ChatGPT Water Usage: Altman's Almond Comparison Explained

Sam Altman put a number on ChatGPT water usage: the water behind 38,000 prompts is about the same as the water it takes to produce a single almond. It is a direct answer to the “AI is drinking the water supply” criticism, delivered in a unit people can actually picture. Below are the four points he made, what can and cannot be verified in them, and how I would use the figure when someone inside a company asks about AI’s environmental footprint.

What Altman said about ChatGPT water usage

The claim has four parts:

• 38,000 prompts use about the same water as producing one almond
• the figure is described as covering the whole picture, not one data center carved out on its own
• evaporative cooling has not been in use for a long time
• the newest large data centers use water on the order of a single office building

What I find interesting is not the size of the number but the choice of comparison. Water debates are usually conducted in liters or cubic meters, and those units leave no picture in the listener’s head. An almond sits in your palm. It hands the judgment of “a lot” or “a little” back to the person you are talking to, instead of asking them to trust a scale they have never handled.

From 0.000085 gallons to one almond

In a blog post in June 2025, Altman wrote that a single ChatGPT prompt used roughly 0.000085 gallons of water, about one fifteenth of a teaspoon. The almond line is a continuation of that. And it works better, for a simple reason: one fifteenth of a teaspoon is unjudgeable. Told that number, nobody knows whether to be alarmed or relieved. Stack the per-prompt figure up to 38,000 and you finally have something on the other side of the scale.

In other words, the same claim with the unit swapped out. That is worth stealing as a communication technique. When you present numbers internally, a figure whose unit is too small to judge will not move anyone. Convert it into something your audience measures in daily life, and the same data suddenly lands.

The missing piece: how the almond math was done

There is a caveat, and it matters. How the 38,000 prompts were matched against one almond — the calculation method — is not shown in the video. There is a statement that the figure includes the full footprint rather than a single facility, but not what “full” includes. Without the assumptions, nobody outside can walk the calculation back and check it.

Altman’s own position is that data center water has become a sticky meme, hard to argue against, and that it does not hold up when you actually check it. Whether you accept that is a separate judgment. But the comparison he offered has not itself been published in a checkable form, and that is the accurate way to hold it. The easier a comparison is to repeat, the further it travels without its sources.

Where I would draw the line

• fine as internal grounding, or as the starting point of a discussion
• not fine as a figure in your own deck or press release presented as your company’s claim — for that, use numbers whose method is public
• if you cite it, cite it with the caveat: this is the provider’s own statement and the method is not published

Evaporative cooling and what “using water” means

The line about evaporative cooling goes after the origin of the mental image. Cooling where water evaporates and leaves the system is a completely different balance sheet from cooling that circulates water in a closed loop, even though both are described as “using water.” The picture of a data center draining a region’s supply is drawn on the first assumption.

The claim that the newest large data centers use water comparable to one office building points the same direction. But which facility, and which year, is not in the statement. Treat it as a claim about the direction the technology has moved, and resist extending it to any specific site near you.

How this lands in Japan

If you are pushing AI adoption inside a Japanese company, the environmental question always arrives. The sources are predictable: the internal approval round, a check from the sustainability team, a questionnaire from a client’s procurement side, and a pointed comment from the colleague who was never comfortable with AI in the first place.

The trouble is that neither side of that conversation has a feel for the quantities. Someone says “it uses enormous amounts of water,” you answer “not really that much,” and the two statements never meet. The value of this comparison is that it moves the discussion from feeling back to quantity. At minimum it becomes an argument about what is being compared to what, which is an argument you can actually make progress in.

There is also a lesson here for overseas AI companies communicating in Japan. Making the unit familiar works in Japan too — but Japanese audiences are unusually likely to ask where the number came from. Ship a vivid comparison without the method behind it, and the vividness converts straight into suspicion. Press and enterprise buyers alike will treat an unsourced figure as a marketing line, and the clarity you worked for becomes a liability. Prepare the comparison and its basis as one package, not two.

A script for the water question at work

My sequence:

1. Decide what you are comparing. One prompt? One employee for a year? The whole company’s annual usage? These are wildly different conversations and people conflate them.
2. Convert it into a unit your audience measures in daily life.
3. Give the source and the assumptions alongside the number. If you cannot, say so plainly rather than smoothing over it.
4. Do not mix water and electricity. They have different cost structures, different local sensitivities, and different mitigations. Mixed together, the discussion blurs and no decision gets made.

The test is simple: are you using this number to make a decision, or to align perceptions? A decision needs figures you can verify. Alignment can run on something like the almond comparison, and that is often enough to unblock the room.

Taking this into your own company

My suggestion is to run your own usage through the math once. Your monthly request count is in your admin console. Multiply it by a published per-prompt estimate and you have a sense of your own order of magnitude. Write the assumptions next to it, park it somewhere internal, and the next time the environmental question comes up you are not starting from zero.

It also helps to choose your comparison unit in advance. An almond, a plastic bottle, the water your own factory or store goes through in a day — any of them will do. Having one unit your audience can actually measure is what keeps your answer consistent from one meeting to the next.