Datacenter cooling is a choice between water, power, and location

Why annual water use is the wrong number for judging datacenter cooling, and why location is the decisive choice.

Robert Lukkenaer, Head of Datacenter at Project Enki | July 2026

Since 16 July 2026, the Netherlands has had an official water shortage. That makes the question of how much water datacenters use a fair one. But cooling is not a simple choice between water or no water. It is a trade-off between air, water, electricity, and location.

The confusion starts with a more fundamental question: why does a datacenter produce so much heat in the first place?

Almost all the electricity a datacenter uses ends up as heat.

This is where most of the confusion begins. You put power in and you expect compute to come out. So where does that energy go?

Compute is not a product that leaves the building. The only thing that physically exits a datacenter is a signal down a fibre-optic cable. That signal carries a negligible amount of energy. A chip computes by switching billions of transistors, and that switching costs energy which is released as heat. This is conservation of energy: it never produces more heat than the energy going in.

In practice this means a 30 MW datacenter is a 30 MW heater. There is no clever technique that changes this. At most, through smart choices, we can allocate a larger share of the power to compute or apply a more efficient cooling technique.

More efficient chips deliver more computation per watt, but at equal installed power the amount of heat to be removed stays the same. At the same time, power density is rising: where a rack used 5 to 10 kW ten years ago, AI racks now run to tens or more than a hundred kilowatts. The required air volumes, pressure differentials, and fan powers become steadily less attractive as a result. That is why the industry is shifting to liquid cooling.

The advantage of liquid cooling is that it allows cooling at a higher temperature. Where air-cooled datacenters are usually designed around a recommended IT inlet temperature of up to 27 °C, suitable warm-water-cooled systems can accept facility water of 40 to 45 °C. This temperature increase inside the datacenter is the biggest development of recent years and delivers a considerable reduction in cooling overhead.

Why a datacenter must not get too hot

The second question I get: what actually happens if you simply do not cool?

If you do not cool, things go wrong within minutes. A chip has a maximum core temperature, usually somewhere around 90 to 100 degrees. When it hits that limit, it lowers its own clock speed to protect itself. You then have hardware worth hundreds of thousands or millions of euros running at half power. If the temperature keeps rising, the equipment shuts itself down.

There is also a vicious circle. Chips that are too hot leak current without doing any computation for it. Performance degrades.

For many electronic components, the ageing rate rises sharply at higher temperatures; as a rough engineering rule of thumb, a doubling of certain degradation processes per 10 °C is sometimes used.

How much power does cooling actually cost?

The industry expresses total overhead as PUE: total energy use divided by the energy use of the IT. At a PUE of 1.54, a datacenter uses roughly 0.54 MW of cooling, electrical losses, and other facility systems for every 1 MW of IT. Modern hyperscale sites come much closer to 1.1.

That difference is enormous. In older or less efficiently designed datacenters, cooling can make up a substantial share of total energy use. In modern facilities that share is usually much lower. Cooling is therefore the biggest lever an operator can pull (once the datacenter is built) without replacing the chips themselves.

An interesting detail: the standardised PUE calculation leaves heat reuse entirely out of account, which means PUE can by definition never fall below 1.00. An additional KPI, the ERE (Energy Reuse Effectiveness) from The Green Grid, does count reused energy and can even reach 0 at full utilisation. That illustrates how large the potential gain is when location and heat offtaker are well matched.

Internal loop and external loop

Almost every datacenter has a closed internal loop that collects the heat at the IT and an external loop that releases this heat to the surroundings. The two loops are usually separated from each other by a heat exchanger.

Ways to get rid of heat

The cooling method a datacenter uses depends on several factors: the outside temperature and the humidity. Those factors determine when you can cool against the outside air and when you have to add water or compressors.

Free cooling.
In free cooling, the water from the external loop flows through large air-cooled heat exchangers. Fans carry the heat away. In the Netherlands this works for a large part of the year, but the fans use power and produce noise.

Adiabatic cooling.
As it gets warmer outside, a small amount of water is misted into the incoming airflow passing over that same radiator. The water evaporates, the air cools, and you extend the period of free cooling. This is the cooling method the much-discussed water use comes from (depending on the system, the weather conditions, and the share of heat removed by evaporation, water use can rise to around 2.2 m³/MWh. For the scenario calculation below, a figure of 1.0 m³/MWh has been used). This method works less well, however, when humidity is high and the water cannot evaporate as easily.

Mechanical cooling.
During the hottest and most humid hours, mechanical cooling is needed. Compressors work on the same principle as a refrigerator but require considerably more electricity. If the full IT load has to be capable of being mechanically cooled, the power connection must also be sized for it, a serious constraint in a country with grid congestion.

Ground cooling.
In the Netherlands this happens more than people think, but still far too little. Aquifer thermal energy storage, or ATES, stores cold and heat in groundwater at a depth of 70 to 180 metres. ATES can be highly efficient but requires suitable soil conditions, permits, and a balanced annual heat budget.

Heat discharge to surface water.
Heat discharge to surface water is bound by permits and discharge conditions, including limits on the discharge temperature, the temperature rise, and the size of the mixing zone. Depending on the permit, temperature and discharge are monitored and reported. This does not remove the ecological question.

If we consider the standard methods for datacenter cooling, we quickly arrive at the table below for the Dutch climate (for both air- and liquid-cooled systems).

Indicative annual cooling operating costs per 1 MW of IT heat to be removed continuously.

With warm-water liquid cooling, the Netherlands can rely on dry free cooling almost all year round. Water and chiller use are thereby all but eliminated.

You can switch chillers on earlier to save water, certainly at moments with very low or negative electricity prices. At the price of €0.07/kWh used here, however, adiabatic cooling remains cheaper. Moreover, a datacenter usually does not pay the bare market price alone: transport costs, taxes, maintenance, and wear remain.

What kind of water is it, actually?

Here lies a misunderstanding that works both ways.

Part of the water use involves drinking water, but that is not technically necessary. Adiabatic systems do require water of controlled quality, because scale, salts, and biological growth can damage the installation. Rainwater, industrial water, or treated wastewater can therefore also be usable. Drinking water is often chosen because the pipe is already there and the quality is guaranteed.

Does the earth warm up because of datacenters?

I get this question often, and the answer has two parts.

Globally, the direct waste heat of datacenters is not a relevant contribution to global warming. The climate impact lies mainly in the emissions of the electricity used (scope 2) and in the production of buildings, installations, and IT equipment (scope 3). Locally, heat release can have an effect, though it is hard to distinguish from paving, construction, and loss of vegetation around large campuses.

So why don't we build them in Norway?

Norway is attractive for some workloads but not for everything. Interactive applications place high demands on latency and connectivity. Grid capacity is scarce in Scandinavia too, while large datacenters need technicians, security, and suppliers nearby. Finally, data sovereignty and jurisdiction play a role. Cold air alone therefore does not determine a suitable location.

Alternative: cooling with seawater

Seawater cooling is not a distant prospect. Google has used water from the Gulf of Finland for its datacenter in Hamina since 2011, via the intake tunnels of a former paper mill. In Stockholm, surface water is likewise used to cool datacenters efficiently.

The principle is simple: instead of removing heat with large volumes of air, evaporating drinking water, or continuously running compressors, the heat is transferred to cold outside water through a heat exchanger.

The execution is less simple. Seawater is corrosive and contains salt, sediment, and organisms. That is why corrosion-resistant heat exchangers, suitable piping, filtration, and measures against biological growth are needed. The intake and return of the water must also be carefully designed. Temperature rise, mixing, and local ecological effects are limited and monitored.

Offshore, conditions are in many cases more favourable than in a river, harbour, or shallow inland water. There is more water volume, current, and natural mixing, while no drinking water has to be evaporated. For a liquid-cooled datacenter with a high return temperature, this means cooling can be done with heat exchangers and pumps almost all year round. Fans and mechanical cooling are then needed only for limited residual heat or exceptional conditions.

This does not automatically make seawater cooling simple, cheap, or free of ecological impact. Intake screens, biofouling, fish larvae, thermal plumes, material selection, and permitting remain important design questions. But at a suitable location, seawater can offer a structural alternative to the choice between drinking-water use and extra load on the electricity grid.

Conclusion

A datacenter delivers compute, but almost all the electricity used ends up as heat that has to be removed continuously.

The debate about datacenter water use is legitimate. Water is scarce in specific places and at specific moments, and it is precisely during warm and dry periods that cooling demand can be highest. But looking only at the annual number of litres gives an incomplete picture. Saving water by cooling more mechanically can increase the load on an already full electricity grid. Conversely, a low PUE can be achieved by evaporating water at the moment when it is socially least desirable.

A responsible assessment must therefore go beyond a single metric. Water use, electricity consumption, grid capacity, noise, local heat load, opportunities for heat reuse, and ecological effects must be considered together. The timing and location of the use are decisive here too.

The most important design choice is ultimately the location. A datacenter in an unfavourable place has to choose between evaporating water, using extra electricity, or discharging low-grade heat to surroundings that can do nothing with it. At a suitable location, warm-water liquid cooling and direct heat discharge to a large natural heat buffer can strongly reduce these dependencies.

The future of efficient datacenter cooling therefore does not lie in one universal technique, but in combining higher cooling temperatures with a location that can responsibly absorb the heat produced. It is not only the efficiency of the datacenter that must be optimised, but the whole system around it.

Robert Lukkenaer is Head of Datacenters at Project Enki, a Dutch company building AI datacenters at sea, directly alongside offshore wind farms, cooled by seawater. Project Enki is working on the first European offshore AI infrastructure.

projectenki.com | contact@projectenki.com