Data center water usage: the numbers, on site and upstream

For the United States this question has an unusually good answer, because the Department of Energy commissioned Lawrence Berkeley National Laboratory to work it out and the resulting December 2024 report is the most complete public account of data centre water that exists. Most of the figures below come from it. The rest come from the two operators that publish their own ratios, and from two AI companies whose per-prompt figures let you work a ratio out.

Two kinds of water

A data centre uses water in two places. On site, it consumes water to cool the servers, mostly by evaporation in cooling towers or adiabatic systems, and that water leaves the local supply for good. Upstream, the power plants that generate its electricity consume water too: thermoelectric plants boil water and evaporate more to condense the steam, and hydroelectric reservoirs lose water to open-surface evaporation. The report keeps the two apart, and any figure you quote should too. Withdrawal is a third thing again: water that is taken and then returned, where consumption is the part that does not come back.

The US numbers

A few things in that table are worth dwelling on. The on-site total tripled in nine years while the share used by small internal data centres fell from 64% to 12, which means the water moved to a handful of large operators. The upstream water is about twelve times the on-site water, so any figure that stops at the building is missing most of the picture. And the ratio per kilowatt-hour is expected to rise rather than fall, because liquid-cooled AI servers and the shift to hyperscale both consume more water per unit of energy even as they use less energy per unit of work.

Why the ratio varies so much

Water usage effectiveness is litres consumed per kilowatt-hour, and the report models it by facility type. In 2023, small data centres and hyperscale sites both came out near 0.32 litres per kilowatt-hour, midsize and colocation sites at 0.67, and AI-specialised sites at 0.61. The spread is the cooling design and the climate. Evaporative systems trade water for electricity, so a site can have a low power usage effectiveness and a high water figure, or the reverse, and the same design in a hot, dry place evaporates far more than in a cool one. The report notes that some hyperscale operators report values of 0.1 to 0.3 for systems it models higher, and that liquid cooling for AI servers pushes the number up.

What the operators say

Microsoft publishes its fleet figure: 0.30 litres per kilowatt-hour in fiscal 2024 and 0.27 in fiscal 2025, measured against the electricity used by the IT equipment, with regions from 0.03 in Europe, the Middle East and Africa to 0.34 in the Americas, and a design target of zero water for cooling in its next generation of sites. Google reports consumption of about 6.1 billion gallons in its data centres in 2024, and says it replenished 4.5 billion gallons, 64% of its freshwater consumption, up from 18% the year before. Google’s research paper on Gemini gives a different kind of number: 0.26 millilitres of water against 0.24 watt-hours for a median text prompt, which works out to about 1.1 litres per kilowatt-hour. OpenAI’s stated 0.000085 gallons against 0.34 watt-hours works out to about 0.9.

Upstream is the bigger draw

The report’s indirect figure, nearly 800 billion litres in 2023, comes from mapping each data centre to its grid’s balancing authority and applying the water consumed per kilowatt-hour of generation there: 4.52 litres on average for the mix supplying data centres, a little above the 4.35 for the country as a whole. That is the water behind the electricity, and it depends on the fuel and the plant rather than on the data centre. A site on a hydro-heavy grid can carry a higher upstream water figure than a site on gas, and the county-level maps in the report show the intensity varying several-fold across the country. This number falls as the grid decarbonises, and better cooling at the data centre does nothing to it.

Which ratio to use for AI

If you are turning AI usage into water, you need a litres-per-kilowatt-hour figure, and there are four defensible choices. Berkeley Lab’s 2023 US average of 0.36 is the best figure for a generic US data centre. Google’s implied 1.1 and OpenAI’s implied 0.9 are the best figures for a frontier model served by a hyperscaler, since they come from those fleets. Microsoft’s 0.27 is the best figure for that fleet. The older 1.8 from the 2016 report is what CrbonFree’s product uses today, and it now sits at the top of the range; the calculator below defaults to it for consistency with the product and lets you pick any of the others. Whichever you pick, add the upstream water separately, and say which ratio you used.

Run it for your own usage:

What would bring it down

On-site water can be cut in two ways that pull in opposite directions. Closed-loop and air-cooled systems cut water to near zero and cost electricity, which is why Microsoft can report 0.03 litres per kilowatt-hour in one region and why its new sites are designed for zero cooling water. Siting in cool, wet climates cuts evaporation without the electricity penalty, which is the UC Riverside team’s point about where and when a model runs. Upstream water falls only as the grid does, plant by plant. And on the demand side, the levers are the same as for energy: fewer tokens, cached prompts, and smaller models for routine work, since a kilowatt-hour that is never used never needs cooling.