How Much Water Does ChatGPT Use? The Real Numbers
The viral claims range from "a bottle per email" to "nothing at all." Here are the actual published figures — from OpenAI itself and from the academic study that started the conversation — with sources you can check.
Per query: a fifteenth of a teaspoon
Sam Altman's June 2025 figure: the average ChatGPT query consumes about 0.000085 gallons of water — roughly 0.32 millilitres, or one-fifteenth of a teaspoon. That covers the evaporative cooling attributable to your query at the data center.
Training: where the big water goes
The widely-cited academic work — "Making AI Less Thirsty" by researchers Pengfei Li, Jianyi Yang, Mohammad A. Islam and Shaolei Ren — estimated that training GPT-3 in Microsoft's US data centers directly evaporated about 700,000 litres of clean freshwater. The same paper projects global AI demand could account for 4.2–6.6 billion cubic metres of water withdrawal in 2027 — more than the annual withdrawal of some countries.
Why data centers use water at all
It's the cooling. Evaporative systems spray water over heat exchangers; the evaporation carries heat away far more cheaply than mechanical chilling — trading water for electricity. Facilities in dry regions increasingly use closed-loop or dry cooling instead, which uses little water but more power. The trade-off is real and local: what matters is which watershed a campus draws from. See how data center cooling works.
How to read the viral numbers
- "A bottle of water per conversation" — traces back to early per-query estimates from the UCR paper's era, before operators published their own figures. OpenAI's current per-query number is ~1,500× smaller than a 500 mL bottle.
- "AI uses no water" — false at infrastructure scale: training runs and evaporative cooling consume real water, concentrated in specific regions.
- The honest frame — per-query use is trivial; aggregate, location-specific use is a legitimate planning issue, which is why several states now weigh water in data center permitting.
The energy numbers →← Cooling explained
Sources (verified August 1, 2026):
• Sam Altman, "The Gentle Singularity" (OpenAI, June 2025) — average ChatGPT query ≈ 0.34 Wh and 0.000085 gal of water.
• Li, Yang, Islam & Ren, "Making AI Less Thirsty" (arXiv:2304.03271) — GPT-3 training ≈ 700,000 L freshwater evaporated; global AI projected at 4.2–6.6 billion m³ water withdrawal in 2027.
Frequently asked questions
How much water does one ChatGPT query use?
About 0.000085 gallons — roughly 0.32 mL, one-fifteenth of a teaspoon — according to OpenAI's Sam Altman (June 2025).
Did training ChatGPT really use 700,000 litres of water?
The peer-reviewed estimate for training GPT-3 — by researchers Li, Yang, Islam and Ren ('Making AI Less Thirsty') — is about 700,000 litres of freshwater directly evaporated in Microsoft's US data centers.
Will AI water use keep growing?
The same study projects global AI demand could reach 4.2–6.6 billion cubic metres of water withdrawal in 2027. Growth is real — but concentrated where evaporatively-cooled facilities sit, which is why siting and cooling choices matter more than individual usage.
Why do data centers evaporate water?
Evaporative cooling removes heat far more cheaply than mechanical chilling — trading water for electricity. Dry and closed-loop systems reverse the trade: more power, almost no water.