What a Data Centre Actually Does With Water

Every watt a server draws becomes heat, and that heat has to leave the building. Cooling is not a side concern in a data centre; it is the second engineering problem after power, and it is the one that decides whether a facility consumes water.

The useful thing to understand is that water is not an extra cost layered on top of electricity. It is largely a substitute for it. A facility can shed heat by evaporating water or by spending energy on refrigeration, and the design choice moves consumption between those two columns.

Why water, specifically

Evaporation carries away a large amount of heat per unit of water, and it does so without a compressor. That physical fact is the whole reason water is in the picture: it is the cheapest way to move heat out of a building, in energy terms.

So a facility that reports low power usage effectiveness may be achieving it partly by evaporating water, and a facility that reports near-zero water consumption is often spending more electricity to do the same job. Comparing two sites on either metric alone tells you which column they chose, not which is more efficient overall.

The three designs

Design	Water	Electricity	Where it fits
Open-loop evaporative	High, consumed on site	Low	Hot climates, water available
Closed-loop / dry cooler	Very low	Higher, rises with ambient temperature	Cool climates, water-stressed regions
Hybrid	Variable, seasonal	Variable	Most large modern builds

Open loop passes warm water across a large surface in an air stream. Some evaporates, carrying heat with it, and the cooled remainder returns to the building. The evaporated portion is consumed. These systems also need periodic blowdown, discharging concentrated water to control mineral build-up, which is a second, less-discussed loss.

Closed loop recirculates the same fluid and rejects heat to air through a dry cooler, or through a chiller when ambient temperature is too high for that to work. Little or no water is consumed. On hot days the chiller runs and the energy cost climbs sharply.

Hybrid designs switch between the two by season or by hour, which is why a single annual figure for a facility can conceal a wide operating range.

The trade never disappears

An air-cooled facility does not remove water from the system. Thermoelectric generation consumes water, so the electricity that runs the chillers carries an embedded water cost at the power plant.

What changes is whose accounts it appears in. On-site evaporation shows up in the data centre's water reporting. Water consumed generating its electricity shows up in the power sector's. This accounting boundary is the single largest reason published per-prompt water estimates differ by three orders of magnitude — some count only the first column, some count both.

Neither boundary is wrong. Reading a number without knowing which one was used is.

Why location decides more than design

Two identical facilities in different places will not behave alike, for two independent reasons.

Climate sets how many hours a year free cooling is available and how fast water evaporates. A cool, humid site can run dry for much of the year. A hot, arid one cannot, and faces the sharpest version of the trade: the places where evaporative cooling is most effective are frequently the places where water is least available.

Grid mix sets the embedded water behind each kilowatt-hour, which varies by region and by hour of day.

This is why siting decisions are made on power and water availability rather than on proximity to users, and why those constraints, alongside interconnect, determine what a large cluster can actually be built to do.

What liquid cooling does and does not change

Direct-to-chip and immersion cooling are often described as a water story. They are mainly a density story.

They move heat away from very dense hardware far more effectively than air, which is what makes current accelerator densities viable at all, and they reduce the energy spent shifting heat around inside the building. What they do not do is determine what happens at the end of the chain. The captured heat still has to be rejected to the environment, and that final step is the same evaporative-versus-dry choice as before.

A facility can be fully liquid-cooled and still consume a great deal of water, or almost none, depending on the heat rejection stage. The two questions are separate and are frequently merged.

What we do not know

The industry-wide mix of these designs is not published in enough detail to state a current average with confidence. Operators report facility totals, sometimes, but rarely the design and almost never the seasonal operating profile, so aggregate estimates rest on assumptions about a distribution nobody has measured.

We also do not know how quickly the mix is shifting. Water-stressed regions are pushing new builds toward dry cooling, and rising density is pushing toward liquid, and those two pressures do not resolve to a single direction. Any figure modelled on the deployment mix of a few years ago may already be wrong, and we cannot say by how much or in which direction.

The practical position

If you want to know a facility's water consumption, the questions are: which heat rejection design, in which climate, and is the reported figure withdrawal or consumption. Those three answers explain most of the variance between sites.

If you want to compare two facilities, you need both columns. A site that consumes no water and a site that consumes a great deal may be shedding the same heat, having made opposite choices about where the cost lands.