Mapping 9,347 operational facilities, researchers found most of them competing with cities rather than farms, and a third landing in water-deficit regions once warming reaches 3C.
The electricity a data centre consumes is a known quantity. It is metered, billed and reported, and the resulting figures have driven years of argument about computing and the grid. Water has had no equivalent accounting, and the reason is mundane rather than sinister.
“The energy footprint of data centers has been quantified in detail because electricity is metered and reported,” said Lorenzo Rosa, the lead author of a study in Environmental Science & Technology. “Water scarcity is measured locally, inconsistently, and often not at all. So, we built the map because it did not exist.”
The map covers 9,347 operational data centres worldwide, placed against local water availability. About a quarter of them sit in areas that experience water stress for at least one month a year, and around 6 per cent in areas stressed for most of the year.
Why data centres drink
The water goes to cooling. Servers convert essentially all the electricity they draw into heat, and that heat has to leave the building.
The cheapest way to move it is evaporative cooling: water is sprayed or trickled across a surface and the phase change carries heat away with the vapour. It is efficient and it is consumptive — the water leaves as vapour rather than returning to the river it came from, which is the distinction between withdrawal and consumption that makes data centre water use more significant than the raw volumes suggest.
The alternative is closed-loop or air cooling, which consumes almost no water and instead consumes more electricity, particularly in hot weather. Operators therefore have a dial between water and power, and in most places the economics have pointed toward water.
Who they are competing with
The study’s more unusual contribution is identifying the other party. Near 41 per cent of these data centres, cities and municipalities are the largest water user. Agriculture is the primary consumer near another 28 per cent.
“These facilities are not landing in empty landscapes,” Rosa said. “They are landing on top of the water supply of cities and farms.”
That split matters because the two competitions resolve differently. Against agriculture, a data centre is competing with a user whose water is cheap, whose volumes are large, and whose political position in a drought is often protected by historical rights. Against a municipality, it is competing with household supply, which is the use that wins public arguments and which regulators are least willing to see curtailed.
The siting logic that produced this is not mysterious either. Data centres want cheap power, fibre connectivity, cool climates where possible, and proximity to the populations they serve. None of those criteria is water availability, and the first and last of them tend to point toward places where people already live.
Three degrees
The forward-looking part of the study considers a world with average temperatures 5.4°F — 3°C — above 1900 levels, which the release places 10 to 40 years from now.
Under that scenario, a third of today’s data centres would sit in regions where water demand exceeds supply. Most of the newly affected facilities are in the United States, which already holds 39 per cent of the global total.
The wide range on the timing, 10 to 40 years, is worth noticing rather than glossing. Three degrees is well above the Paris Agreement’s targets and within the range of current trajectories depending on which assumptions are made about emissions and climate sensitivity, so the uncertainty in the date is real and the scenario is not an outlier.
The concentration of risk in the United States follows partly from arithmetic — with 39 per cent of the world’s facilities it would take a large share of any new exposure — and partly from geography, since much of American data centre construction has gone to the arid and semi-arid West and Southwest, where land and power have been cheap and water has not been the binding constraint it is becoming.
What the study does and does not settle
This is a siting analysis, not a consumption audit. It establishes where facilities sit relative to water stress and who else draws on that water; it does not report how much water the facilities actually use, which varies enormously with cooling design, climate and operator practice.
That distinction cuts both ways. A data centre in a water-stressed region running closed-loop cooling may consume almost nothing locally, while one in a wet region using evaporative cooling consumes a great deal. Proximity to stress is a risk indicator rather than an impact measurement, and the study is better read as a map of where conflict is likely than as an accounting of harm already done.
The release is also thin on the research team. Rosa’s institutional affiliation is not given, and his collaborators are mentioned but not named — unusual for a study of this profile, and an omission that makes the work harder to situate.
One disclosure deserves attention. The authors acknowledge support from Saoradh Enterprise Partners LLC for access to the data centre dataset, which the release frames as data access rather than a funding grant. Where the underlying facility list comes from matters for a study whose central claim is a count, and a commercial data provider is a different proposition from a public register. No other funding is stated.
What the map supplies, for the first time in one place, is a basis for asking the question locally. Whether a particular facility is a problem depends on its cooling system, its region’s hydrology and who else is drawing from the same source — and until now there was no systematic way to know which facilities were even worth asking about.
Notes
- Rosa, L., et al. Environmental Science & Technology, 2026. DOI: 10.1021/acs.est.6c08088.
- American Chemical Society. “Many data centers are located where water resources are limited.” EurekAlert!, 19 October 2026. https://www.eurekalert.org/news-releases/1148339.
Leave a Reply