp elpriset, 29 November 2025.
  32. Microsoft, 3 June 2024, EcoDataCenter on Borlänge, September 2025, and Mistral AI, February 2026, and Google on Horndal, 2 June 2026.
  33. Stockholm Exergi, Heat recovery, read 14 November 2026, and Fortum, 6 May 2026.
  34. Swedish Tax Agency, Slopad skattenedsättning för datorhallar, December 2022.
  35. The Social Democrats’ demand as reported by TT, 18 March 2026, and the Government Offices of Sweden, Handlingsplan för Sveriges AI-strategi, 20 February 2026.
  36. Electricity Maps, Grid in Review 2025 for Sweden, Norway, France, Ireland, Germany and Poland, and Ember, Global Electricity Review 2026, April 2026.
  37. The EU Energy Efficiency Directive, Directive (EU) 2023/1791, Article 12, the EU AI Act, Regulation (EU) 2024/1689, Article 53 and Annex XI, the European Commission’s consultation on measuring AI’s energy use, April 2026, and its press release on ratings for data centres, September 2026.

Terms in this article

How we work: the articles are written with the help of AI, and every fact is checked against the sources listed above. Found a mistake? Write to hej@sibolaget.se.

More questions?

New to AI? Start with What is AI? The glossary explains the terms, and the front page has straight answers to the most common questions. Missing something? Write to us.