The water question behind every data center

Picture a mayor in southern Europe at the end of a hot summer.

Reservoir levels are being watched. Farmers are concerned about irrigation. Residents have been asked to conserve water. Then a proposal for a new data center arrives.

The developer may want to talk about investment, jobs, artificial intelligence, or digital sovereignty. The mayor has a more immediate question: how much water will it use?

At first, this sounds like a request for one number. It is not.

A 2025 study from Lawrence Berkeley National Laboratory found that the water used to perform a computing workload can vary by more than 10,000 times. Server efficiency, electricity supply, utilization, cooling technology, climate, inactive equipment, and hardware age can all change the result. The researchers concluded that there is no single recipe for minimizing water use.

That variation is large enough to make sweeping claims about data centers and water unreliable. It also reveals the more important story: a data center’s water footprint is not a fixed property of computing. It emerges from a chain of physical and operational choices.

Heat turns into a water question

Servers do not need water to process information. They need electricity. The water question begins with what happens to that electricity afterward.

Processors turn electrical power into computation and heat. Thousands of servers operating continuously, particularly dense GPU systems built for AI, create a large thermal load. The heat has to leave the processors, the servers, the room, and eventually the facility. The full route is explained in our article on data-center cooling systems.

One widely used method moves heat from the IT equipment into room air, then into chilled water, through a condenser-water circuit, and finally to a cooling tower. At the tower, some water evaporates into the atmosphere, carrying the heat with it.

The U.S. Department of Energy’s cooling guide describes this evaporative stage as the final point of heat removal in a typical cooling-tower system. Some circulating water must also be discharged to control mineral concentrations. Fresh makeup water replaces what is lost through evaporation and discharge.

There is nothing irrational about this design. Evaporation can reject large amounts of heat efficiently. Its efficiency has a physical consequence: some of the water leaves the system.

This is the origin of the familiar image of the thirsty data center. The image captures a real issue, but often turns one cooling architecture into an unavoidable law of computing.

Data centers can reject heat through outside air, dry coolers, mechanical refrigeration, evaporative systems, circulating liquids, or combinations that change with temperature and load. Water plays a different role in each.

A cooling tower consumes water through evaporation. A closed circuit can circulate the same working fluid repeatedly. Liquid may carry heat away from processors before that heat is ultimately rejected through air, evaporation, or another system.

That is why the vocabulary can mislead. “Liquid-cooled” does not tell us how much freshwater a facility consumes. “Air-cooled” does not prove that no liquid exists anywhere in the thermal system. “Closed loop” describes circulation within a boundary, not the complete water footprint. Closed-loop cooling needs its own boundary explained.

For the mayor, the practical question is not whether water appears somewhere in the engineering diagram. It is how much new water enters the facility, what happens to it, and what must replace it during normal operation.

A water number is only as useful as its boundary

Even a precise answer can conceal part of the system.

A data center can use water directly at its site and indirectly through the electricity it consumes. Some forms of power generation withdraw or consume water for cooling and other processes. Berkeley Lab ranks the water intensity of the electricity supply among the main factors determining the water used by a computing workload.

A facility with low on-site water use may therefore retain an upstream water footprint. That does not make direct water consumption unimportant. It makes the boundary of the claim important.

If a data center says it uses no water, does that mean no freshwater enters the cooling system? No water is consumed at the site? No water is used by the electricity supply? Or no water is involved anywhere in manufacturing, construction, and operation? Those are different claims.

Water Usage Effectiveness, or WUE, makes one part of the picture easier to compare by relating annual site water use to the energy consumed by IT equipment. It is a companion to Power Usage Effectiveness, not a replacement for it.

The pairing matters because energy and water efficiency can pull in different directions. Evaporative cooling may reduce the electricity needed to reject heat while consuming more water. A dry system may reduce direct water demand but use more electricity in hot conditions. A hybrid system may switch between the two.

A good result on one metric does not settle the question raised by the other. PUE needs its limits visible, while WUE needs its volume, source, and location beside it.

Measurement language matters too. Water withdrawal describes water taken from a source. Water consumption generally refers to the portion that is not returned to the same local system in a usable way. A facility that withdraws water and returns most of it creates a different condition from one that loses water through evaporation, even when the initial intake looks similar.

A liter has a location

The mayor’s question is local because water itself is local.

Electricity can move through interconnected grids. Data can cross continents through fiber. Water remains tied closely to watersheds, reservoirs, aquifers, treatment systems, weather, and competing users.

The European Environment Agency reports that water scarcity affects large parts of Europe seasonally, with the greatest pressure in southern Europe and stressed river basins elsewhere. A liter consumed in a region with abundant supply does not create the same consequence as a liter consumed during summer stress in an overused basin.

AI makes the location question more urgent, though not always in the way popular headlines suggest. GPUs can concentrate far more computation and heat into a rack than conventional systems. That increases the thermal challenge, but it does not create one universal AI water footprint. Hardware efficiency, utilization, cooling, electricity, and location still determine the result. The gap between memorable per-prompt estimates is examined in our AI water-use article.

Europe is also demanding more visibility. The EU already operates a reporting database for energy-performance and water-footprint indicators from covered data centers. In 2026, the European Commission consulted on a common rating scheme intended to make data-center performance easier to compare. Its data-center energy-performance page tracks that work.

Regulation will not remove every ambiguity. Metrics still depend on definitions, boundaries, reporting quality, climate, and operating conditions. But the direction is clear: water can no longer remain hidden inside a general claim that a data center is efficient.

Design from the place outward

For decades, the language of “the cloud” made computing sound almost placeless. The infrastructure underneath never was.

A data center occupies land. It connects to a particular electrical network. Fiber must reach it. Cooling equipment rejects heat into a real climate. Roads, foundations, security, permits, and operating teams have to exist around it. Water belongs in that same physical description.

Policloud develops and deploys physical, modular data-center infrastructure for defined sites. That approach makes water part of the design question before the infrastructure is committed.

A site facing seasonal scarcity should bring that constraint into the cooling decision from the beginning. A dense workload should shape the thermal architecture before someone decides how many racks will fit. A location with a practical use for captured heat presents a different set of possibilities from one where all heat must be rejected outdoors.

The same principle applies to power, network access, land, noise, and regulation. Those conditions come together in data-center site selection.

The relevant question is not whether one cooling technology is always better than another. It is whether the selected system fits the workload and the place without creating a resource problem that someone else must solve later.

The mayor does not need to know whether data centers, as a global category, use a large or small amount of water. She needs to know what the proposed facility will require from her town and its watershed.

How much water enters the site? How much is consumed? How does the answer change during the hottest weeks of the year? What source supplies it? What happens if that source becomes constrained? Which parts of the footprint sit outside the property boundary?

A trustworthy data-center proposal should be able to answer those questions plainly.

If it cannot, the water problem has not been solved. It has only been hidden inside the word “efficient.”