AI uses water mainly to keep hot processors cool inside data centres, and some of that water is lost to the air as it evaporates. How much any one model or query uses is another matter. The answer changes with the building, the local climate and the time of year, and there is still no agreed way to measure it.
more
Where the water actually goes
Chips in a data centre run hot. One common way to cool them is to pump water past them, then carry the warmed water to a cooling tower, where part of it evaporates away.
That water is often drinkable. WIRED reports that salty or brackish water corrodes equipment, so many operators draw on potable municipal supplies. Amazon, Meta and Apple have been turning more to treated municipal wastewater instead.
Cooling is not the whole footprint. Speakers at an OECD and IEEE event said AI’s water use also reaches into making AI hardware and into generating the electricity that powers it, since some power plants use a lot of water.
A trade between water and power
Every data centre makes a choice. Using more water lets a site run its electric chillers less, while leaning on electricity cuts the water bill but raises the power bill and the greenhouse gas emissions that come with it.
There are cooling liquids that save on both. WIRED notes that some large tech firms have been wary of them, because they could bring so-called forever chemicals into the process.
The weather matters too. Hot summers push cooling demand up, so a data centre may draw more water, or more power, in those months. Fengqi You, a professor of energy systems engineering at Cornell, told WIRED that every location and every state is different.
Why the popular numbers wobble
You may have seen the claim that an email written with AI costs a whole bottle of water. WIRED argues that pinning a figure on an “average” query is far harder than such claims suggest.
One high-profile figure went badly wrong. Journalist Karen Hao wrote in her book Empire of AI that a proposed Google data centre near Santiago, Chile, might need more than a thousand times the water of the town’s whole population. A mix-up over units appears to have inflated the number by a factor of 1,000, and Hao acknowledged the error in a Twitter thread. Andy Masley, who runs an effective altruism group in Washington, DC, flagged it after months of questioning AI water statistics on his Substack.
Both sides of the argument use numbers loosely. In a Fox News op-ed, the co-chairs of the AI Infrastructure Coalition, an industry group, called AI’s water use minimal and smaller than that of America’s golf courses. More than 230 environmental groups, meanwhile, wrote to Congress warning that data centres threaten Americans’ water security.
Why nobody agrees on a measure
The measuring problem is older than the rows about it. On 12 February 2025 the OECD and IEEE held a session on AI’s water footprint alongside the French AI Action Summit. Its moderator said water had drawn less attention than energy, partly because the sums are complex and there are no commonly accepted metrics. Water use is also absent from regulatory compliance rules, which may be why it stays under-measured.
The IEEE P7100 working group is trying to change that. It is developing a technical standard for measuring AI’s environmental impact, and the event’s findings feed into that work.
Location may matter more than any global total. Unlike carbon, whose harm spreads everywhere, water stress is local, so the same data centre could strain one region and barely register in another with fuller reservoirs. Speakers at the OECD event made the same point about places where water is already scarce.
So the useful question is rarely how much water AI uses in general. It is how much a particular data centre draws, from which supply, in which summer.
