How Much Water Does AI Use? 0.26 mL to 47 mL a Prompt

In brief

A chatbot prompt uses between 0.26 mL of water, Google's figure for a median Gemini text prompt, and about 17 mL, a university estimate for an older GPT-3 request that also counts power plant water, as of November 2026. Sam Altman put ChatGPT at about 0.32 mL. What each number counts explains the gap.

How much water does AI use per prompt?

About a quarter of a milliliter to about 47 milliliters, depending on who measured and what they counted. This table lists each published figure, with the date and the boundary of what it covers.

A clear glass dropper holding a single drop above a small beaker on a black plinth, beside a terracotta Claude starburst, a brushed silver OpenAI knot and a rainbow glass Gemini star on their own plinths, in a charcoal studio under white rim light.
FigureWho and whenWater per promptWhat it countsMethod
Median Gemini Apps text promptGoogle, 21 Aug 2025, May 2025 data0.26 mLData center water, from energy per prompt times Google's fleet water usage effectiveness of 1.15 L per kWhMeasured on production traffic
Average ChatGPT querySam Altman, June 20250.000085 gallons, about 0.32 mL by our conversionScope not statedOne line in a blog post
400 token Le Chat replyMistral, with Carbone 4 and ADEME, July 202545 mLData center water plus upstream impacts such as making the serversLifecycle study
Medium GPT-3 request, US averageLi, Yang, Islam and Ren, arXiv 2304.03271, v5 March 202516.9 mL, of which 2.20 mL on site and 14.70 mL off siteOn site cooling water plus water used to generate the electricityModel: 800 words in, 150 to 300 words out, about 0.004 kWh
Same request, by locationSame paper7.11 mL in Ireland to 47.51 mL in Washington stateSame as aboveMicrosoft data center locations
Published prompt figures run from 0.26 mL to 47.5 mL
0.3 mL1 mL10 mL50 mLGoogle, GeminiAltman, ChatGPTRen, GPT-3 in IrelandRen, GPT-3 US averageMistral, Le ChatRen, GPT-3 in Washington

The newer company figures are the small ones. The Google and OpenAI numbers are for 2025 serving, while the university paper models GPT-3, a model from 2020, so it does not describe a model you use today. It's still the clearest public example of how much a power plant adds. Use it for that. AI data center energy consumption statistics has the matching energy figures, so they are not repeated here.

Does ChatGPT use water?

Yes. ChatGPT runs in data centers, and a data center that cools its chips with evaporation consumes water every time it works. Altman wrote that the average query "uses about 0.000085 gallons of water; roughly one fifteenth of a teaspoon." That converts to about 0.32 mL.

The line comes with no stated model or boundary for the water, so it is a claim you can quote but not audit. The university estimate for GPT-3, 16.9 mL on a US average, is about 53 times larger. Neither number is a typo. They count different things, and that's the whole story.

Does ChatGPT use water to cool down? Indirectly, yes. The chips sit in a data center, and the building rejects heat with chillers, cooling towers or both. Some of that water evaporates, and that part counts as consumed. A closed loop or an air cooled hall uses much less on site but more electricity, which moves water use to the power plant.

What are three real cases of AI water use?

  1. One day of Gemini. Google's 0.26 mL, times 1,000 prompts, comes to 260 mL, a little over one US cup. Altman's 0.32 mL gives 320 mL. The same 1,000 prompts at the university estimate come to 16.9 liters. That is our arithmetic on their figures.
  1. One request, two states. In the Ren paper the same GPT-3 request costs 7.59 mL in Texas and 47.51 mL in Washington. The on site part is 1.00 mL in Texas and 3.80 mL in Washington. Most of the gap is the off site part, 6.59 mL against 43.71 mL, which is the water behind the local electricity.
  1. One company, one year. Google said in its 2026 Environmental Report that it replenished about 7.7 billion gallons of water in 2025, roughly 78% of its freshwater consumption. That implies about 9.9 billion gallons consumed, our division, across offices and data centers. No per prompt figure can be multiplied up to that total, because the total covers all of Google's operations, not one product.

Why do AI water estimates differ by 100 times?

Five choices decide the number, and each one moves it by a factor of several.

  1. On site or off site. On site water cools the data center. Off site water is what the power plants evaporate to make the electricity. In the Ren paper the US average request used 2.20 mL on site and 14.70 mL off site, so the second part is about 87% of the total.
  2. Consumption or withdrawal. Consumption is water that does not come back. Google defines it as water input minus water returned. Withdrawal is water taken in, much of which returns. The Ren paper projects 4.2 to 6.6 billion cubic meters of global AI water withdrawal in 2027, a withdrawal figure that is not comparable to a consumption figure.
  3. Which model and how long a reply. The Ren request is 800 words in and 150 to 300 words out on GPT-3. Mistral used a 400 token reply. Google used the median prompt.
  4. Where and when. The paper's own table runs from 7.11 mL in Ireland to 47.51 mL in Washington.
  5. Scope of the lifecycle. Mistral adds upstream impacts such as manufacturing. Google counts energy for chips, host machines, idle capacity and overhead, and says its figure is "substantially lower than many public estimates."

The widest published pair is Google at 0.26 mL and the Washington case at 47.51 mL, about 180 times apart. Mistral at 45 mL against Google is about 170 times. Those ratios are our arithmetic. The honest reading is that the figures don't conflict. They answer different questions.

The often repeated "500 ml bottle" is the Ren paper's summary that GPT-3 would drink a 500 ml bottle for roughly 10 to 50 medium responses. It is a count of responses per bottle for a 2020 model, not a statement about every chatbot today, and the paper's own table spans about 10.5 to 70.4 responses per bottle.

How does a data center use water to cool AI chips?

Chips turn electricity into heat, and the building has to get that heat out. One common way is evaporative cooling, where warm water meets air and part of it evaporates. That evaporated part is the water the figures above call consumed.

The industry measure is water usage effectiveness, liters per kWh of electricity used by the computers. Google's paper used 1.15 L per kWh for its fleet. The Ren paper used 0.550 L per kWh on site for Microsoft's US average. Both are small next to the off site factor in the same paper, 3.142 L per kWh of electricity for the power plants. Cooling with chillers instead saves water on site but uses more electricity, which can shift the water to the power plant.

So a lower on site number does not by itself mean a lower footprint. Ask where the electricity comes from too.

How much water do AI data centers use in total?

Nobody publishes a clean total for AI alone, because AI shares buildings with every other cloud workload. The best public pieces are broader. The Environmental and Energy Study Institute (EESI) cites a federal report for an indirect water footprint of about 211 billion gallons in 2023 for US data centers, meaning water used at the power plants that supply them. That is EESI's report of the figure, and it excludes on site cooling water.

On site, EESI attributes to the Washington Post that a large data center can use up to 5 million gallons a day, about the demand of a town of 10,000 to 50,000 people. It also cites the Financial Times for about 2 billion gallons used in Northern Virginia in 2023, up 63% from 2019.

Those local numbers are why the question is not settled by a per prompt figure. A drop per prompt can still be a large share of a small town's supply when a hundred buildings sit in one county. The per prompt number tells you about efficiency. The local numbers tell you about impact. Training adds a one time cost: the Ren paper put on site water for training GPT-3 at about 700,000 liters in Microsoft's US data centers.

Is AI water use actually a problem?

It depends on the place and the water source, not the average. Per prompt, even the largest figure here is a few tablespoons. Locally, a data center in a dry county competing with farms and homes is a real planning issue. EESI describes Northern Virginia as one such case.

Efficiency is moving fast on the company side. That's real. Google wrote that its median prompt figures were "substantially lower than many public estimates," and its paper reports a 33 times drop in energy per prompt over one year. Nothing guarantees the same pace for the total, since more prompts can offset a lower cost per prompt. Google's electricity use rose 37% in 2025, per its 2026 report.

Read any claim with three questions: does it count on site water, power plant water or both, is it consumption or withdrawal, and which model and year. A figure that answers none of them is a slogan. Don't repeat it.

What questions should you ask an AI about its water use?

Models can summarize the literature, but they repeat old numbers, so make each question force the boundary and the date. Paste these into Claude, ChatGPT, Gemini or Whizi with web search on.

  1. "List every published per prompt water figure for a chatbot. For each, give the author, the date, the model, and whether it counts on site water, off site water or both. Use only the original paper or company post and link it."
  2. "Convert [figure] into liters for [number] prompts a day for a year. Show the arithmetic and the unit at each step."
  3. "Which of these figures is consumption and which is withdrawal? Quote the sentence in the source that says so, or write not stated."
  4. "I live in [region]. What is the water source for the local data centers and the local power grid, and which public report says so?"
  5. "Argue the strongest case that AI water use is a serious problem and the strongest case that it is not. Give a source I can open for every number, and say which side the evidence favors."
  6. "Find a figure in this text that claims AI uses a bottle of water per prompt: [paste]. Is it per prompt, per conversation, or per session, and for which model?"

Open the source behind every number the model gives you. How to fact check AI answers shows a two minute routine, and why AI gets things wrong explains why a confident old figure keeps coming back.

What mistakes ruin an AI water calculation?

Four mistakes cause most of the bad numbers in circulation.

  1. Comparing a per prompt figure with a per day or per year total. The fix: multiply prompts by the figure and state which figure.
  2. Mixing consumption with withdrawal. The fix: use consumption for "water used up" and keep withdrawal for supply questions.
  3. Using the GPT-3 number for a current model. The fix: label it as a 2020 model estimate, or use a 2025 company figure and say it counts less.
  4. Quoting one number with no boundary. The fix: say whether it counts the power plant, and give a range.

Where does Whizi fit when you want to check these figures?

Whizi is one app instead of a pile of single purpose tools, so you can put the same water question to Claude, GPT and Gemini and read only where they disagree. Web search is on by default in each chat. On Pro, Compare runs two models on one prompt in two columns, and Pro's first week is $1.99.

The free plan gives you 10 messages a day on Whizi AI, Auto and Claude Haiku, with no card. That is enough for the six questions above. It doesn't replace opening the original paper, and a model can still repeat the 500 ml figure with no date on it.

What do these figures not tell you?

They do not give the footprint of a whole conversation, an image, a video, a file analysis or a long reasoning run. Google's figure is a median text prompt, and it says the energy distribution is skewed, so the average is not the median. Altman's is an average with no stated method. The university paper is a model of a 2020 model, using Microsoft locations and a stated request size.

No company here published a water figure that we could verify for the Claude models, and we are not going to guess one. If a vendor publishes one, the same three questions apply: what is counted, consumption or withdrawal, and which model. A sustainability report is a company's own accounting, not an audit.

Workflow checklist
  • Write down which figure you used and whether it counts power plant water
  • Keep consumption and withdrawal apart
  • Label any GPT-3 number as an estimate for a 2020 model
  • Open the original paper or company post for every figure
  • Report a range, not a single number
Common questions

Frequently asked questions

How much water does one ChatGPT query use?

About 0.32 mL on Sam Altman's June 2025 figure of 0.000085 gallons, about one fifteenth of a teaspoon. A 2023 university study put a medium GPT-3 request at 16.9 mL once power plant water is added. The first has no stated method, and the second models an older model.

How much water does AI use per day?

No one publishes a per day total for AI alone. For your own use, multiply your prompts by a figure: 50 prompts a day is about 13 mL at 0.26 mL each, or about 0.85 L at 16.9 mL each. That is our arithmetic on the published figures.

Is the 500 ml bottle of water per AI prompt true?

No. The Ren paper said GPT-3 would use a 500 ml bottle for roughly 10 to 50 medium responses, which is about 10 to 50 mL a response including power plant water. Google measured 0.26 mL for a median Gemini prompt in 2025.

Does Gemini use less water than ChatGPT?

On the published figures, Google says 0.26 mL for a median Gemini text prompt and Altman says about 0.32 mL for ChatGPT, as of November 2026. They are close, but the methods are not equal. Google published its method and OpenAI gave one line.