How OpenAI Is Trying to Rewrite the Water Story

The Almond Defense: How OpenAI Is Trying to Rewrite the Water Story

KENNETH MUNOZ · SEPTEMBER 10, 2026

What Happened

On the launch episode of the Sources podcast with Alex Heath, Sam Altman was asked about the growing backlash over how much water AI data centers use. His answer, delivered from memory and without a citation in hand, has since become the industry's go to talking point: 38,000 ChatGPT queries, he said, use about as much water as growing a single California almond. He went further, claiming that a modern large data center uses no more water than an office building, just people running sinks and flushing toilets, and dismissed the entire water concern as a "robust meme" that doesn't "hold up to any scrutiny."

The clip spread fast. Screenshots of the quote racked up hundreds of thousands of views on Reddit and X within days, and the almond line in particular gave critics of AI's water footprint a tidy image to push back against, along with a tidy image for OpenAI's defenders to rally behind.

The Math, Checked

The almond comparison isn't invented from nothing. A 2019 study affiliated with the U.S. Geological Survey put the water footprint of a single California almond at around 3.56 liters. Altman has previously estimated a ChatGPT query at roughly 0.32 milliliters of water. Divide one by the other and you land close to 11,000 queries per almond, not 38,000. That's a threefold overstatement in OpenAI's favor, according to an analysis by IBTimes UK, and PolitiFact rated the claim "Mostly False" for a related reason: it also leans on a relatively low, unconfirmed water estimate for ChatGPT itself, since OpenAI hasn't published the methodology behind that 0.32 milliliter figure.

There's a bigger gap in the number, too. Altman's estimate appears to count only the water used on-site for cooling, not the water consumed upstream at power plants to generate the electricity a query actually runs on. Shaolei Ren, an engineering professor at UC Riverside who studies AI's resource footprint, has found that this power-generation layer can account for up to 75 percent of a query's true water cost. Leave that out, and any comparison to an almond, or to anything else, is measuring a fraction of the real number.

The Office Building Claim

Altman's second claim, that a modern large data center uses roughly what an office building uses, in sinks and toilets, is harder to check because there's so little public data to check it against. It's true that some newer facilities have moved from evaporative cooling toward closed-loop systems that recirculate water rather than losing it to the air. But "some newer facilities" isn't the same as "the industry," and independent reporting has found messier realities on the ground. A data center campus in Fayette County, Georgia drew about 29 million gallons of water over 15 months through connections local officials didn't know existed, prompting the county to ask residents to cut back on watering their lawns. A separate Georgia project drew complaints from neighbors that their tap water had turned muddy after a nearby facility began operating.

Nationally, direct water use by data centers rose from about 5.6 billion gallons in 2014 to roughly 17 billion gallons in 2023, according to the Congressional Research Service. That's the direct figure only, and it's still climbing as AI-driven construction accelerates.

Why the Transparency Fight Matters

The reason so much of this debate runs on estimates and modeling instead of hard numbers is that data centers generally don't have to disclose their water use, and the industry has spent real political capital keeping it that way. California Assemblymember Diane Papan has tried twice now to pass a law requiring data center operators to report their water sources and consumption. Governor Gavin Newsom vetoed her first attempt last year, citing a reluctance to impose "rigid reporting requirements" before understanding the impact on business. Business groups, including the Data Center Coalition, opposed that bill and are opposing this year's version too. Two similar measures, AB 2469 and AB 2619, are now sitting on Newsom's desk with an October 12 deadline for him to sign or veto them again.

Until legislation like that passes somewhere and actually takes effect, comparisons like Altman's almond math are close to unfalsifiable. Nobody outside these companies has the raw numbers to check the estimates against, which is precisely why a former OpenAI employee turned water-policy expert, Alex Biering of the California Farm Bureau Federation, has suggested using the almond figure as a starting point rather than an endpoint: if regulators can pin down what an almond actually costs in water, maybe that gives them a reference point for pinning down what a data center costs too.

The Bottom Line

The almond line is a good soundbite because it's specific, mildly funny, and makes AI sound trivial next to agriculture, which genuinely does use far more water overall than data centers do. But specific doesn't mean accurate. The best available math suggests Altman understated ChatGPT's water use by roughly three times even on his own terms, before accounting for the electricity side of the ledger at all. And the office building comparison, offered with total confidence, can't currently be verified by anyone outside the companies making it, because those companies are also the ones fighting to keep the underlying numbers private.