Enter how you actually use AI. This shows the modeled energy and water behind it, using published ranges rather than one viral number. Nearly every figure here is disputed, and that gap is itself part of the story.
Why ranges, not one number. Only two companies, OpenAI and Google, have published any per-query figures, and both cover on-site cooling only. Everything else on this page is a third-party estimate built from hardware assumptions or a single 2023 university study. Treat every number here as an order of magnitude, not a bill.
| Activity | Energy (typical) | Water (typical) | Basis |
|---|---|---|---|
| Quick chat | 0.3 Wh | 10 mL | Epoch AI 2025 / Altman 2025 / Google 2025 disclosure; water blends on-site + grid, per UC Riverside "Making AI Less Thirsty" |
| Complex / reasoning | 18 Wh | 150 mL | University of Rhode Island AI Lab, 2025, GPT-5 reasoning mode; water scaled to the same ratio as the chat estimate |
| Image generation | 8 Wh | 30 mL | Hugging Face / Carnegie Mellon measured energy; water is not independently studied, so it's derived from energy using a standard grid water-intensity factor |
| Agent / coding session | 300 Wh | 1,500 mL | Most speculative row. Built from Epoch/Climate Brink findings that agent tasks run roughly 1,000x the tokens of a single chat reply, scaled to a 30-minute session |
Sources: Epoch AI, "How much energy does ChatGPT use?" (2025). Sam Altman, "The Gentle Singularity" (June 2025). Google, Gemini per-prompt methodology (Aug 2025). Li, Yang, Islam, Ren, "Making AI Less Thirsty" (UC Riverside / UT Arlington, 2023, expanded 2025). University of Rhode Island AI Lab, GPT-5 energy analysis (2025). Zeke Hausfather, "The real energy use of agentic AI," The Climate Brink (2026).