How much water does a data centre drink?

By George Clay · 2026-06-19 · written with AI assistance, reviewed before publishing

Ask how much water a data centre uses and you'll get answers that disagree by a factor of thirty. That isn't because the science is hard. It's because almost nobody measures it, and the one number that decides the whole answer is rarely stated out loud. So let's state it, and let you turn the dial.

The arithmetic

A data centre's water use tracks its electricity. Servers turn power into heat; getting the heat out — usually by evaporating water in cooling towers — is most of the on-site draw, and generating the electricity in the first place evaporates more water at the power station. Put both together and the whole thing collapses to one line: water equals electricity times a water-usage-effectiveness figure, measured in litres per kilowatt-hour.

The convenient part: a terawatt-hour of electricity at one litre per kilowatt-hour is exactly one million cubic metres of water. So the modeled water for any market, in million m³ a year, is just its data-centre electricity in TWh times whatever water intensity you believe. Everything hard about this estimate lives in that one multiplier.

So what's the multiplier?

This is where the factor of thirty hides. Berkeley Lab puts the average US data centre at about 0.4–0.5 L/kWh for on-site cooling alone. The best hyperscalers run closer to 0.2; an evaporative cooling tower in a hot, dry summer can pass 4 and touch 9. None of that counts the water evaporated back at the power plant, which the IEA reckons is the larger share — about two-thirds of the total footprint. Add it in and the all-in intensity lands near 1.4 L/kWh, which is the default below.

The markets we can actually source

Here's the honest limit. There's no free, complete, redistributable dataset of data-centre capacity or power by country — the comprehensive numbers sit behind expensive commercial reports. So this isn't 180 countries; it's the handful where a credible public figure exists. Electricity is reported by the bodies named; the water column is modeled, and it moves when you move the slider.

Electricity figures: IEA Energy and AI (2024), Berkeley Lab (2023), and national statistics offices as noted. Water column: modeled at the intensity set above. Years differ by source.

Is the number big?

At the default, the world's data centres come out around 580 million m³ a year — and that the model lands within a whisker of the IEA's own published estimate of about 560 million m³ is the point: the sum is sound, it's the inputs that wobble. Set against the roughly four trillion cubic metres people withdraw worldwide every year, it's a rounding error — about a hundredth of one percent. Globally, data centres are not draining the planet.

But water isn't a global resource; it's a local one. The trouble is that data centres cluster — in Ireland, where they already pull 21% of the nation's electricity; in the Arizona and Nevada desert; in a few counties of Virginia. A draw that vanishes in the world total can be the marginal straw on a stressed local aquifer, and a cooling tower evaporates most fiercely on exactly the hot, dry days when the surrounding farms and towns want water too. The right denominator, as ever on this site, isn't the planet. It's the catchment.

Read it for what it is

Move the slider to 0.4 and you're looking at efficient on-site cooling; push it to 9 and you're modeling a desert tower in a heatwave. That the answer swings so far on one assumption is the real finding here. Anyone quoting a single, confident, precise figure for data-centre water — in either direction — is hiding the multiplier. We'd rather hand it to you.