How much water and electricity does one AI prompt actually use?
Published answers to this differ by a factor of about sixty. The gap is not mainly a disagreement about AI. It is a disagreement about where you stop counting.
Key figures
Google measured its median Gemini text prompt at 0.24 Wh of electricity and 0.26 mL of water.
A researcher whose work Google cited puts a comparable prompt nearer 15 mL once the water used to generate the electricity is included.
Both can be correct. One counts water inside the building, the other counts water at the power station as well.
The published numbers
Three figures are quoted more than any others, and they are not measuring the same thing.
| Source | Energy per prompt | Water per prompt | What the water covers |
|---|---|---|---|
| Google, median Gemini text prompt | 0.24 Wh | 0.26 mL | Cooling inside the data centre |
| Sam Altman, average ChatGPT query | 0.34 Wh | about 0.32 mL | Not stated |
| Ren and colleagues, GPT-4 style prompt | not stated | about 5 mL | Cooling inside the data centre |
| Ren and colleagues, same prompt | not stated | about 15 mL | Data centre plus the power station |
Google also reports that the AI accelerator itself accounts for only 58% of the energy a prompt consumes. The rest is host processors and memory, idle capacity held ready, and the overhead of running the building. Any estimate that measures the chip alone is therefore low by roughly two fifths before anything else is argued about.
Why they differ by sixty times
Electricity has to be generated, and most generation is thermal: something is burned or fissioned, water is boiled, steam turns a turbine and then has to be condensed. That condensing evaporates water, and it happens at the power station, not at the data centre.
So a data centre operator reporting water can honestly report only what its own cooling towers evaporate. A researcher asking what a prompt costs the water system has to include the power station too. The International Energy Agency has put the indirect share of data centre water consumption at around three fifths of the total.
The dispute is sharper than a footnote. Shaolei Ren, whose earlier work Google cited, objected publicly that Google compared its own in-building figure against his in-building-plus-power-station figure, which is not a like-for-like comparison.
What it adds up to
The IEA puts data centre electricity at roughly 415 TWh in 2024, about 1.5% of world consumption, growing to roughly 945 TWh by 2030 in its base case, just under 3%. That is growth of about 15% a year, more than four times the rate of everything else on the grid.
Training is a separate line. Ren and colleagues estimated that training GPT-3 in Microsoft's United States data centres evaporated 700,000 litres of clean freshwater, and projected global AI water withdrawal of 4.2 to 6.6 billion cubic metres by 2027.
So what is the answer
For one text prompt today, somewhere between a quarter of a millilitre and roughly fifteen millilitres of water, depending entirely on whether you count the power station, and about a third of a watt-hour of electricity. On the low reading a prompt is five drops. On the high reading it is about three teaspoons.
Neither number makes an individual prompt a meaningful personal footprint. What matters at scale is the second chart, not the first: the question is not what one prompt costs but what a sector growing 15% a year does to a grid that is not growing at 15% a year.
One caveat on all of these. Google's figure is a median for short text prompts in one product in one month. Image and video generation, long documents and reasoning models are heavier, and none of the published per-prompt figures cover them.
Sources and method
- Measuring the environmental impact of delivering AI at Google Scale, Google, submitted 21 August 2025. The 0.24 Wh, 0.26 mL, 0.03 gCO₂e figures and the 58% accelerator share are from this paper, for the median Gemini Apps text prompt.
- The Gentle Singularity, Sam Altman, June 2025. Gives 0.34 watt-hours and 0.000085 gallons, which is about 0.32 mL.
- Making AI Less "Thirsty": Uncovering and Addressing the Secret Water Footprint of AI Models, Li, Yang, Islam and Ren, version 5 dated 26 March 2025. Source of the 700,000 litre training figure and the 2027 withdrawal projection. The 5 mL and 15 mL per-prompt figures are the author’s revised estimates as reported in the press following the Google paper.
- Energy and AI, International Energy Agency, 2025. Source of the 415 TWh, 945 TWh and 15% a year figures.
- Every source here was opened and checked on 25 September 2026. Where two credible sources disagree, both are shown with their scope, rather than averaged into one number.