Sam Altman Downplays ChatGPT's Water Thirst With an Almond Claim That Overstates the Real Gap Threefold

Sam Altman wants you to stop feeling guilty about ChatGPT. The OpenAI chief told a new podcast this week that 38,000 queries use as much water as producing one California almond. His own past figure says otherwise.

Speaking on the launch episode of the Sources podcast with journalist Alex Heath, Altman pushed back on what he called a meme about artificial intelligence (AI) draining local water supplies.

A modern data centre, he said, uses 'the equivalent amount of water as an office building' with its sinks and toilets running. He argues newer sites rely on closed-loop cooling that recirculates water rather than evaporating it. He also conceded he was quoting from memory and did not have the exact figure to hand.

Where the Almond Math Breaks Down

Altman's ratio doesn't hold up against public data. Growing a single California almond takes about 3.56 litres (3.13 imp. quarts) of water, according to research affiliated with the US Geological Survey.

Altman himself has estimated a typical ChatGPT query at roughly 0.32 millilitres. Divide one by the other, and you get close to 11,000 queries per almond, not 38,000. The 38,000 figure only works if you assume around 12 litres (10.55 imp. quarts) per almond, the top of the range cited by water researchers. Altman's ratio sits well above every figure he has publicly cited.

His Own Blog Post Undercuts the Boast

The 0.32 millilitre estimate is not an outside number. Altman published it in June 2025 in a blog post titled 'The Gentle Singularity', and it has been repeated widely since.

He maintained the almond figure was still in the right ballpark. Two independent fact-checks released this week by explainx.ai and frontiernews.ai ran the same calculation and reached the same conclusion: the comparison oversells the gap between AI and agriculture by roughly a factor of three.

What the Snappy Soundbite Skips

A per-query figure also hides the scale. OpenAI now serves about 900 million weekly users sending billions of prompts, so tiny fractions of a millilitre add up fast, and that is before the water used to train models.

Much of the original worry traces to a 2023 study, 'Making AI Less Thirsty', from the University of California, Riverside, which estimated a short GPT-3 chat could drink roughly a 500-millilitre (0.44-imp. quart) bottle.

Altman's number likely counts only on-site cooling, not the water burned at power plants to make the electricity behind each answer. Researchers estimate that power layer can reach up to 75% of a query's true water footprint.

Real Data Centres Tell a Messier Story

On the ground, the picture is less tidy than an office building. A data centre campus in Fayette County, Georgia, drew about 29 million gallons (131.8 million litres) over 15 months through connections the county did not know existed, prompting officials to ask residents to stop watering lawns.

A separate Georgia project drew complaints that neighbours' tap water had turned muddy. Shaolei Ren, an engineering professor at the University of California, Riverside, says the industry is too secretive to trust any single tidy statistic.

His team found data centre cooling could demand up to 1.45 billion extra gallons (5.49 billion litres) of peak capacity a day nationwide. Altman may be right that one prompt is cheap. The reassurance rests on a number he can't quite stand behind.