We tend to think of AI as a software problem, weightless and waterless, but even a 100-word email written with ChatGPT has been estimated to require roughly a bottle’s worth of water once data-centre cooling and electricity generation are counted — and scaled to global demand, researchers project AI could withdraw 4.2 to 6.6 billion cubic metres of water a year by 2027, roughly half the United Kingdom’s annual freshwater withdrawals.

We tend to picture artificial intelligence as something weightless, a process happening in the abstract, with no physical cost beyond a power bill. It is not weightless, and it is not waterless. By one widely cited estimate, asking a chatbot to write a single 100-word email can account for roughly a bottle of water once the cooling of the data centre and the generation of its electricity are both counted.

That figure comes from work by Shaolei Ren and colleagues at the University of California, Riverside, the group that has done the most to put numbers on a cost the industry rarely discusses. It is worth understanding where the water goes, and also where these numbers are firm and where they are not.

Where the water goes

There are two main places. The first is direct: large data centres throw off a great deal of heat, and many are cooled by evaporating water, which carries heat away as it turns to vapour. The second is indirect: generating the electricity that runs the servers also consumes water, in the cooling towers of thermal power stations and in evaporation from hydroelectric reservoirs.

The bottle-per-email figure counts both. So does the often-quoted estimate that training the GPT-3 model in Microsoft’s United States data centres consumed on the order of 5.4 million litres of water, several hundred thousand litres of it evaporated on site.

This is the part the software framing hides. The query feels like pure information. The hardware behind it sits in a building that has to be kept cool, drawing on a grid that itself runs partly on water.

The numbers are estimates, and they swing

The precise figures deserve care, because they are estimates built on assumptions rather than meter readings.

The water cost of a given task depends heavily on where the data centre is, what time of year it is, and how the local electricity is generated. The same email that, in a 2024 Washington Post analysis with Ren, works out to around 519 millilitres on average, slightly more than a standard bottle, can be far higher in a hot region cooled by evaporation and far lower in a cool climate on a low-water grid. An average across those conditions is useful for a sense of scale, not as a precise reading of any single message.

The forward projection should be read the same way. Ren’s group, in a study titled Making AI Less “Thirsty” and later peer-reviewed in Communications of the ACM, estimates that global AI demand could account for 4.2 to 6.6 billion cubic metres of water withdrawal a year by 2027, which the authors put at around half the annual water withdrawal of the United Kingdom. That is a striking comparison, and it is a projection from one research team, with a range wide enough to signal how much is still uncertain.

The estimates are not the only evidence

Because so much of this rests on one group’s modelling, it is fair to ask whether anything corroborates it from outside.

Something does. The companies running the data centres have reported sharp rises in water use in their own environmental reports, independent of any academic projection. Microsoft’s global water consumption climbed about 34 per cent in 2022, to roughly 6.4 million cubic metres, and Google’s rose about 20 per cent, to around 19.5 million cubic metres, increases both companies and outside analysts have linked to the growth of AI and cloud computing. A United Kingdom government review of water use in data centres and AI reached the same broad conclusion, and made the same complaint about how little is disclosed.

The modelled headline numbers should be held loosely. The underlying trend, rising water use as AI scales, is visible from more than one direction.

Withdrawal is not the same as consumption

One distinction does a lot of work here, and it is easy to miss.

Water withdrawal means water taken from a source. Some of it is returned, warmer or slightly altered, and can be used again downstream. Water consumption means water that leaves the local cycle, mostly through evaporation, and does not come straight back. The headline figure of 4.2 to 6.6 billion cubic metres, and the comparison to the United Kingdom, are about withdrawal. The amount actually consumed is smaller.

That does not make the figure meaningless. It means the fair reading is “water passing through the system on this scale”, not “this much water gone for good”.

Why it still matters

Set against agriculture, which accounts for the great majority of human water use, AI’s global share is small. That comparison is sometimes offered as if it settles the matter. It does not, for two reasons.

The first is concentration. A data centre does not draw its water from the global average. It draws from one local watershed, and a growing number of them sit in dry regions where freshwater is already contested. A use that looks trivial worldwide can be a real strain on a single town’s supply.

The second is the rate of growth, paired with how little gets disclosed. Demand for AI computing is climbing quickly, and the companies running the data centres publish only limited information about how much water each site uses. Most of the public figures are estimates precisely because the measured ones are not shared.

That is the part worth watching, more than any single bottle-per-email statistic. Not whether AI uses water, which it plainly does, but whether the firms building it will report where and how much, so the cost can be weighed where it actually lands.