OpenAI CEO Sam Altman has once again pushed back hard against claims that ChatGPT is draining the world's water supply, calling the idea "completely untrue, totally insane" and saying it has "no connection to reality." The comments, made at an AI summit in India and repeated days later on the Sources podcast with Alex Heath, have reignited a long-running debate over the real environmental cost of running large language models at scale.
Here's what Altman actually said, what independent researchers and fact-checkers have found, and how much water and energy a single ChatGPT query really uses.
What Sam Altman Actually Said
Speaking at an event hosted by The Indian Express, Altman dismissed viral claims that each ChatGPT query consumes gallons of water. According to TechCrunch's report on the comments, Altman said: "Now that we don't do that [evaporative cooling], you see these things on the internet where, 'Don't use ChatGPT, it's 17 gallons of water for each query' or whatever. This is completely untrue, totally insane, no connection to reality."
He argued that modern OpenAI data centers have moved away from evaporative cooling, the water-intensive method that fueled many of the original viral water-usage statistics, and now use closed-loop or air-based cooling systems that consume water at levels closer to an office building than a factory.
The Almond Comparison That Went Viral
Altman went further in a separate interview, offering a specific comparison: as reported by Yahoo Finance, he claimed that "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." The comparison was meant to make AI's water footprint sound trivial next to everyday agriculture, but it quickly drew scrutiny from fact-checkers and researchers.
Fact-Check: Is the Almond Comparison Accurate?
PolitiFact examined the claim directly and rated it "Mostly False." Their reporting found that Altman's comparison relies on a relatively low estimate for ChatGPT's water use paired with a relatively high estimate for how much water a single almond requires to grow. Independent estimates cited by PolitiFact suggest it would actually take somewhere between 1,000 and 10,000 conversational AI prompts, not 38,000, to match the roughly 6 liters of fresh water used to grow one California almond.
Crucially, PolitiFact also notes that OpenAI has not published enough detailed operational data for outside researchers to independently verify its own water-use estimates, which makes any specific number, including Altman's, difficult to confirm with certainty.
How Much Water Does One ChatGPT Query Actually Use?
In a June 2025 blog post cited by TechRadar, Altman put a specific figure on it: the average ChatGPT query uses about 0.34 watt-hours of electricity, roughly what an oven uses in a little over one second, and about 0.000085 gallons of water, or roughly one-fifteenth of a teaspoon. Individually, that is a tiny amount. But ChatGPT has more than 400 million weekly active users sending multiple prompts a day, which is where the debate over cumulative, global-scale impact comes in.
Why the Debate Isn't Going Away
Part of the disagreement comes down to what's being measured. Altman's per-query figures reflect inference (the cost of answering a single prompt), not the far more resource-intensive process of training a model in the first place. He has separately argued that comparing AI training costs to human costs is unfair, telling the Indian Express crowd that "it also takes a lot of energy to train a human," pointing to roughly 20 years of food, care, and education behind every adult capable of independent thought.
Meanwhile, the debate is playing out against real regulatory pressure: California Governor Gavin Newsom is currently considering bills that would force data center operators to disclose far more detailed information about their water consumption, which could make it possible to verify claims like Altman's for the first time.
The Bottom Line
Sam Altman is right that the viral "17 gallons per query" figures circulating online are almost certainly exaggerated, and that modern data centers have largely moved away from the most water-intensive cooling methods. But his own almond comparison overstates just how favorably AI compares to everyday agriculture, according to independent fact-checkers. The honest answer, for now, is that neither side has enough transparent data to settle the debate definitively, which is exactly why pending disclosure laws in California matter.
For more on how AI companies are scaling their infrastructure and what it's costing consumers and the planet, see our coverage of why RAM prices are skyrocketing in 2026 and how AI platforms like Povotra are being built for real-world use cases.
Frequently Asked Questions
1. Did Sam Altman really say ChatGPT's water usage is "fake"?
Yes. Speaking at an event in India, Altman called claims that ChatGPT consumes large amounts of water per query "completely untrue, totally insane, no connection to reality," specifically pushing back on viral figures like "17 gallons per query."
2. Is the almond comparison accurate?
Not entirely. PolitiFact rated Altman's claim that 38,000 ChatGPT queries equal the water used to grow one California almond as "Mostly False," noting it combines a low estimate of AI water use with a high estimate of almond water use. More independent estimates suggest the real ratio is closer to 1,000–10,000 queries per almond.
3. How much water does a single ChatGPT query use?
According to Altman's own June 2025 blog post, roughly 0.000085 gallons, about one-fifteenth of a teaspoon, per average query.
4. Why can't this be independently verified?
OpenAI has not published detailed enough operational data on its data centers for outside researchers to fully verify its water and energy claims, according to PolitiFact's reporting.
5. Could this become regulated?
Possibly. California is considering legislation that would require data center operators to disclose far more information about their water consumption, which could make these claims independently testable in the near future.