A few years ago, nobody asked how much water an AI chatbot drinks. Now it's one of the most repeated questions about the technology. You've probably seen the claims: a single ChatGPT query "uses a bottle of water," or "runs a shower for hours." Then you've seen the other side, where OpenAI's CEO Sam Altman says the real number is tiny.
Both camps can't be right. So let's look at what has actually been said, why the figures are so far apart, and what you should and shouldn't conclude from them.
What Sam Altman has claimed
In a June 2025 blog post, Altman wrote that an average ChatGPT query uses about 0.34 watt-hours of electricity and about 0.000085 gallons of water, which he described as roughly one fifteenth of a teaspoon. In metric terms, that's around 0.32 milliliters.
He has returned to the subject since then. In public conversations he has pushed back on viral comparisons like the shower claim, calling many of them exaggerated. In one podcast appearance he compared ChatGPT use to growing an almond, saying it takes something like 38,000 queries to match the water used for a single California almond. In early 2026 he went further, rejecting a widely shared claim of 17 gallons per query and saying the idea that water is a big per-query problem is outdated.
Two things are worth noticing here. First, the numbers are small. Second, they came with very little explanation of how they were calculated. Altman didn't publish the underlying method, the hardware, or which data centers and which model were measured. That doesn't make him wrong, but it does mean the figures are a claim, not a dataset.
Why the almond math doesn't quite add up
Here's a small example of how slippery this gets. A commonly cited 2019 estimate says growing one California almond takes around 3.56 liters of water. If a query uses 0.32 milliliters, then dividing gives roughly 11,000 queries per almond, not 38,000.
That's my math, using the figures above, and it shows the problem. Back-of-the-envelope comparisons are easy to make and easy to get wrong, and they depend heavily on which almond study, which query, and which water you count. The almond line makes a good soundbite. It's not a measurement.
Where does the water actually go?
To judge any number, you need to know what it counts. AI water use usually falls into three buckets.
On-site cooling. Servers get hot. Many data centers remove that heat with evaporative cooling, where water absorbs heat and evaporates, a bit like sweating. That water is consumed: it leaves as vapor and doesn't return to the local supply as liquid. Other designs use closed loops, air cooling, or liquid cooling that recirculates water and loses very little. The difference between these designs is huge.
Electricity generation. Power plants use water too, especially thermal plants that boil water for steam and cool it afterward. So the electricity a data center draws carries a hidden water cost that depends on the local energy mix. Wind and solar use very little. Coal, gas and nuclear can use a lot.
Manufacturing and training. Making chips takes enormous amounts of ultra-pure water. Training a large model also uses plenty of energy, and therefore water, before anyone sends a single query. This is often left out of per-query numbers entirely.
When someone says "a query uses X milliliters," the first question is which of these they included. A figure that counts only on-site cooling can be tiny. A figure that counts everything can be far larger.
The metric behind the numbers: WUE
Researchers often use a measure called Water Usage Effectiveness, or WUE. It's simply liters of water used per kilowatt-hour of energy consumed by the data center. A facility with a low WUE sips water; one with a high WUE gulps.
Real-world values vary widely, from close to zero for efficient closed-loop or air-cooled sites to several liters per kilowatt-hour for evaporative ones. A figure of around 1.8 liters per kWh is often quoted as an industry average, though measured values spread all over the place.
Now here's an interesting check. Altman's energy figure is 0.34 Wh per query, which is 0.00034 kWh. At a WUE of 1.8 L/kWh, that works out to about 0.6 milliliters. His stated 0.32 milliliters implies a WUE closer to 0.94 L/kWh. Again, that's my arithmetic, but it suggests his water number assumes fairly efficient cooling, better than that commonly quoted average. That's plausible for modern facilities, and it's also the kind of assumption that deserves to be stated out loud.
Why other estimates are much higher
Independent researchers have come up with bigger numbers, for a few reasons.
Some start from a higher energy estimate. Earlier studies put a ChatGPT-style response at around 2 to 4 watt-hours, with 2.9 Wh often cited. Plug that into the same 1.8 L/kWh figure and you get roughly 5 milliliters per query, more than ten times Altman's number. One academic analysis estimated about 1.2 milliliters for a short prompt under typical conditions, nearly four times the CEO's figure.
Others count indirect water from electricity, which can add substantially, depending on the grid.
And some measure different things. A short question, a long document summary, a reasoning-heavy task, and an image generation request are not the same workload. A "typical query" is an average over wildly different jobs, and the average depends on who picks the mix.
The honest summary is that published per-query water estimates span more than two orders of magnitude. Nobody outside the companies has the complete picture, and there is no accepted standard for how to report it.
Scale changes the story
Even if you accept the smallest number, scale matters. Take a purely hypothetical one billion queries in a day. At 0.32 milliliters each, that's about 320,000 liters. At 5 milliliters each, it's about 5 million liters. Neither figure is the headline-grabbing disaster that some viral posts imply, and neither is nothing.
For comparison, agriculture, power generation and manufacturing use vastly more water than AI does globally. So when people say AI water use is a tiny share of the total, that can be true.
But totals hide a second issue: location. Water isn't a single global pool. A data center drawing from a river in a wet region is a different situation from one pulling from a stressed aquifer in a drought-prone area. Local communities care about the local effect, and a small global share can still be a big local one. This is where much of the real-world debate sits.
What per-query numbers leave out
Even a perfectly measured per-query figure doesn't answer everything.
It says nothing about training, which is front-loaded and very resource-heavy.
It doesn't capture the growth trend. If the number of queries and the size of models both rise quickly, total consumption can climb even while each query gets more efficient.
It doesn't cover the whole stack: chip manufacturing, networking, and backup cooling.
And it usually reflects one company's own infrastructure, which may be better than the industry average. Smaller providers, or the same model hosted elsewhere, could look very different.
How to read claims like these
A few habits will serve you well whenever you see a number about AI and the environment.
Ask what's counted. Direct cooling only, or electricity too? Training included or not?
Ask for the energy figure. Water and energy are linked through WUE, so if you know one and the cooling method, you can sanity-check the other.
Watch the units. Gallons, liters and teaspoons make the same amount sound very different. A figure like 0.000085 gallons looks tiny in part because of all the zeros.
Be wary of vivid comparisons from either side. "A shower" and "an almond" are both rhetorical tools. They're built to land a point, not to inform.
And separate the per-query question from the system question. It's possible for one chat message to be nearly negligible while the total buildout of AI infrastructure is still a serious planning problem for water and power.
Can it improve?
Yes, and it's already happening in some places. Closed-loop liquid cooling, siting data centers in cooler climates, using recycled or non-potable water, and scheduling heavy workloads around local conditions all reduce consumption. More efficient models and better chips lower the energy needed per answer, which cuts water too.
The catch is that efficiency gains often get eaten by growth. When something gets cheaper to run, people use more of it. Whether the net effect goes up or down depends on how fast each side moves.
The takeaway
Altman's figure, about 0.32 milliliters per query, is small, and it may well be accurate for the best-run facilities. But it's one number from one source with limited disclosure, and independent estimates run from a bit higher to many times higher depending on what's counted.
The most useful position is neither "AI is draining the planet" nor "water is a non-issue." It's closer to this: per-query water use is probably modest, the total is real and growing, the local impact matters more than the global average, and we'd all be better off if companies published their methods and measurements so the argument could be settled with data instead of soundbites.