A single ChatGPT query uses about 0.32 millilitres of water in the data centre, by OpenAI’s own June 2025 figure, roughly a fifteenth of a teaspoon. Google reports 0.26 millilitres for a median Gemini text prompt. Older estimates ran to half a litre per email because they counted the water behind the electricity, on 2023 hardware.
Three numbers get quoted for this question and they differ by a factor of a thousand, yet each of them is correct within what it counts. So the useful answer is a number with its boundary attached, and this post gives all three with theirs.
The three figures and what each one counts
OpenAI’s figure is 0.32 millilitres per query. In June 2025 Sam Altman wrote that the average ChatGPT query uses about 0.34 watt-hours of electricity and about 0.000085 gallons of water, which converts to 0.32 millilitres. That is the water the data centre itself consumes, mostly by evaporating it to carry heat away from the servers. OpenAI has not published the method behind the figure.
Google’s figure is 0.26 millilitres per prompt. Two months later it published a methodology paper for its own fleet. The median Gemini Apps text prompt uses 0.24 watt-hours, emits 0.03 grams of CO₂e and consumes 0.26 millilitres of water. Google counts the accelerator, the host machine, idle capacity and the data centre overhead, and reports that energy per prompt fell 33 times in a year as models and hardware improved.
The large figure is 519 millilitres per email. In September 2024 the Washington Post worked with researchers at UC Riverside to estimate what a 100-word email written by GPT-4 costs. The answer, a little over a bottle of water, included the water used at power plants to generate the electricity, and it reflected the hardware and models of 2023. The same research group’s 2023 paper had put GPT-3 at a 500 millilitre bottle per 10 to 50 medium responses, again depending on where and when the model ran.
So the small figures are on-site water on current hardware, and the large one is on-site plus upstream water on older hardware. Neither is wrong, but a company that wants a defensible number has to say which of the two it is reporting.
Where the water actually goes
Servers turn almost all their electricity into heat, and the cheapest way to move a lot of heat out of a building is to evaporate water. A data centre with evaporative cooling consumes water in proportion to the energy it uses, and the ratio is called water usage effectiveness, litres per kilowatt-hour of site energy. Lawrence Berkeley National Laboratory’s 2016 survey of US data centres put the average at 1.8 litres per kilowatt-hour; its 2024 report puts the 2023 average at 0.36, with hyperscale sites lower and AI-specialised sites higher, and the operators’ own figures range from Microsoft’s 0.27 to about 1.1 implied by Google’s per-prompt numbers. Air-cooled and closed-loop sites sit far below the average, and sites in hot, dry places sit above it, which is why the same model can be ten times thirstier in one region than another. The data-centre water post has the full set of figures, and the calculator below lets you pick the ratio.
The second draw is upstream. Most electricity still comes from plants that boil water to spin turbines and use more water to condense the steam, and that water is consumed on your behalf whether the data centre uses any or not. The Berkeley study found that this off-site water exceeds what data centres use on site. It is the reason the 519 millilitre figure is so much larger than OpenAI’s, and it is the part that depends on the grid rather than on the AI company.
What our own methodology says
CrbonFree measures AI usage per token rather than per query, so it can put a number on a specific exchange. Under methodology v1.2 a medium tier model charges 0.3 joules per input token and 0.5 joules per output token on the accelerator, scaled up by the host power and cluster utilisation factors and the facility’s power usage effectiveness to give site energy. Multiply by the 1.8 litre average and you get on-site water:
| Exchange | Model tier | Site energy | Water at 1.8 L per kWh |
|---|---|---|---|
| Short chat, 200 tokens in and 150 out | Medium | 0.18 Wh | 0.32 mL |
| Typical task, 1,000 tokens in and 300 out | Medium | 0.59 Wh | 1.06 mL |
| Typical task, 1,000 tokens in and 300 out | Small or distilled | 0.21 Wh | 0.38 mL |
| Typical task, 1,000 tokens in and 300 out | Reasoning, extended thinking | 1.36 Wh | 2.45 mL |
The short chat lands on 0.32 millilitres, the same figure OpenAI published, from an independent set of factors. The longer exchange is three times that, and a reasoning model doubles it again, because output tokens cost more than input tokens and reasoning models produce far more of them. That spread is the answer to "how much": it depends on the length of what you send and get back, and on which model answers.
Try your own numbers. The calculator runs the same factors for any exchange size, model tier and number of queries a day.
20
200 tokens in, 150 tokens out, uncached.
0.3 J per input token, 0.5 J per output token. PUE 1.2.
Water per query
0.32 mL
from 0.18 Wh of site energy for the exchange
Per day and per year
Take OpenAI’s 0.32 millilitres and a person who sends 20 queries a day.
| Scale | Per day | Per year |
|---|---|---|
| One person, 20 queries a day | 6.4 mL | 2.3 L |
| A team of 50 | 320 mL | 117 L |
| A company of 5,000 | 32 L | 11,700 L |
These are on-site figures. Add the water behind the electricity and, on the Berkeley finding, the totals more than double, with the exact amount depending on where the electricity comes from. Swap the short query for a 1,000-token task and the totals triple again. A company that runs AI at scale will not find any of this in its water accounting today unless someone meters the tokens.
How to get your own number
- Meter the tokens, per model, from the provider’s usage data rather than from spend.
- Convert tokens to site energy with per-tier factors and the facility’s power usage effectiveness. CrbonFree does this under methodology v1.2 and publishes every factor.
- Multiply site energy by a water usage effectiveness. Use the provider’s published figure where one exists and the 1.8 litre average where it does not, and say which you used.
- Report on-site water and upstream water as separate lines, the way the researchers do, so the number survives the first question an auditor asks.
What actually reduces it
Water follows energy, so anything that cuts energy per answer cuts water. Shorter prompts and shorter answers help directly. Prompt caching helps more than people expect: a cached input token costs 0.03 joules on the medium tier against 0.3 for an uncached one, a tenth of the energy. Choosing a small tier for tasks that do not need a large model cuts the site energy of a typical task from 0.59 to 0.21 watt-hours. And the UC Riverside team’s central point still stands: where and when a model runs changes the water it consumes, so the same prompt served from a cool, closed-loop site is a different number from the same prompt served from a desert.
Sources
- 01Sam Altman, The Gentle Singularity (June 2025)
The average ChatGPT query uses about 0.34 watt-hours and about 0.000085 gallons of water, roughly one fifteenth of a teaspoon.
- 02Elsworth et al., Measuring the environmental impact of delivering AI at Google Scale (August 2025)
A median Gemini Apps text prompt uses 0.24 watt-hours, 0.03 grams of CO2e and 0.26 millilitres of water, with a 33 times reduction in energy per prompt over twelve months.
- 03Li, Yang, Islam and Ren, Making AI Less "Thirsty" (2023, Communications of the ACM 2025)
GPT-3 consumes a 500 millilitre bottle of water for roughly 10 to 50 medium-length responses depending on where and when it runs; training it evaporated about 700,000 litres; global AI demand could withdraw 4.2 to 6.6 billion cubic metres of water in 2027.
- 04The Washington Post with UC Riverside, A bottle of water per email (September 2024)
A 100-word email with GPT-4 was put at 519 millilitres of water, counting the water used to generate the electricity.
- 05Shehabi et al., United States Data Center Energy Usage Report, Lawrence Berkeley National Laboratory (2016)
The source of the 1.8 litres per kilowatt-hour average water usage effectiveness for US data centres, and of the finding that water used at the power plant exceeds water used on site.
- 06Shehabi et al., 2024 United States Data Center Energy Usage Report, Lawrence Berkeley National Laboratory (December 2024)
US data centres used 176 terawatt-hours in 2023, 4.4% of the country’s electricity, heading for 6.7 to 12% by 2028.
Figures attributed to CrbonFree come from methodology v1.2, published in full with every factor and formula. Read the methodology.
About the author
Cory Bergh
Cory leads Crbon Labs, which originates its own climate projects and builds CrbonFree, the platform that measures the footprint of AI usage per token. He was previously VP of Technology and Innovation in the energy industry and holds a BComm, an MBA and the CFA.