Claim report card
How much water does an AI prompt use?
“Every AI prompt uses a bottle of water.”
It depends on the model, the task, the date and what is counted. Google reports 0.26 mL for a median Gemini text prompt (2025); OpenAI reports about 0.32 mL per ChatGPT query (2025). The bottle came from one modeled GPT-3 scenario: about 500 mL for 10 to 50 responses (2023), not today’s models. A long reasoning or agentic task can use far more than a short answer. Ruling: false.
| Value | Measures | Source | Grade |
|---|---|---|---|
| 0.26 mL | water for a median Gemini text prompt, comprehensive method | Google Cloud (2025) | Background |
| 0.32 mL | water for an average ChatGPT query, per OpenAI | Sam Altman (2025) | Background |
| ~500 mL / 10 to 50 responses | the modeled GPT-3 scenario that became the “bottle” claim | Li, Yang, Islam & Ren (arXiv) (2023) | Background |
| >10,000× | variation in measured workload-level data-center water use, per LBNL | Lawrence Berkeley National Laboratory (2025) · U.S. Government Accountability Office (2025) | Background |
Why do water-per-prompt estimates differ by 1,000 times or more?
Estimates count different things. On-site cooling water (Google, OpenAI) is a fraction of a millilitre. Adding the water used to generate the electricity, plus a pessimistic climate and cooling design, pushed the original 2023 estimate to roughly 500 mL per 10 to 50 responses. Measured workloads vary more than 10,000-fold (LBNL, 2025), so no single figure applies everywhere. Annual totals such as de Vries-Gao’s 312.5 to 764.6 billion litres for 2025 are system-level estimates, not an average prompt.
Why this grade
What the sources show
- 1
LBNL reports more than 10,000-fold variation in workload-level water use, so one universal estimate does not fit every prompt.
- 2
Google’s and OpenAI’s figures are operator-level averages for on-site cooling. Google applies a fleet-wide water-use efficiency; OpenAI does not publish its method.
- 3
A per-prompt estimate can be valid for a stated model, workload, period, infrastructure mix and accounting boundary. The error is presenting any one as representative of all AI use.
- 4
GAO says company reporting does not separate generative AI’s energy and water use from other data-center workloads.
- 5
Water used directly for cooling must be separated from water used to generate electricity.
- 6
Cooling design trades water against power: evaporative cooling saves electricity, while dry or closed-loop systems, including those serving direct liquid-cooled racks, can consume little or no water. Uptime Institute calls general per-query water figures not meaningful.
- 7
“One prompt” is not a fixed unit of work. Efficiency per token has improved fast (Google reports a 33-fold cut in energy per median Gemini prompt over 12 months), but one prompt can return a short answer or trigger extended reasoning, tool use and many model calls.
- 8
Per-prompt estimates expire. Given how fast models and serving systems change, a figure that was valid two years ago is likely outdated today.
Attribution boundary
What would settle it
The facility, model, task, date and local grid conditions for the specific workload in question.
Sources
- The water use of data center workloadsLawrence Berkeley National LaboratoryPublished 2025-06 · checked here 2026-07-17 ↗
- Generative AI’s Environmental and Human EffectsU.S. Government Accountability OfficePublished 2025-04-22 · checked here 2026-07-17 ↗
- Measuring the environmental impact of AI inferenceGoogle CloudPublished 2025-08-21 · checked here 2026-07-17 ↗
- Water is local: generalities do not applyUptime Institute (Jay Dietrich)Published 2025-05-28 · checked here 2026-07-17 ↗