GRID · WATER STRESS
The Water Cost of Compute
A data centre is a thirsty machine — the servers run hot, and much of the world cools them with water. So the question of where the AI build-out is going up has a second half: is there water there?This layer recolours the world’s watersheds by baseline water stress— how much of a basin’s renewable supply is already being withdrawn — from WRI Aqueduct 4.0. Switch on Data Centres too and the contrast is stark: Northern Virginia’s Ashburn, the densest data-centre cluster on Earth, sits on abundant water, while the fast-growing campuses around Phoenix sit in extremely high stress. Across the 15,833 assessed sub-basins, 17.2% already face medium-high stress or worse.
From abundant to overdrawn
Aqueduct sorts each sub-basin into one of six classes by the share of its renewable water already withdrawn. Most of the world runs low— but stress is concentrated, and where it bites it bites hard. Arid & low water use is a class apart: deserts with little water and little demand, neither safe nor overdrawn. The bars below are the share of assessed sub-basins in each class.
Where the stress concentrates
Ranked by the sheer area of land under high or extremely highstress, the map is led by the big, dry, heavily-irrigated economies — and, tellingly, by the three countries building the most compute. China, the United States and India top both lists, which is exactly why the overlay matters: the new data centres and the scarce water are landing in the same places.
How to read it
The unit here is the watershed, not the country — and that’s deliberate. Water stress is a property of a river basin, and a national average hides the story: the United States as a whole is moderate, but Phoenix’s basin is extremely high while Virginia’s is low. Because Aqueduct measures local supply against localwithdrawals, a few results look surprising and are worth understanding rather than “fixing”: Las Vegas reads low because it is a small demand in a vast, near-empty desert basin (its real constraint is Colorado River allocation, a matter of politics and reservoirs, not local supply), and Singaporereads low because it imports much of its water — an exposure this indicator doesn’t capture. So treat the map as a map of physical basin stress, not of every kind of water risk.
This is a baseline risk index (Aqueduct’s 2023release), not a live feed — it doesn’t change minute to minute, and we badge it as exactly that. Sub-basins with insufficient data are dropped rather than shown as a value, so unfilled land means low stress or not assessed. The data is WRI Aqueduct 4.0, used under CC BY 4.0.