AI data center energy consumption statistics 2026: use at data centers built for AI rose 50% in 2025

Data centers used 485 TWh in 2025, and AI drove the growth

Each of these AI data center energy consumption statistics links to the agency, lab or company that published it. The date is the period a figure covers. For a company's per prompt figure, it's the month the company published it.

Don't add these rows up. The IEA counts every data center, including the ones that run email, video and banking, and it splits out the ones packed with AI chips. The per prompt rows cover one chatbot answer, and none of them include training. Google's are text only, and OpenAI didn't say what its average covers.

Most of the growth is recent. The IEA's chart data puts data center use at 269 TWh in 2020, 416 TWh in 2024 and 485 TWh in 2025, so it's grown about 1.8 times in five years. The IEA's own forecast for 2030 is about double 2025. For data centers built for AI, it's triple.

How much electricity does ChatGPT use a day?

ChatGPT's prompts used about 850 MWh of electricity a day, by our count from two figures OpenAI gave in mid 2025. That's about 310 GWh a year. It's roughly what 29,000 US homes use in a year, at the US Energy Information Administration's average of 10,791 kWh per home.

That's small next to the grid as a whole. It's about 0.06% of the 485 TWh that data centers used in 2025. The IEA ran a similar check in April 2026, and it found that 10 billion queries a day at a generous 1 Wh each would come to about 3.6 TWh a year.

It's also more than a widely repeated 2024 estimate. In March 2024, The New Yorker reported an estimate that ChatGPT handled about 200 million requests a day using more than half a million kWh, and reporters turned that into 17,000 US households. OpenAI's own numbers give a daily total up to 1.7 times larger. Prompt volume grew more than twelvefold, while each prompt uses about a seventh of the 2.5 Wh the old estimate implied.

Training one big model can use about as much. Epoch AI estimated in September 2025 that training xAI's Grok 4 took about 310 GWh of electricity, the same as our yearly ChatGPT figure. Epoch built it from Grok 3's known resources, not from an xAI disclosure, and the two figures are counted differently.

Treat the 850 MWh as a snapshot from mid 2025, not as a measurement. Altman didn't say what his 0.34 Wh includes, such as cooling, idle servers, or image and voice requests. OpenAI hasn't published a daily prompt count since July 2025, and its user numbers have grown since then, as our ChatGPT statistics page shows. ChatGPT is still the biggest chatbot on every count we track in AI chatbot market share.

How much energy does one AI prompt use?

A typical prompt to a big chatbot uses about 0.24 to 0.34 Wh, on the figures Google and OpenAI published in 2025. Google put its median Gemini Apps text prompt at 0.24 Wh in August 2025, which it said is like watching TV for less than nine seconds. Sam Altman wrote in June 2025 that the average ChatGPT query uses "about what an oven would use in a little over one second."

What a figure counts can change the answer a lot. Google's two numbers for the same prompts differ 2.4 times, only because the bigger one adds host machines, idle capacity and data center overhead such as cooling. Google's is also a median and Altman's is an average, and a few long prompts pull an average up. So don't call 0.34 Wh 40% more than 0.24 Wh. They aren't measured the same way.

Each prompt is getting cheaper to run. Google said the energy of its median Gemini text prompt fell 33 times over 12 months, and its carbon fell 44 times. Microsoft Research cut its own estimate from 0.34 Wh to 0.31 Wh when it revised its paper in June 2026. We found no per prompt energy figure from Anthropic, xAI or Meta as of October 2026. Gemini's user numbers are on our Gemini statistics page, and Anthropic's are on Claude statistics.

Which AI tasks use the most energy?

Video, reasoning and very long inputs use the most energy, often hundreds of times a short text reply. The IEA said in April 2026 that video, reasoning and agent tasks can use hundreds or thousands of times more energy per query than simple text. A high resolution image can take about 10 times a short reply, it added.

Video is the extreme case. MIT Technology Review measured some open models itself in May 2025. One response from Meta's Llama 3.1 405B took about 6,706 joules once it allowed for cooling and other demands, or 1.86 Wh. A standard image from Stable Diffusion 3 Medium took about 0.63 Wh. A five second video from CogVideoX took about 3.4 million joules, or 944 Wh. That's as much as running a microwave for over an hour.

Long inputs cost more, too. Epoch AI estimated in February 2025 that if you give GPT-4o an input of 10,000 tokens, the query uses about 2.5 Wh, and that one with 100,000 tokens uses almost 40 Wh. Microsoft Research puts a reasoning length query, with 15 times the tokens, at 3.91 Wh. That's 13 times a typical one. Our LLM API pricing statistics show what those tokens cost.

How much water does an AI prompt use?

A chatbot prompt uses about a quarter to a third of a milliliter of water, on Google's and OpenAI's 2025 figures. Google put its median Gemini prompt at 0.26 mL, about five drops. Altman put the average ChatGPT query at 0.000085 gallons. That's about 0.32 mL by our conversion, or in his words "roughly one fifteenth of a teaspoon."

Mistral's figure is far higher, at 45 mL for a 400 token reply from Le Chat, because it counts more. It adds upstream impacts, such as making the servers, to the data center's own use. Mistral did the July 2025 study with the consultancy Carbone 4 and ADEME, France's ecological transition agency. It said training Mistral Large 2 and running it for 18 months used 281,000 cubic meters of water.

At Altman's rate, ChatGPT's 2.5 billion daily prompts come to about 212,500 gallons of water a day. That's our arithmetic. Google said in June 2026 that its water projects put back about 7.7 billion gallons in 2025, which it called roughly 78% of its freshwater use that year.

How much of US electricity goes to data centers?

US data centers used 4.4% of the country's electricity in 2023, or 176 TWh, Lawrence Berkeley National Laboratory found in a December 2024 report for the Department of Energy. That was up from 58 TWh in 2014. The lab's 2025 update, published in June 2026, puts the share at 11.8% by 2030, or 649 TWh. Its range runs from 9.5% to 15.3%. The share would more than double.

The Electric Power Research Institute (EPRI) put it at 9% to 17% by 2030 in its February 2026 report, up from 4% to 5% today. That range is 60% higher than its earlier scenarios. The US is the biggest market by far. It used 45% of the world's data center electricity in 2024, ahead of China at 25% and Europe at 15%, the IEA said in April 2025. The IEA expects data centers to make up nearly half of US electricity demand growth to 2030.

How does AI energy consumption compare with other industries?

AI is still a small user of the world's power, but it's one of the fastest growing. Data centers of every kind used 1.5% of the world's electricity in 2025, and the IEA expects about 3% by 2030. They'll make up slightly less than 10% of the growth in world electricity demand to 2030, by the IEA's count. Our World in Data puts AI alone at about 0.5% of the world's electricity in 2025, based on IEA figures.

The 2030 figure is big in country terms, though. The IEA wrote in April 2025 that its 945 TWh forecast for 2030 is slightly more than all of Japan uses today. Locally, data centers already took as much as 20% to 30% of electricity demand in the regions where they cluster in 2025, the IEA said.

Emissions go up with the power. The IEA expects data center emissions to double to about 350 million tonnes a year by 2035. That would still be about 2% of the power sector's emissions. The IEA also notes that a server rack the size of a fridge could have the peak power demand of 65 households by 2027.

How fast is Big Tech's electricity use growing?

Google's electricity use rose 37% in 2025, which it called the largest growth in its history, in its 2026 Environmental Report on 30 June 2026. Its operational emissions fell 2%. Its supply chain emissions rose 25%. Google also signed deals for more than 12 GW of new clean energy in 2025.

Microsoft used 37.0 TWh of electricity in its fiscal 2025, which ended in June 2025, per its 2026 environmental data fact sheet. That's up 24% from 29.8 TWh a year earlier and 3.4 times its fiscal 2020 use. Its total emissions rose 25% in the year, to 20.29 million tonnes. That's about 58% above its 2020 baseline.

The IEA estimated that Amazon Web Services, Microsoft, Alphabet and Meta used 110 TWh between them in 2024, which was up 20% on the year before. The largest tech companies' capital spending passed $400 billion in 2025. The IEA expects that to jump by another 75% in 2026.

Which nuclear and gas deals are powering AI?

Meta is behind the two biggest figures in the table below. It announced up to 6.6 GW of nuclear power by 2035 on 9 January 2026. Then, on 27 March 2026, Fortune reported that Meta would fund seven more gas plants for its Louisiana campus, for 10 plants and 7.5 GW in all.

Gas is growing next to the nuclear deals. The IEA expects 15 to 27 GW of gas power on data center sites by 2030, mostly in the US. Orders for gas turbines jumped 70% in 2025, it said. Most of the new reactors in these deals aren't due before 2030.

How we compiled these numbers

We opened every figure at its source on 7 October 2026. Those are the IEA's reports from April 2026 and April 2025 and its chart data, plus Lawrence Berkeley National Laboratory, EPRI and the US Energy Information Administration. We also used Epoch AI, Microsoft Research, MIT Technology Review's own tests and the companies' posts and reports. Where a company told a reporter, we name the outlet. The 850 MWh, 310 GWh, 29,000 homes, 212,500 gallons and 0.32 mL figures are our arithmetic, with the inputs shown.

We left out what we couldn't open at its origin. That cut a widely repeated 2024 figure for AI servers, Google's data center total for 2025, and EPRI's state shares, where two write ups of the same release disagree. The IEA's report body says data centers grew "more than 15%" in 2025. We use the 17% and 485 TWh in its summary, which match its chart data. We'll update this page when the IEA, Google or OpenAI publishes new figures.

How to cite these AI data center energy consumption statistics

You're free to use any figure or table on this page, as long as you link back to it. Please credit the original source for the numbers that came from someone else, too. A plain credit line like this one works:

Source: Whizi Research, "AI data center energy consumption statistics 2026: use at data centers built for AI rose 50% in 2025", whizi.io/resources/ai-data-center-energy-consumption-statistics, updated October 2026.

More numbers from the same series: AI privacy statistics, AI in education statistics and AI hallucination statistics.