OpenAI CEO Sam Altman says the water footprint of artificial intelligence has been overstated as opposition to new data centres grows.
Speaking on the Sources Podcast with Alex Heath, Altman rejected claims that one ChatGPT query consumes water equal to a six-hour shower. He cautioned that he was recalling his calculation from memory and might be wrong.
“For every 38,000 ChatGPT queries, that is the same amount of water that is used in the production of a single almond in California,” Altman said.
A widely cited estimate places the water requirement for one almond at about 1.1 gallons. However, experts say limited industry data makes Altman’s comparison difficult to verify.
Altman points to modern cooling systems
Altman acknowledged that older data centres relied heavily on evaporative cooling. He argued that operators have since adopted technologies that sharply reduce direct water consumption.
“Like if you look at a modern very large data center, it uses the equivalent amount of water as an office building in terms of, you know, people like running the sinks and the toilets and whatever,” he said.
“So, that has been a robust meme and difficult to disprove, but I don’t think it holds up to any scrutiny.”
A Virginia legislative review partly supports that argument. It found that most data centres used as much water as, or less than, an average large office building. However, a small number of facilities consumed substantially more.
Microsoft introduced a closed-loop design in 2024 that uses no water for cooling during operation. The company says each centre could avoid more than 125 million litres of annual water use. Its Wisconsin AI site expects yearly water consumption similar to a typical restaurant.
Nvidia has promoted liquid-cooling technology that can cut cooling-water use by up to 100 percent. Amazon says its US data centres consume 0.075 percent of the water Americans use on lawns and gardens.
Research produces sharply different estimates
A 2023 academic study estimated that training GPT-3 consumed about 700,000 litres, or 185,000 gallons, directly at Microsoft’s US facilities. Its broader estimate reached 5.4 million litres after including water linked to electricity generation.
The researchers also estimated 500 millilitres of water for every 10 to 50 medium-length responses. Actual consumption changes with location, weather, cooling technology and workload.
Altman offered another estimate in 2025. He said an average ChatGPT query used 0.000085 gallons, or roughly 0.32 millilitres. That rate would equal about 13,000 queries per 1.1-gallon almond, rather than 38,000.
Public concern has not disappeared. One Georgia data centre reportedly consumed 29 million gallons over 15 months. Residents near another project alleged that construction turned their well water muddy.
Experts told CalMatters that companies disclose too little operational data to confirm universal per-query estimates. AI agents may further increase demand because they process far more tokens than ordinary chatbot exchanges.
