AI water footprint calculator
Data centers use water to cool their servers, and power plants use water to make the electricity. The published figures per prompt range from a few drops to several spoonfuls.
The gap comes from the scope. Google and OpenAI count the running of the data center. Mistral counts the whole life of the model, training included.
Result
- Water per day
- 6.4 mL
- Glasses of water per year (250 mL)
- 9.34
These are estimates. Real figures depend on the model, the data center and the length of the answer.
How this is calculated
- Google (2025). A median Gemini text prompt uses 0.24 Wh, 0.03 g CO2e and 0.26 mL of water. arxiv.org
- OpenAI (2025). Sam Altman states an average ChatGPT query uses about 0.34 Wh and 0.32 mL of water. blog.samaltman.com
- Mistral AI with Carbone 4 and ADEME (2025). Life cycle analysis of Mistral Large 2. A 400 token answer accounts for 1.14 g CO2e and 45 mL of water, training included. mistral.ai
Questions
Does ChatGPT really use a bottle of water per conversation?
Not with the figures the companies publish today. OpenAI gives 0.32 mL per query, so a 500 mL bottle would cover about 1,500 queries. With the Mistral life cycle figure of 45 mL, the same bottle covers about 11 answers.
Which figure should I trust?
They answer different questions. The running figures tell you what one more prompt costs today. The life cycle figure spreads the cost of training and hardware over every answer.