How Much Water Does One AI Question Use?

weirdly accurate.

How Much Water Does One AI Question Use?

How much water does one AI question use?

This AI water and energy test asks you to guess, before you see any figures, how much electricity and water one AI question uses, along with an AI image and three everyday yardsticks. Then it shows the published figures: Google’s and OpenAI’s own 2025 per-question numbers, and the much-quoted 2023 estimate by Li, Yang, Islam and Ren behind the “bottle of water” line. They count different things, and the result sheet says what each one includes.

This test turns those figures into six guesses on a lab bench. You charge a glass battery for electricity or fill a beaker for water, on a log scale from 0.01 to 1,000 Wh or mL, while a live readout turns your guess into everyday terms such as how long a 10 W LED bulb would run. Three questions are about AI: one text question’s electricity, its cooling water, and one generated image. Three are everyday yardsticks: a web search, a minute of streaming video and a litre boiled in a kettle. Each answer is a published range, and a guess counts if it lands within a set multiple of it, ×4 on the first question tightening to ×1.5 on the last. The result tells you whether you overestimate or underestimate AI, and by how many times.

It is not a calculator for your own footprint. The 2025 company figures describe a typical short text prompt, they are self-reported and nobody outside has audited them, and OpenAI has not published its method. Long inputs and heavier models use more, and Google’s figure leaves out training the model, the network and your own device. There is no percentile: nobody has measured how people usually guess these numbers, so you are scored only against the published ranges.

What the test measures

One thing: how well your sense of AI’s electricity and water use matches the published figures.

The task

  • Six guesses: an AI text question’s electricity and cooling water, an AI image, a web search (Google’s 2009 figure), a minute of streaming video including your screen and router, and a litre boiled in a kettle
  • Drag up or down on a glass battery (Wh) or a beaker (mL), or use the bar; the scale runs from 0.01 to 1,000 on a log scale
  • A live yardstick while you set it: a 10 W LED bulb’s running time, a share of a phone battery, drops, teaspoons or glasses
  • Each round says how close counts, from within ×4 of the published range on the first to within ×1.5 on the sixth, and shows the range right after you lock in

Your result

  • How many of the six guesses landed within range
  • Whether you overestimate or underestimate AI, and by how many times, averaged over the three AI questions; a separate line for the three everyday things
  • Your guess against the measured range and the figure that spread online, for each AI question
  • A table of who measured what: Google, OpenAI, Epoch AI, Li et al., Mistral, Luccioni et al. and MIT Technology Review, with what each one counts and leaves out
  • A shareable card with your score only, and a link that challenges a friend to guess closer

Before you start

  • About two minutes
  • Works on a phone or a computer; the bench is 3D, so the browser needs WebGL
  • Sound is optional: only short effects, with a mute button
  • Free, no account

AI water and energy FAQ

Is this test free? Do I need to sign up to see my result?

Yes, it’s free, and there’s nothing to sign up for: no account, no email, no payment. Your result appears on this page the moment you finish, and you can share it or play again straight away.

How much water does ChatGPT use per question?

The test asks you to guess that first, then shows OpenAI’s and Google’s own 2025 figures side by side. Only OpenAI can measure ChatGPT’s use, and it has given a figure without a method: Sam Altman stated an average per-query electricity and water figure in his June 2025 post “The Gentle Singularity”. Outside estimates vary with what they count, such as cooling water only or also the water used to generate the electricity.

Is it true that ChatGPT drinks a bottle of water for every 20 questions?

One of the test’s six rounds asks exactly that: guess how much cooling water one text question uses, and the measured range appears straight after; the result sheet sets it beside the bottle line. That line comes from Li, Yang, Islam and Ren’s 2023 paper “Making AI Less Thirsty”, an estimate for GPT-3 rather than a measurement, revised in 2025. The sheet also explains what the estimate assumed and counted, and how it compares with Google’s and OpenAI’s figures.

Why do estimates of AI’s energy and water use disagree so much?

Estimates of AI energy and water use disagree mostly because they draw the boundary in different places. Google’s 2025 study counts the AI chips, the host machines, idle spare capacity and building overhead, but not the network, your device or training. Mistral’s July 2025 figure for its Le Chat assistant is a life-cycle number that also counts making the servers. The model, the length of the prompt and the year matter too.

How is the AI water and energy test scored?

The AI water and energy test scores each guess against a published range on a log scale. A guess counts if it lands within that round’s multiple of the range, from ×4 on the first round to ×1.5 on the sixth; a miss is reported as how many times too high or too low. The headline averages the three AI questions, and two times or more counts as overestimating or underestimating. There is no percentile, because there is no data on how people usually guess.

How much electricity do data centres use in total?

About 415 TWh in 2024, around 1.5% of the world’s electricity, according to the International Energy Agency’s April 2025 report “Energy and AI”. The IEA expects about 945 TWh by 2030. Not all of that is AI, and the total depends on how much everyone uses it, which a per-question figure cannot show.

How long does the test take, and is anything I enter sent anywhere?

The test takes about two minutes, is free and needs no account. Your guesses are worked out in your browser and are not sent to a server. If you share a challenge link, it carries your name, your score and which rounds you got right after the # in the address, so a friend’s page can compare.

Where the numbers come from

Every answer is a range between published figures, checked against the original texts on 22 October 2026; company figures are self-reported. The kettle range is the physics of heating a litre by 80 °C up to the German Blue Angel kettle label’s ceiling. There is no percentile and the site does not publish its own players’ data.