The Water Cost of Compute — Why Data Centres and Water Stress Collide

There's a number that rarely makes the press release. When a new AI data centre is announced, the headline figure is power — hundreds of megawatts, sometimes more than a small city draws. But all of that electricity ends up as heat, and a great deal of the world still carries that heat away with water. A single large site running evaporative cooling can turn millions of litres a day into vapour. So behind every megawatts-and-jobs announcement sits a quieter question: is there water where this is going?

This guide is about reading the water-stress map for that answer — and about why the new wave of data centres so often lands in exactly the places where water is already tight.

Compute is a thermal problem

It's easy to think of a data centre as a building full of computers. It's more useful to think of it as a building full of heaters. Every watt of electricity a server draws comes back out as heat, and the dense racks used for AI training throw off far more heat per square metre than the web servers of a decade ago. Keeping the chips inside their safe temperature range is half the engineering of a modern facility.

There are a few ways to do it, and they trade water against energy. Evaporative (or "swamp") cooling sprays or trickles water so that, as it evaporates, it pulls heat out of the air — cheap and effective, but it consumes the water. Air cooling uses fans and chillers instead, sparing water but spending more electricity. Newer closed-loop liquid systems circulate the same coolant over and over, and some sites run on recycled or non-potable water. Across the fleet that already exists, though, a large share of cooling still leans on fresh water — and that's before counting the water used at the power plants supplying the electricity, many of which need cooling water of their own. (That second link is why the power-plants layer belongs in the same conversation.)

The upshot is simple: more compute, in most of today's world, means more water demand — and where you put the compute decides whose water it draws on.

What "water stress" actually measures

The map colours the world by baseline water stress, the headline indicator from the World Resources Institute's Aqueduct project. The idea is a ratio: how much water people withdraw from a basin, divided by how much renewable water that basin has. Aqueduct sorts every watershed into bands — low is under a tenth of the available water withdrawn; extremely high is over four-fifths, meaning almost all the renewable supply is already spoken for and there's almost no slack for a drought or a thirsty new user.

A class apart is arid & low water use — the deserts. These basins have very little water, but they also have very little demand, so they're neither comfortably supplied nor overdrawn; the map gives them their own neutral shade rather than forcing them onto the stress ramp. And where the underlying data runs out, a basin is simply left unfilled rather than shown as a guess, so empty land means low stress or not assessed, never "we made something up." That's the honesty rail this whole project runs on.

Across the roughly 15,800 assessed watersheds, a little over 17% already sit at medium-high stress or worse — a minority of the map, but a consequential one, because that's where almost all the friction lives.

Why the map is drawn by watershed, not by country

The single most important design choice here is the unit. It would be easy to colour countries — the data exists at that level — but a national average is exactly the wrong lens for this question, because water doesn't follow borders.

Take the United States. As a whole it's moderate. But that average quietly blends the water-rich East with the chronically overdrawn basins of the Southwest, and a country-coloured map would paint Phoenix the same shade as Pittsburgh. The entire point of the overlay — that a specific cluster of data centres sits in abundant or in scarce water — would vanish into the average.

So this map keeps Aqueduct's native unit, the sub-basin: a watershed defined by the land that drains to a common point, not by a line on a political map. It's both the honest choice and the practical one. It's the resolution at which you can put a data-centre dot down and read its real water situation — which is the whole reason to fuse the two layers.

That sub-basin view does throw up a couple of results that look surprising until you understand them. Las Vegas reads low, despite its reputation, because it's a small demand sitting in an enormous, near-empty desert basin — its genuine constraint is its slice of Colorado River water, a matter of reservoirs and interstate law rather than local supply. Singapore reads low because it imports much of its water and recycles relentlessly — a real exposure that a basin-supply number can't see. Neither is a bug; both are reminders that this is a map of physical basin stress, one lens on water risk, not the whole of it.

The collision: where compute meets scarcity

Now lay the two layers over each other, and the reason for all of this becomes obvious.

The data centres cluster where land, power and fibre are cheap and where the customers are — and those places are not chosen for their water. The result is a set of striking contrasts. Northern Virginia's Ashburn — "Data Centre Alley," the densest concentration of server halls on Earth — sits on comfortably low stress. The fast-growing campuses around Phoenix, drawn by cheap land and solar power, sit in extremely high stress. Same country, opposite ends of the water budget.

Zoom out and the pattern hardens. Rank the world by the sheer area of land under high or extremely high stress and the leaders are China, the United States and India — which are also, not coincidentally, the three countries building the most compute. The new data centres and the scarce water are converging on the same map. That's not a forecast of disaster; plenty of those projects will use air cooling, recycled water or sites with genuine supply. But it's precisely the tension this overlay exists to make visible — and it's why "how many megawatts" is only ever half the question.

How to use the layer

Open the Grid map, switch on Water Stress, and then switch on Data Centres as well. Read the watersheds first — the dark globe stays calm where water is abundant, and glows amber to red where withdrawals are pressing against supply. Then read the data-centre bubbles against that backdrop: a dot on dark is a campus with water to spare; a dot on deep red is one competing for a basin that's already nearly fully drawn.

It pairs naturally with the rest of the Grid. The power-plants layer adds the generators that also drink water; the carbon-intensity layer adds how clean the electricity behind it all is. Together they're a way of asking the fuller question about the build-out — not just where the compute is going, but what it costs the place it lands in.