How much electricity does it take to ask an AI chatbot a question? According to OpenAI’s chief executive, Sam Altman, not much. “The average query uses about 0.34 watt-hours,” he wrote in a blog post in June 2025, “about what an oven would use in a little over one second, or a high-efficiency lightbulb would use in a couple of minutes.”
That is a company estimate, and OpenAI has not published how it was worked out. But independent estimates land in the same range, and the number is small next to everyday appliances. The cost becomes real when you multiply it by billions of questions a day. Our 44-second explainer above sets out the comparison; this article gives the sources and the caveats.
An independent estimate lands close
The research group Epoch AI made its own estimate in February 2025, before Altman’s figure appeared. It found that “typical ChatGPT queries using GPT-4o likely consume roughly 0.3 watt-hours”, about a tenth of an older and widely quoted 3 Wh figure. Epoch said it had tried to err on the side of higher energy use, and that different assumptions could put the figure anywhere from about 0.1 Wh to between 1 and 4 Wh.
Google has published a figure for its own chatbot. In August 2025 it said that “the median Gemini Apps text prompt uses 0.24 watt-hours (Wh) of energy”. Google’s method counts more than the chips doing the work: it includes machines kept idle for spare capacity, the servers’ own processors and memory, and data-centre overheads such as cooling.
For comparison, Google said in 2009 that a web search used “0.0003 kWh of energy per search”, or 0.3 Wh. That figure is 16 years older and was calculated differently, so it is a rough reference point rather than a like-for-like test.
Small next to a kettle or a phone charge
Household comparisons help. Heating a litre of water from tap temperature to boiling takes about 93 Wh of heat, and a kettle running at about 90% efficiency uses roughly 100 Wh to do it. At 0.34 Wh each, that is about 300 AI questions per litre boiled.
The US Environmental Protection Agency puts a full smartphone charge at “0.019 kWh”, or 19 Wh. That is equal to about 55 questions.
Billions of prompts add up
The scale is what changes the picture. In July 2025, The Verge reported that ChatGPT receives more than 2.5 billion requests a day, a figure first obtained by Axios and confirmed by an OpenAI spokesperson.
TSN multiplied the two. At 0.34 Wh, 2.5 billion prompts is about 850 megawatt-hours a day, or about 310 gigawatt-hours a year. Ofgem, the energy regulator for Great Britain, uses 2,500 kWh a year as the typical electricity use of a medium-use home from 1 July 2026, down from 2,700 kWh. On that basis, a year of ChatGPT prompts would use about as much electricity as around 120,000 typical British homes.
That is TSN’s order-of-magnitude estimate, not a measurement. It treats every prompt as an average one, and it mixes company figures from 2025 with Ofgem’s 2026 value.
Data centres are a bigger story than chatbots
The International Energy Agency looks at the wider system. In its Energy and AI report, it says data centres used about 415 terawatt-hours in 2024, around 1.5% of the world’s electricity, and projects that will more than double to around 945 TWh by 2030. Those figures cover all data centres, not just AI, though the IEA calls AI the most important driver of the growth.
By TSN’s arithmetic, a year of ChatGPT prompts at Altman’s figure would be about 0.07% of the IEA’s 2024 data-centre total. Governments are already writing rules for the buildings that serve that demand, from South Korea’s faster permits for AI data centres to Donald Trump’s call for big data centres to make their own power. The harder question is who pays for that power.
Every number here is an estimate made a different way
No shared standard lies behind the figures in this article. Altman’s number has no published method. Epoch’s is a modelled estimate built on assumptions about chip use and model size, which OpenAI keeps private. Google’s counts overheads that others may leave out. Epoch itself says that “more transparency from OpenAI and other major AI companies would help produce a better estimate”.
The question also matters. Epoch estimates that a long input of about 10,000 tokens, roughly a short paper, raises the cost to around 2.5 Wh, and that a very long input of 100,000 tokens could need “almost 40 watt-hours”. It says reasoning models, which write out long chains of working, “may be substantially more energy-intensive”. All the per-query figures here are for text; none of these sources gives a figure for generating images or video.
Efficiency gains are real, and so is growth
The encouraging news is that the cost of a single question appears to be falling fast. Epoch’s estimate is a tenth of the older 3 Wh figure, and Google says the energy used by its median Gemini text prompt fell 33-fold over a recent 12-month period. A question that costs a fraction of a second of oven time is not a reason for guilt over everyday use.
The honest caveat is that efficiency and volume are moving at the same time. Cheaper questions tend to mean more questions, longer answers and heavier tasks, and the IEA’s forecast is a reminder that the total is still rising. The useful test for AI’s energy claims is disclosure: published methods, comparable measurements and figures for the heavier uses, not just the average query.
Sources
- Sam Altman, “The Gentle Singularity”, 10 June 2025
- Epoch AI, “How much energy does ChatGPT use?”, 7 February 2025
- Google Cloud, “Measuring the environmental impact of AI inference”, 21 August 2025
- Google, “Powering a Google search”, 11 January 2009
- US EPA, Greenhouse Gas Equivalencies Calculator: calculations and references
- The Verge, “OpenAI says ChatGPT users send over 2.5 billion prompts every day”, 21 July 2025
- Ofgem, Review of typical domestic consumption values (decision published 27 May 2026)
- Ofgem, Review of typical domestic consumption values: decision, 27 May 2026 (PDF)
- IEA, Energy and AI, executive summary
Related stories
- South Korea to Cut Permits and Parking Rules for AI Data Centres
- Trump Says Big Data Centres Are ‘Mandated by Me’ to Make Own Power; No Order Yet
- Who Pays for AI’s Power? Duke’s Data-Centre Deal, Google in Indiana and the Rush to On-Site Gas
Follow @tsnmediahub on X for breaking AI news as it happens.
Follow on X
