Water per prompt is energy times a cooling factor
A prompt draws a small amount of electricity; each kilowatt-hour of data-center energy evaporates about 1.9 litres of water on the global average (its Water Usage Effectiveness, or WUE). Multiply the two and you get the water per prompt. Because the energy per prompt is tiny, so is the water — grams of water, not litres.
The famous "500 mL" number is for a conversation, not a prompt
The gap between headlines and reality is the real story here. The 2023 UC Riverside study reported about 519 mL, but that was for a full conversation on an older model, counting power-plant water. Newer figures from the model makers are far smaller.
Same activity, wildly different figures — the bars are that far apart because the studies measure different things and different model generations.
Common per-prompt estimates
| Source | Water | What it counts |
|---|---|---|
| OpenAI (Sam Altman, 2025) | 0.3 mL | Per prompt |
| Google (Gemini, 2025) | 0.26 mL | Per prompt |
| Independent estimate, 2024 | ~5 mL | Per prompt |
| Literature average | 10–25 mL | Per prompt, older models |
| UC Riverside, 2023 | 519 mL | Per conversation |
For scale: a 0.5 L water bottle is about 500 mL, so at 0.3 mL a prompt you could run more than 1,600 prompts on one bottle. At the old 25 mL figure, about 20 prompts.
What moves the number
- Cooling method. Water cooling evaporates 2–5 L/kWh; air cooling uses 0.3–0.5 L/kWh on-site but burns more power, shifting the water to the power plant.
- Location. A data center in Iceland or the Pacific Northwest (hydro, cool air) uses far less water per kWh than one in Arizona.
- Season. Hot summer days need more evaporative cooling, so the same prompt can cost two to three times more water in July than in January.
- Training vs use. Training a large model is the big cost — GPT-3\'s training was estimated near 700,000 L — but that is a one-time expense spread across billions of later prompts.
Common questions
How much water does one ChatGPT prompt use?
Estimates range from about 0.3 mL to 25 mL for a single prompt, depending on the study. The widely shared 500 mL figure was for a whole 5-to-50 exchange conversation in a 2023 model, not one prompt. OpenAI and Google now put a typical prompt near 0.3 mL.
Why do the numbers vary so much?
They measure different things. Some count only water evaporated in the data center; others add the water used at the power plant that makes the electricity. The location, the cooling system, and the season all move the figure, so a 100x spread between estimates is normal.
Where does the water actually go?
Most of it evaporates. Data centers use water to carry heat away from the servers, and the power stations feeding them evaporate more water for their own cooling. Roughly 80% of an AI query's water footprint is off-site, at the power plant, not in the data center itself.
