How much water does an AI data center use? The numbers, and the two different questions

Peer-reviewed research estimates that training the GPT-3 language model in Microsoft's US data centers could directly evaporate 700,000 litres of clean freshwater, and projects that global AI demand will account for 4.2 to 6.6 billion cubic metres of water withdrawal in 2027. Those are the best public numbers available. They are also modelled estimates, and they measure two different things that are routinely confused.

Why a data center uses water at all

To get rid of heat. Every watt of electricity a facility draws becomes heat that has to leave the building, and water is an unusually good way to carry it.

The reason is physics rather than engineering preference. Evaporating water absorbs a large amount of energy, so passing warm air over wet surfaces and letting the water evaporate removes a great deal of heat for very little electricity. That is evaporative cooling, and it is efficient on power and expensive on water.

The alternatives invert the trade. A closed-loop system circulates the same coolant repeatedly and rejects heat elsewhere, and air cooling moves heat with fans alone. Both lose far less water and consume more electricity to remove the same heat.

So the question how much water does a data center use has no single answer. It is determined by the cooling design, and the design is chosen against local electricity prices, local climate and local water availability.

The two numbers people mix up

This is the single most important distinction in the subject, and almost every alarming headline about it gets it wrong.

Withdrawal is water taken from a source. Much of it is returned afterwards, often warmer than it arrived, which has its own environmental implications but is not a loss of water.

Consumption is water that does not come back, almost always because it evaporated. A facility can withdraw an enormous volume and consume a small fraction of it.

Reporting a withdrawal figure as though it were consumption overstates the loss by a wide margin, and it happens constantly. When you encounter a number for AI and water, the first thing to establish is which of the two it is.

What the research actually estimates

The main work in this area is by Pengfei Li, Jianyi Yang, Mohammad Islam and Shaolei Ren, who set out to quantify a footprint they argue had been ignored while attention went to carbon.

Their headline figures: training GPT-3 in Microsoft's state-of-the-art US data centers can directly evaporate 700,000 litres of clean freshwater, and global AI demand is projected to account for 4.2 to 6.6 billion cubic metres of water withdrawal in 2027.

They scale the second figure deliberately. That projected withdrawal is more than the total annual water withdrawal of four to six countries the size of Denmark, or about half that of the United Kingdom.

Their stated point is not that AI is uniquely thirsty. It is that the number was unknown, and that the information had been kept out of public view while the carbon figure was being scrutinised.

Direct and indirect, and why the distinction matters

A facility's water footprint has two halves and they move in opposite directions.

Direct water is used on site for cooling. Indirect water is used to generate the electricity the facility consumes, because thermal power generation withdraws and consumes substantial water for its own cooling.

That coupling produces a result people find counterintuitive. A data center that switches from evaporative to air cooling reduces its direct water consumption and increases its electricity draw, which increases its indirect water footprint. Whether the total improves depends entirely on how the local grid generates power.

Any complete figure therefore has to say which halves it counts. A site-only number flatters a facility on a fossil-heavy grid, and a full supply-chain number is much harder to establish.

What operators do and do not publish

Very little at the level that would settle any of this, which is the gap the research was written to expose.

Several large operators publish annual environmental reports with company-wide water figures. Those are useful for a trend and close to useless for a specific question, because they aggregate facilities across many climates, cooling designs and water sources into one number.

What is almost never published is per-facility water withdrawal and consumption, broken out by source, with the cooling technology stated. That is the level at which a local authority, a neighbour or a journalist could actually assess a site, and it is also information that would cost an operator nothing in competitive terms to disclose.

The practical consequence is that researchers have to model what could be measured. Every estimate in this article exists because the measured number is not public, and the error bars on all of them would shrink substantially if it were.

Why cooling design is the whole answer

Because the same compute can produce wildly different water numbers depending on one engineering choice.

An evaporative system is cheap to run on electricity and consumes water by design, since evaporation is the mechanism. A closed-loop system recirculates coolant and rejects heat through a dry cooler or chiller, consuming very little water and drawing more power. Air cooling sits at the far end of that trade.

Hybrid designs switch between modes with the weather, using evaporation on the hottest days when the electricity cost of dry cooling is highest, and running dry the rest of the year. A facility like that has a water figure that is seasonal rather than constant, which is also when water availability is tightest.

That is the part of the subject that rewards attention. A commitment to reduce water use tells you almost nothing unless it says which design is being adopted and what happens to the electricity draw.

How much to trust a per-query water figure

Less than the precision implies, and the reason is methodological rather than motivated.

A per-query figure is produced by dividing a modelled facility-level water use by a modelled query volume. Both inputs rest on assumptions about cooling technology, local climate, seasonal variation, electricity mix and utilisation that the operators do not publish.

The direction of the claim is sound: inference consumes electricity and cooling, and both have a water footprint. The specificity is not. A claim that one query consumes a precise volume of water is a modelled estimate with large error bars presented as a measurement, which is a familiar pattern in AI claims generally.

Why the local question matters more than the global one

Globally, AI water use is small against agriculture and thermal power generation. Locally, it can be the dominant new demand on a stressed system, and that is where the real arguments happen.

Three variables decide whether a facility's water use is a problem. The watershed it draws from and how stressed that source already is. The season, because a drought-year draw matters differently from an annual average. And the water quality, because evaporating treated drinking water is a different act from using recycled or non-potable supply.

That is why two facilities with identical annual figures can be entirely uncontroversial and the subject of a planning fight respectively. The volume is the least informative part of the number.

What to ask about a specific facility

Four questions, in order, and all four are answerable if an operator chooses to answer them.

Is this figure withdrawal or consumption. Does it cover the site only, or the electricity supply chain as well. What cooling technology is in use, since evaporative, closed-loop and air produce very different numbers for the same compute. And what is the source, its stress level, and whether the water is potable.

The reason the researchers framed their paper around disclosure is that these four questions are usually unanswerable from public information, and nothing about them is commercially sensitive.

The short version

The best estimate in the literature is that training GPT-3 in Microsoft's US facilities could directly evaporate 700,000 litres of clean freshwater, with global AI water withdrawal projected at 4.2 to 6.6 billion cubic metres in 2027, more than four to six Denmarks' worth or about half the United Kingdom's.

Those are modelled figures, they describe withdrawal and consumption as separate things, and they split between the building and the power station behind it. Any number quoted without those distinctions is not wrong so much as unreadable.

This article reports published research estimates. Figures were verified on 20 October 2026.