How Much Water Does ChatGPT Use? (2026 Guide)

You have probably seen the claim that asking ChatGPT to write a short email costs a bottle of water. You may also have seen the counter-claim that a query uses a fifteenth of a teaspoon. Those two figures differ by a factor of roughly fifteen hundred, and both get cited confidently.

They cannot both be right, and the gap between them is not a rounding disagreement. It comes from measuring different things, at different points in the supply chain, for different models, in different years.

This guide sets out what each published number actually counts, where the viral figure came from, and what a per-query water estimate can and cannot tell you.

Every figure below is attributed, because in this topic the source matters more than the number.

The Short Answer

The best available per-query estimates cluster in the range of a fraction of a millilitre to a few tens of millilitres, depending on what is counted.

OpenAI's own published figure is the lowest. In a June 2025 post, Sam Altman wrote that "the average query uses about 0.34 watt-hours" of energy and "about 0.000085 gallons of water; roughly one fifteenth of a teaspoon." That works out to roughly 0.32 millilitres.

Google published a comparable figure for its own assistant. In August 2025 it reported that the median Gemini Apps text prompt consumed 0.24 watt-hours of energy and 0.26 millilitres of water, which it described as about five drops.

Academic estimates run higher because they count more. A widely cited paper puts the United States average at around 16.9 millilitres per request for GPT-3, of which only about 2.2 millilitres is consumed at the data centre itself.

The "bottle of water per email" claim is not supported by the paper it is usually traced to. That paper said 500 millilitres per ten to fifty responses, not per response.

Publishing content about AI and wondering whether it is actually reaching anyone? Get a free SEO audit and see where your visibility really stands.

Where the Water Actually Goes

The single biggest source of confusion is that "water used by ChatGPT" describes two completely separate flows. Researchers label them scope 1 and scope 2, and most published arguments are comparing one against the other without saying so.

Scope 1: cooling the building

Data centres generate heat and many of them shed it by evaporating water in cooling towers. That water leaves as vapour and does not return to the local system, which is why researchers call it consumed rather than merely withdrawn.

This is the flow people picture when they imagine a server drinking water. It happens at the facility, it is measurable, and it is the number an operator has most direct control over.

Operators track this with a metric called water usage effectiveness, which expresses litres consumed per unit of computing energy delivered. It is the water-side equivalent of the power efficiency ratio the industry has reported for years.

It also varies enormously by site and season. A data centre using closed-loop or air cooling may consume almost none on site. One using evaporative cooling in a hot, dry month consumes a great deal.

Scope 2: generating the electricity

Power stations also consume water, mostly for cooling their own thermal cycles. Every watt-hour a data centre draws therefore carries an invisible water cost somewhere upstream.

In the academic estimates this is the larger share by some distance. Of that roughly 16.9 millilitres per GPT-3 request for the US average, about 14.7 millilitres is off-site water attributed to electricity generation rather than water used at the data centre.

That single split explains most of the apparent disagreement between company figures and academic ones. A number that counts only the building will always be far smaller than one that counts the power station too.

Why location changes the answer

The same query costs different amounts of water in different places, because both the cooling method and the local generation mix differ.

The same research reports a spread across US states, from roughly 7.6 millilitres per request in Texas to around 47.5 millilitres in Washington. A six-fold difference for identical computation.

Counter-intuitively, a colder and wetter state can score worse. What drives the off-site share is the generation mix and the cooling technology at the power stations, not the weather outside the data centre.

Routing also moves during the day. A provider balancing load across regions is quietly changing the water cost of identical queries hour by hour, which is another reason a single figure is a simplification.

So "how much water does a query use" has no single answer even in principle. It depends where the query was served.

What Each Published Number Measures

Once you know the scope distinction, the headline figures stop contradicting each other.

OpenAI's figure

Altman's 0.000085 gallons is a per-query average with no published methodology behind it. OpenAI has not released the calculation, the model mix it covers, or whether it includes off-site water.

It is also a single average across every query type. A one-word reply and a long reasoning task are not comparable units of work, and averaging them hides most of the variation that matters.

Treat it as a company statement rather than a measurement you can audit. It is directionally useful and not independently verifiable.

Google's figure

Google's number is better documented. Its August 2025 methodology covers full system power including idle machines provisioned for reliability, CPU and RAM, and data-centre overhead such as cooling.

Google is also explicit about the limits. It describes the result as a point-in-time analysis based on May 2025 data, says the figure does not represent every prompt, notes that it will change as models and usage evolve, and states that the data has not been verified by an independent third party.

Crucially, the 0.26 millilitres is data-centre water. It is a scope 1 style number, which is why it lands so close to OpenAI's and so far below the academic estimates.

The academic estimates

The most cited peer-reviewed work here is "Making AI Less Thirsty" by Pengfei Li, Jianyi Yang, Mohammad A. Islam and Shaolei Ren, revised in March 2025.

It models GPT-3 rather than current models, and it counts both scopes. It also puts training in context: roughly 700,000 litres of on-site water for GPT-3's training, rising to about 5.4 million litres once off-site water is included.

Two things follow. The per-query numbers in that paper describe a model from 2020, and hardware and model efficiency have improved substantially since. And because it counts off-site water, it is not comparable with Google's figure at all.

Not sure whether your content is being cited by AI assistants at all? Get a free SEO audit and find out how visible you are in AI search.

Where the Bottle-of-Water Claim Came From

The claim that a single short email costs roughly half a litre of water is the most repeated figure in this debate and the least defensible.

It appears to come from reading the "Making AI Less Thirsty" estimate as per response when the paper said 500 millilitres for ten to fifty medium-length responses. That alone inflates the figure by between ten and fifty times.

There is a second compounding error. Analysts retracing the arithmetic have argued that the underlying energy figure was per page of generated text, and that the calculation assumed something like ten to seventy pages of work per query. A real conversation is closer to one or two pages.

Stack the conversation-length error on top of efficiency gains since GPT-3 and the corrected estimate for a typical interaction lands near five millilitres rather than five hundred.

That does not make the topic trivial. It does mean the specific viral number is an artefact of arithmetic rather than a finding.

It is also a useful reminder about how statistics travel. The inflated figure spread because it was vivid and quotable, while the correction is a paragraph of unit analysis that no one screenshots.

Chasing a topic where the facts keep shifting? Get a free SEO audit and see which of your pages are actually earning rankings.

How to Read the Number Sensibly

A per-query figure is a useful unit and a poor argument. Three things keep it honest.

• Scope, always. Ask whether a figure includes off-site water for electricity. A tenfold difference usually means someone changed the scope, not the technology.
• Model and year. GPT-3-era estimates describe a model several generations old. Company figures describe current systems but without auditable methodology.
• Per-query versus aggregate. A tiny number multiplied by an enormous number of queries is not necessarily a tiny number. The individual and infrastructure questions are separate.
• Local conditions. Water consumed in a water-stressed region matters differently to the same volume consumed where supply is abundant. Litres are not fungible.
• Training versus inference. Most per-query figures cover inference only. Training is a large one-off cost amortised across an unknown number of future queries.

None of this means the question is unanswerable. It means the honest answer is a range with its assumptions attached, which is less satisfying than a headline and considerably more useful.

Anyone quoting a single decisive number, high or low, is leaving at least one of those out.

What the Per-Query Figure Does Not Settle

Even a perfectly accurate per-query number leaves the interesting questions open.

It says nothing about total consumption, which depends on query volume that neither company publishes in a form you can multiply. It says nothing about where that consumption lands geographically, which is what determines whether it affects anyone's supply.

It also says nothing about the counterfactual. The relevant comparison is not a query against nothing, but a query against whatever it replaced, which might be several web searches, a video, or a task that would have consumed more.

And it says nothing about growth. Per-query efficiency has improved while total usage has risen sharply, and those two trends point in opposite directions.

Want to know whether AI search is sending you traffic or taking it? Get a free SEO audit and get a straight answer on your search performance.

Final Thoughts

On the published evidence, a single ChatGPT query consumes a very small amount of water. OpenAI puts it at roughly a fifteenth of a teaspoon and Google's comparable figure for Gemini is about five drops, both counting water used at the data centre.

Academic estimates running to tens of millilitres are not contradicting those. They include the water consumed generating the electricity, and they model an older, less efficient generation of models.

The bottle-of-water claim does not survive contact with its own source. It came from reading a per-ten-to-fifty-responses figure as per response, and the correction is roughly a hundredfold.

It is worth noticing which direction the uncertainty runs. The figures that look alarming are mostly old or broadly scoped, and the figures that look reassuring are mostly unaudited. Neither camp is being straightforwardly dishonest.

None of which settles whether AI infrastructure water use is a problem overall. That is a question about totals, locations and growth rates, and a per-query number is simply the wrong instrument for it.

Model choice affects the energy and water cost of a query too, since a heavier reasoning model does more work per answer, and our guide on which ChatGPT model you should use covers when the heavier options are actually worth it.

If you take one thing away, make it the habit of asking what a water figure counts before repeating it.