Why Data Centres Use Water to Cool the Internet

You open a video, retrieve a photograph or ask a computer to generate an answer. Somewhere, processors switch, memory moves data and electrical equipment does work. Almost all the electricity entering that computing equipment eventually becomes heat.

The information can travel across the world in a moment. The heat cannot be emailed away. It must cross a succession of physical boundaries until it reaches somewhere that can receive it: usually the outside air, sometimes a useful heating system.

Water can make this journey efficient. But water used to transport heat and water lost from the local supply are different quantities. Confusing them makes both alarming claims and reassuring promises harder to judge.

First, find out where the heat goes

A cooling system does not delete heat. It moves it. If a steady-running rack uses 1,000 kilowatt-hours of electricity over an interval, roughly that much thermal energy must ultimately leave the equipment. The cooling machinery can add its own heat along the way. Recovering useful heat changes its destination, not the need to remove it.¹

This is an energy balance, not a claim that every rack needs a particular volume of water. A building can reject the same heat through different equipment, in different weather, with very different water demands. Nor is a kilowatt-hour a measure of computing output: two workloads can use equal electricity and accomplish different things.

“Liquid cooled” is only part of the route

Air-cooled servers release heat into air that must then be cooled or replaced. In direct liquid cooling, a coolant collects heat close to the components. A heat exchanger can pass that heat into another water circuit without mixing the two fluids. The US Department of Energy illustrates systems with separate rack and building-side loops.²

That arrangement answers how heat gets away from the chip. It does not yet answer how heat gets out of the building. The final step might use a cooling tower, a dry cooler, refrigeration equipment or a combination. A closed circuit beside a processor can therefore coexist with evaporation on the other side of a heat exchanger.

The tower deliberately lets some water leave

In a cooling tower, water meets moving air. A portion evaporates, carrying energy away; the remaining water cools and circulates again. The deliberate loss is part of the cooling mechanism, not necessarily evidence of a leak.

Evaporation leaves dissolved minerals behind. Operators therefore discharge some concentrated water, called blowdown, and add replacement or makeup water. Tiny entrained droplets, called drift, can also escape. Treatment and equipment design control these effects.³

“It recirculates” is thus compatible with “it needs a continuing supply”. A roundabout keeps traffic circulating without sealing every exit. The accounting question is what crosses the boundary, not how many times water passes a sensor inside it.

Consumption is not the same as withdrawal

Withdrawing water means taking it from a source. Consumptive use describes the portion no longer available for immediate reuse in the relevant water environment—for example, because it has evaporated. Water molecules have not vanished from the planet. Their availability has changed in a particular place and time.⁴

For a data centre supplied by a utility, the site may measure delivered water rather than withdraw directly from a river. Blowdown may go to treatment and eventually return to a water body. Its destination, quality and timing determine what that return means. It should not automatically be counted as either total consumption or an immediately reusable supply.

Why not just use outside air?

You can. A dry heat exchanger transfers heat to outside air without intentionally evaporating cooling water. But its useful temperature difference depends on the weather and the required coolant temperature. A hot afternoon makes that task harder; other equipment, more airflow or a different design may be needed.

Evaporative systems can cool towards the air’s wet-bulb temperature, which reflects humidity as well as temperature. Dry conditions create more evaporative opportunity; humid conditions narrow it. This is why the local climate belongs in the engineering decision, alongside the permitted equipment temperatures.⁵

Hybrid installations can switch between approaches. A DOE-described supercomputing project used a thermosyphon cooler to reduce reliance on its cooling towers when conditions allowed.⁶ That is an example of changing the final heat-rejection step, not proof that every location can achieve the same result.

Water can be hidden beyond the fence

A cooling design can reduce its direct water demand while changing electricity demand. Generating that electricity may itself use water. The Green Grid’s source-inclusive water metric explicitly recognises the difference between the site boundary and the electricity-supply boundary.⁷

The relationship is not a universal trade of one litre for a fixed number of kilowatt-hours. It depends on the cooling technology and the electricity system serving it. Berkeley Lab’s 2024 US data-centre report separates direct cooling water from electricity-related water in its analysis; its projections remain projections, not measured 2026 outcomes.⁸

So a “water-free cooling” claim needs a follow-up question: free of water consumption where? A precise on-site statement may be valid while leaving power generation, equipment manufacture and other water uses outside its scope.

A better ratio can accompany a bigger total

Power usage effectiveness, or PUE, compares total facility energy with IT equipment energy. It describes infrastructure overhead, not how useful the computation is.⁹ Water usage effectiveness, or WUE, relates site water usage to IT energy, commonly in litres per kilowatt-hour.²

For a fair comparison, the reporting period, water definition and energy boundary must match. “Water usage” should not quietly switch between supplied volume and consumptive use. A lower ratio can reflect a genuine efficiency improvement while growth in activity pushes the absolute total upwards.

The same caution applies to a neat “water per prompt” number. A prompt is not a standard unit of work: model, output length, hardware, location and cooling conditions change the inputs. Without those assumptions, apparent precision can conceal more than it explains.

The meaningful question is local

One litre is a unit of volume, not a complete measure of impact. A research review in npj Clean Water emphasises the importance of water stress when assessing data centres.¹⁰ An annual total cannot by itself describe demand during a dry spell or competition at a particular source.

Useful scrutiny therefore asks for the location and source, seasonal demand, return flows, cooling design and electricity boundary. Reclaimed water can reduce demand for drinking-quality supply, but its use still needs a local accounting: that reclaimed flow may already have another destination. Efficiency improvements are valuable; their significance depends on what the improved system is being asked to do.

The cloud is not made of vapour. It is made of equipment, energy connections and places that must accept heat. Water enters the story because one particularly effective way to remove heat is to let some water leave with it.

A cooling loop can be closed while its water budget remains open.

Sources & further reading

Research checked September 4, 2026. Sources include foundational technical guidance and explicitly dated case studies. All quantities in the interactive examples are hypothetical; this article makes no claim to measured industry-wide 2026 water totals.