“38,000 ChatGPT queries use the same water as growing one California almond.” The instinctive reaction is relief. Almonds are small. Then the math starts, and things get slippery fast.
Sam Altman made the claim on the Sources podcast with journalist Alex Heath in early September 2026, pushing back against viral social media posts claiming a single ChatGPT query consumes as much water as a six-hour shower. Those posts were absurd — off by orders of magnitude. His replacement statistic, however, wasn’t exactly bulletproof. Altman himself noted he was recalling the figure from memory. That caveat matters more than it might seem.
Where Altman’s Number Actually Comes From
The arithmetic holds together, but only with carefully chosen ingredients.
His team later confirmed the inputs: 0.000085 gallons (roughly 0.32 ml) per query, multiplied against a 3.2-gallon almond water footprint. Run the division and you get 37,647 — close enough to 38,000. Neat. Except the Almond Board of California and US Geological Survey-linked research both cite roughly one gallon per almond as the representative figure, not 3.2.
Using one gallon with Altman’s own per-query estimate yields closer to 11,000–12,700 queries per almond. The 3.2-gallon figure exists within some methodology ranges but sits well above mainstream estimates. There’s a second problem: Altman’s 0.32 ml covers direct, on-site cooling water only. It excludes water consumed by power plants to generate the electricity running the servers in the first place.
As Axis Intelligence noted, “Sam Altman’s 0.3 ml per query and the UC Riverside 10–25 ml per query are both technically correct… They are not measuring the same thing.”
The Water Nobody’s Counting
Scope changes everything — and AI companies get to choose their scope.
Independent analyses using “scope-2” accounting — water consumed during electricity generation, not just on-site cooling — put typical ChatGPT queries at 3–10 ml, with earlier studies reaching 10–25 ml. At 3 ml per query and a one-gallon almond, you get roughly 1,260 queries per almond. Same almond. Very different story. The methodology isn’t a footnote; it’s the entire answer.
Altman also argues modern AI data centers use water “like an office building” — bathrooms and sinks, not cooling towers. That reflects a real shift in some newer facilities. The broader picture is more complicated. A Fayetteville, Georgia data center reportedly consumed 30 million gallons during a drought, contributing to residential water pressure drops. Per-query averages and local infrastructure stress are not the same metric, and treating them as interchangeable is how a podcast talking point glosses over lived reality.
PolitiFact rated Altman’s statement “Mostly False,” arguing his figure “overstates the water consumed by almond agriculture and uses a relatively low ChatGPT water use estimate, which is unconfirmed and disputed.”
The Scale Problem Worth Paying Attention To
Small numbers multiplied by billions stop being small.
OpenAI has said ChatGPT handles roughly 2.5 billion user prompts per day. Even at Altman’s own 0.32 ml figure, that works out to approximately 800,000 liters of direct cooling water daily — before accounting for electricity. New AI agents that execute complex, multi-step tasks require substantially more compute than a simple chat prompt, making the per-query baseline an increasingly conservative anchor as workloads grow more sophisticated.
Neither Altman’s almond nor the six-hour shower is a reliable number. No standardized, publicly verified methodology exists for measuring AI’s water footprint. Until one does, every vivid comparison — however memorable — is really just a choice about which math to show you.
