Sam Altman says 38,000 ChatGPT prompts equal one almond. The math only works on his terms

The OpenAI CEO used a snack to kill a six-hour-shower meme. U.S. search spiked past 100,000. The ratio is real only if you pick the thirstiest almond study and the smallest query he has ever published.

Sam Altman went on Alex Heath’s Sources podcast this week to swat a meme: that a single ChatGPT prompt uses as much water as a six-hour shower that never comes back.

His replacement image was an almond.

“I don’t have the exact calculation in front of me, but I think the real number is something like — doing this from memory, it might be wrong, but it’s close — 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.”

He added that this was “full-on, total true water accounting, not just what’s running in one data center,” and that people “scarfing down 12 almonds at a time don’t feel like they’re doing something horrible from a water perspective, for the most part.”

By Friday night that line was the #4 search in the United States — sam altman chatgpt water usage, 100,000+ queries, related search: sam altman almonds. GPT-6 Astra launched in the same window. The snack outranked most of college football.

The six-hour shower is garbage. The almond is a lawyered comparison. Both can be true.

Where 38,000 comes from

Altman did not invent a new measurement on the podcast. He reached for a number he published himself.

In a June 2025 essay, The Gentle Singularity, he wrote that the average ChatGPT query uses 0.000085 gallons of water — about 0.32 milliliters, a fifteenth of a teaspoon. OpenAI does not publish a current, audited figure. That blog post is still the company’s only public per-query number. Google has put a Gemini text prompt at 0.26 ml. Same order of magnitude.

Almonds are the other half of the fraction. A 2019 paper affiliated with the U.S. Geological Survey is the one that gets you close to Altman’s headline:

• ~12 liters (3.2 gallons) per kernel if you count irrigation, rainfall, and water-quality impact
• 12,000 ml ÷ 0.32 ml ≈ 37,500 queries per almond

That is 38,000. It is also the top of the range.

Use the irrigation-only numbers everyone else cites and the boast collapses:

PolitiFact rated the 38,000 claim Mostly False. Not because ChatGPT drinks like a shower. Because Altman picked the thirstiest almond and the smallest query, then presented the ratio as a fact he happened to remember.

If his 0.32 ml is stale — evaporative cooling, older models, short text prompts — the gap is worse. Independent estimates for a ChatGPT reply have run from 1 to 50 ml when you include the electricity that made the power. At 1 ml you are in the low thousands of prompts per almond. At 50 ml you are in the dozens. GPT-6 Astra, Codex, and “computer use” agents run longer than a 2025 average query. The blog number is a snapshot of a product that no longer exists.

The office-building line

Altman’s second claim is the one that actually matters for zoning fights.

“If you look at a modern, very large data center, it uses the equivalent amount of water as an office building in terms of people running the sinks and the toilets.” He called the evaporative-cooling panic “a robust meme and difficult to disprove,” and said those systems have not been used “in a long time.”

Partly right, partly sales.

A 2024 Virginia state commission — Virginia is the densest data-center corridor in the country — found most facilities used as much water as a large office, or less. Some used far more. Geography and cooling design decide, not the word “AI.” Closed-loop and air-cooled halls sip. Evaporative towers in hot, dry counties still dump water to the air. New hyperscale campuses are not uniformly “modern” on day one, and they are not uniformly in places with spare aquifer.

The meme Altman is killing is per-prompt guilt. The fight cities are in is cumulative load: power plants, cooling towers, and a buildout that is still accelerating because agents don’t send one prompt.

California almond orchards still dwarf the sector. Reason has cited on the order of 4.2 billion gallons a day for those farms versus tens of millions for data centers nationally. That is a real comparison. It does not tell a town in northern Virginia whether the next hall should get a water permit.

What the search is actually asking

Related queries were not “is AI evil.” They were the almond. People wanted to know if the CEO was lying.

Short answers:

• One ChatGPT prompt is not a six-hour shower. That viral math is wrong by orders of magnitude.
• 38,000 prompts = one almond only if you use Altman’s 2025 per-query figure and the 12-liter almond. A more typical almond (~1 gallon) lands nearer 11,000–12,000.
• OpenAI has not published a 2026 water number for GPT-6 Astra, computer-use sessions, or image jobs.
• A modern hall can look like an office. An evaporative campus in a drought county does not.
• Scale beats the snack. Billions of prompts, longer agent runs, and new sites are the water story. The almond is a talking point.

Altman is correct that the shower meme made people feel filthy for using a chatbot. He is also the vendor. Until OpenAI puts a current, third-party water figure on Astra the way it puts FrontierMath scores on Astra, the 38,000 number is a memory of a blog post, not a measurement of the model that just launched.

Eat the almond. Ask for the spreadsheet.