Skip to main content
Jamie Watters

AI water use: the bottle of water per 20 ChatGPT questions was withdrawn by its own authors in 2023. It's still quoted in 2026

Jamie Watters

Operational resilience and AI delivery practitioner.

Published: 29 September 2026•9 min read
#ai-water#fact-check#data-centres#claim-hunter
Card headed 'The most quoted AI water number was withdrawn by its own authors in October 2023', with two dated quotes from the same paper: 6 April 2023, 'ChatGPT needs to drink a 500ml bottle of water for a simple conversation of roughly 20-50 questions and answers', and 25 October 2023, 'GPT-3 needs to drink (i.e., consume) a 500ml bottle of water for roughly 10-50 responses'. Footer: the first sentence is the one still quoted in 2026; source arXiv 2304.03271, versions 1 and 2.

The most repeated AI water number is that ChatGPT "drinks" a bottle of water for every 20 to 50 questions. Its authors withdrew that sentence in October 2023, six months after it went round the world. The change is in the paper's own version note (Li, Yang, Islam and Ren, arXiv 2304.03271, version history). The replacement names an older model, GPT-3, and a wider range, 10 to 50 responses. The sentence still in circulation is the first one. A May 2026 explainer gives "One ChatGPT conversation (20-50 exchanges): ~500ml", with no source (Oreate, 7 May 2026).

Two dated quotes from the same paper, one above the other. 6 April 2023, version 1: ChatGPT needs to drink a 500ml bottle of water for a simple conversation of roughly 20 to 50 questions and answers. 25 October 2023, version 2: GPT-3 needs to drink, that is consume, a 500ml bottle of water for roughly 10 to 50 responses. Below, what changed: the model, the range, and the change note the authors wrote on the arXiv page.

Every AI water number I traced has the same pattern: the document gives a scope, a range and a date, and the number that travels has lost all three. Here is each one beside its document, and a human beside the arithmetic.

How much, with a person standing next to it

A brain runs on about 20 watts, its metabolic budget rather than an electricity bill (Balasubramanian, PNAS, 2021). Over a day that's 480 watt-hours. Google estimates that the median text prompt to its Gemini app used 0.24 watt-hours in May 2025 and "consumes 0.26 milliliters (or about five drops) of water" (Google, August 2025). That counts the water used at the data centre, worked out from Google's fleet-wide average, so the power station isn't in it (Google 2026 report, footnote 50). Run Gemini prompts at a brain's power for a day, about 2,000 of them, and they use up about half a litre of water on site. Google calls it a point-in-time estimate for one product's median prompt, not every prompt.

On the human side, the US Geological Survey puts domestic water use at 82 gallons per person per day in 2015, about 310 litres (USGS, 2018). That is water delivered to the home, not water used up. Most of it goes back down the drain, and the USGS did not estimate how much a household consumes. So this is half a litre consumed against 310 litres delivered, and the label matters. The figure of about 3,800 litres a day per person you may have seen, the water behind food included, is 74 per cent rain falling on crops (Hoekstra and Mekonnen, PNAS, 2012). Rain on a wheat field and evaporation from a cooling tower don't belong in the same bar chart, so it isn't in this piece.

Two limits. First, watts against watts is not thinking against thinking, and no source I found puts a number on how many prompts equal a day of human thought. Second, the half litre counts only the data centre. The power station feeding it consumes about 3.1 litres per kilowatt-hour for average US electricity, on the UC Riverside paper's consumption figure (Li and colleagues, 2025 revision). Add it and the half litre becomes roughly two litres a day, still against 310 delivered.

And one thing no average can show: where the water is drawn. Half a litre from a full reservoir and half a litre from a stressed aquifer are not the same half litre. The bank note behind the scariest total says so too, as you'll see. That is a local argument about particular basins, which averages can't settle. This piece is about what the numbers count.

The measured total, for scale: Google's data centres consumed 10,523 million gallons of water in 2025, about 39.8 billion litres, on 42,415,800 megawatt-hours of electricity (Google 2026 Environmental Report). That's about 0.94 litres per kilowatt-hour on site, Google's own estimate, and a count of last year. A rising total isn't answered by a small per-prompt figure; both can be true at once.

Who said it, and where it came from

These numbers reached six million views together in Hank Green's video "Why is Everyone So Wrong About AI Water Use??", uploaded 8 December 2025, at 5,990,306 views when my claim-hunting script pulled its captions on 27 September (the video). It is a debunking video: at 18:13 it says "Corn uses orders of magnitude more water than AI." I read the captions, not the film; the quotes below are the captions, each with a timestamp link. What the video doesn't do is trace each number to its document, and one of its own figures is wrong. That's the job here.

The teaspoon and the five drops are close in size, and only one of them says what it counts. At 0:00 the video opens on Sam Altman's line that "the average query uses about 0.34 watt-hours" and "about 0.000085 gallons of water; roughly one fifteenth of a teaspoon" (The Gentle Singularity, June 2025). That's 0.32 millilitres, no model, no method. Google's five drops come with the method above, at the data centre only (Google 2026 report, footnote 50). Put 0.32 next to 0.26 and they are within a quarter of each other. They look like the same kind of count, and only Google has said so.

The bottle was a 2023 estimate at seventeen times today's energy per prompt, and seven-eighths of it was the power station. The Riverside paper assumes 0.004 kilowatt-hours per request, 4 watt-hours. For average US electricity that gives 2.2 millilitres on site plus 14.7 millilitres at the power station, 16.9 millilitres a request (Li and colleagues, 2025 revision). Google's May 2025 prompt uses 0.24 watt-hours, about a seventeenth of that, and Google says the median prompt's energy fell 33-fold in the twelve months before (Google, August 2025). So the bottle wasn't absurd when it was written. It was an estimate for a 2020 model on 2023 assumptions, and most of it was water the newer per-prompt figures don't count.

Two bars to one scale, water per request. The Riverside paper's 16.9 millilitres for a GPT-3 request at 4 watt-hours, split into 2.2 millilitres on site and 14.7 millilitres at the power station. Google's 0.26 millilitres for a median Gemini prompt at 0.24 watt-hours, on site only, drawn as a hairline. Caption: seventeen times the energy, and seven-eighths of the bottle is the power station.

And the authors changed the sentence, six months after the April 2023 coverage. Version 1, 6 April 2023: "ChatGPT needs to 'drink' a 500ml bottle of water for a simple conversation of roughly 20-50 questions and answers, depending on when and where ChatGPT is deployed" (v1). Version 2, 25 October 2023: "GPT-3 needs to 'drink' (i.e., consume) a 500ml bottle of water for roughly 10-50 responses, depending on when and where it is deployed" (v2), with a version note reading "new estimates of GPT-3's water footprint based on Microsoft's new WUE and PUE data". The current version, 26 March 2025, adds "medium-length" to the responses (v5). The low end of the range halved, which doubles the worst-case water per response, from 25 millilitres to 50. I couldn't find a report of the change, and the 2023 sentence is the one a 2026 explainer still prints.

The trillion litres came from a bank note that says the total is not the point. At 0:12 the video says a Morgan Stanley projection "could reach around a,000 billion lers [1,000 billion litres] by 2028. That's a trillion liters, an 11fold increase from 2024 estimates." The report itself is for the bank's clients and I haven't read it. The newswire that broke it quoted it directly and kept three things the headlines dropped (ANI via The Tribune, 8 September 2025). The figure: "approximately 1,068 billion liters by 2028 (our base case)", a base case, not a midpoint. The range: "between 637 billion litres and 1,485 billion litres annually by 2028". And the point: more than half of the world's top data-centre hubs "are in areas already facing medium basin physical risk ... to threats from drought, flooding, and declining water quality." By the time the number reached the video it was "a trillion", 11-fold, with the range and the warning gone. Neither newswire copy prints the 2024 starting point in litres; divide 1,068 by 11 and it's about 97 billion, my arithmetic and not theirs.

Four boxes in a chain showing how the trillion-litre figure travelled. The Morgan Stanley research report, September 2025, for clients: a 2028 base case of 1,068 billion litres, a range of 637 to 1,485, and a line that more than half the top data-centre hubs sit in basins at medium risk. The ANI newswire, 8 September 2025, quoting all three but not the 2024 starting point in litres. The headlines that week: 1 trillion litres, 11-fold, range and local warning gone. The video, 8 December 2025, at 0:12: a trillion litres, 11-fold, naming Morgan Stanley, no range. Footer: the document says the total is not the point; the headlines kept the total.

Why the numbers disagree: three mechanisms

Withdrawal or consumption. Withdrawal is water taken from a river or an aquifer. Consumption is the part that evaporates or is otherwise not returned. US thermoelectric power stations withdrew 133 billion gallons a day in 2015 and consumed 3 per cent of it (USGS Circular 1441). The video makes this point at 11:50: "It's about 2 to 3%."

Forecast or fact, and how old. Google's 39.8 billion litres is a count of last year. Morgan Stanley's trillion is a base case for 2028 with a range from 637 to 1,485 billion litres (ANI via The Tribune). Quoting the base case as "a trillion" is how a scenario turns into a fact. And an estimate has a date: the bottle is a 2023 estimate for a 2020 model.

Total water or freshwater. At 11:12 the video says: "according to the US Geological Survey, electricity generation accounts for, get this, 40% of all freshwater withdrawals in the United States." The USGS says thermoelectric power took 41 per cent of all withdrawals in 2015, saline water included, and 34 per cent of freshwater withdrawals (USGS Circular 1441). Ten years earlier the freshwater share was 41 per cent (USGS Circular 1344). So the 40 is either the total share mislabelled or the 2005 figure repeated in 2025; the video lists no sources, so I can't tell which. The latest USGS freshwater share, for 2015, is 34.

Three questions, next time you meet a number

First: withdrawal or consumption? If the source doesn't say, you have a number without a label.

Second: where does the count stop? The cooling at the data centre, the power station feeding it, or the factory that made the chips? Altman's teaspoon doesn't say. Google's five drops stop at the data centre. Riverside's bottle includes the power station, and that is most of it.

Third: measured or forecast, from when, and what's the range? A 2015 measurement, a 2023 estimate and a 2028 scenario don't carry the same weight. If a forecast has a range, quote the range. If an estimate has a date, say the date.

Pass: the number you're about to repeat has an answer to all three, from the source, not from the person who told you about it. Fail: any one is missing. Then say "reportedly", or say nothing.

Run on the five numbers in this piece:

  • Five drops a prompt (Google). Repeat it, with its label: on site only, the median Gemini text prompt, Google's own estimate for May 2025.
  • A fifteenth of a teaspoon a query (Altman). "Reportedly." No model, no method, no scope.
  • A bottle per 20 to 50 ChatGPT questions (Riverside, 2023). Don't. Its authors replaced it in October 2023 with GPT-3, 10 to 50 responses, power station included.
  • A trillion litres by 2028 (Morgan Stanley). A base case, cooling and electricity, range 637 to 1,485 billion. Quote the range and the year, and the line about the basins, which is the bank's own point.
  • 40 per cent of freshwater (the video). Mislabelled. 34 per cent of freshwater withdrawals in 2015, and withdrawals, not consumption.

Sources

No paywall, no sponsors. If this saved you some time, you can buy me a coffee.

☕ Buy me a coffee

Get the next one in your inbox

I build with AI in the open and write up what held and what didn't. Real numbers, the failures before the wins.

Share this post

← Previous Post

Jev is a narrow tool with four clear uses. Greg Isenberg's $10m framing is not one of them.

Greg Isenberg's episode on Jev promised $10m businesses. His own guest says use it for routing and keep it away from your money. Here is each claim against the data, and the four conditions under which the tool earns its place.

Next Post →

OpenAI's Texas data centre is built to barely use water. The power plants it runs on use 300 times more water than it does

The company building OpenAI's Stargate campus in Abilene says a 200 megawatt block uses 910,000 gallons of water a year on its website, and about 2 million in…

← Back to all posts