Water Is the Data Center Constraint Nobody Puts in the Press Release
Every announcement of a new AI campus leads with megawatts. Power is the number that gets negotiated in public, fought over at the utility commission, and quoted in the analyst note. It’s also the constraint that money can eventually solve, because you can build generation, sign a nuclear offtake, or put turbines on the site.
Water is harder, and it decides more siting arguments than the public record suggests.
A large data center running evaporative cooling can consume millions of litres a day at full load. The word consume is doing real work there. Some of the water returns to the system, and a meaningful fraction leaves as vapour and does not come back to the local basin. In a wet region with a big river that’s an accounting detail. In the arid west, in central Spain, in parts of Chile and Arizona and northern Mexico, it’s the reason a project dies in a county meeting.
Operators have three ways out and all of them cost something. Closed-loop and air cooling cut water use sharply while consuming more electricity, which converts a water problem into a power problem in a market where power is already the binding constraint. Recycled and non-potable water works well where the municipal infrastructure exists to deliver it, and that infrastructure is uncommon outside a few metros that built it deliberately. Direct-to-chip liquid cooling, which the current generation of accelerators is pushing everyone toward anyway, changes the thermal math substantially and is arriving fastest in new builds rather than retrofits.
The political dynamics deserve attention separately from the engineering, because they don’t move at the same speed. Electricity consumption is invisible to a resident until the bill arrives. Water is visible immediately and it’s personal. A community that shrugs at a gigawatt will organise against a project that draws from the same aquifer as the farms, and it will do so faster, with better local press, and with legal standing that power opposition rarely has. Water rights in the American west are property, adjudicated, senior and junior, and a developer buying them is buying them from somebody who used to farm.
This is why Texas asked for water plans alongside ownership disclosure when it froze new grid connections. It’s why Chile and the Netherlands have both slowed projects that had power lined up. The pattern is consistent: the megawatt negotiation happens with a utility and a regulator, the water negotiation happens with a community, and the second one has no clear counterparty who can simply say yes.
For anyone modelling where capacity actually lands over the next five years, the useful screen isn’t a map of cheap power. It’s the overlap of cheap power, cool climate, and a basin with slack in it. That overlap is smaller than the power map alone, which is part of why the Nordics, the upper Midwest and parts of Canada keep appearing in siting decisions that look expensive on a pure energy basis.
The buildout is being modelled almost entirely on electricity. The constraint that stops individual projects is frequently the other one.