Data centre cooling is the set of systems that remove the heat servers generate so equipment stays inside safe operating limits. Almost every watt of power a server draws turns into heat. Cooling moves that heat out of the hall and rejects it outside, using air, water, refrigerant, or a fluid in direct contact with the chip. It is one of the largest costs an operator carries after the power bill itself.
This guide explains how cooling works, the main methods in use, the two efficiency numbers buyers and planners track, and why AI workloads are forcing a shift from air to liquid. It is written to be read in order or jumped through by question. Data Centre Axis is neutral on vendors and takes no position in any transaction. We track the public record, we do not sell cooling kit.
Why do data centres need cooling at all?
Servers, storage and network gear convert electricity into computation and heat in roughly equal measure. A rack drawing 10 kW of power rejects close to 10 kW of heat into the room. Left unmanaged, that heat pushes hardware past its thermal limit, which triggers throttling, shortens component life, and eventually causes failure.
Cooling keeps inlet air or coolant within the range the hardware is rated for. Most equipment runs comfortably with supply air in the low-to-mid 20s Celsius, though guidance has widened over the years to allow warmer set points and save energy. The job is constant. A hall never stops producing heat while it is powered, so the cooling system runs every hour of every day.
How does data centre cooling work?
Cooling follows a chain. Heat leaves the chip, moves into the surrounding air or a liquid, travels to a heat-rejection stage, and is finally dumped outside the building.
In a conventional air-cooled hall, server fans push hot air into a contained aisle. Computer room air handlers draw that air across cooling coils, then return cold air to the front of the racks. The coils carry chilled water or refrigerant that takes the heat to a chiller or a cooling tower on the roof or in the yard. From there it is rejected to the outside air or to evaporating water.
The same chain applies to liquid systems, but the first step changes. Instead of air carrying heat off the components, a fluid does, and it does so far more efficiently because liquids hold and move much more heat per litre than air does per cubic metre.
What are the main data centre cooling methods?
There are four approaches in common use. Most real sites run a mix.
Air cooling. Cold air is supplied to the rack front and hot air is captured at the rear, usually with hot-aisle or cold-aisle containment to stop the two mixing. It is simple, well understood, and cheap to deploy. It struggles above roughly 20 to 30 kW per rack, which is where AI hardware starts.
Rear-door heat exchangers. A water-cooled coil bolts onto the back of the rack and captures heat as air passes through it. This raises the density a single rack can handle while keeping much of the existing air design in place. It is a common bridge between pure air and full liquid.
Direct-to-chip liquid cooling. Cold plates sit directly on the hottest components, the CPUs and GPUs, and a coolant loop carries heat away at the source. This is the leading method for high-density AI racks because it deals with the heat where it is most concentrated. Air still handles the rest of the rack.
Immersion cooling. Whole servers sit in a bath of non-conductive fluid that absorbs heat directly. It removes server fans entirely and handles very high density, but it changes hardware handling, maintenance, and hall design significantly, so adoption is more specialist.
What is PUE in a data centre?
Power Usage Effectiveness, or PUE, is the ratio of the total energy a facility draws to the energy that reaches the IT equipment. A PUE of 1.0 would mean every watt went to computing and none to cooling, lighting or losses. Real numbers sit above that.
Cooling is usually the largest non-IT load, so it is the single biggest lever on PUE. The Uptime Institute's annual survey has put the global fleet average around 1.56 for several years, while the most efficient hyperscale sites report figures near 1.1. A lower PUE means more of the power bill goes to useful work and less to keeping the room cool, which is why efficient cooling design is a commercial issue, not only an engineering one.
What is WUE and how much water do data centres use?
Water Usage Effectiveness, or WUE, measures litres of water consumed per kilowatt-hour of IT energy. It matters because many cooling systems reject heat by evaporating water in cooling towers, which trades electricity for water.
The trade-off is the heart of the planning debate. Evaporative cooling cuts power use and improves PUE, but it consumes water, and in dry regions that water is contested. Closed-loop and air-assisted designs use little or no water but draw more power. In Australia, where several large sites sit in water-stressed catchments, this is a live planning constraint that shows up in environmental referrals and state assessments. The industry-average WUE is often cited at around 1.8 litres per kilowatt-hour, while the most water-efficient operators report well below half a litre. Buyers underwriting a site, and developers seeking approval, both have to model cooling water as a real risk rather than a footnote.
Why is AI changing data centre cooling?
AI training and inference run on dense clusters of GPUs that draw far more power per rack than traditional servers. A rack of conventional servers might pull 5 to 15 kW. An AI rack can pull 50 to 150 kW or more, which is well beyond what air alone can carry.
That density is forcing the move to liquid. Direct-to-chip cooling and, in some cases, immersion are becoming the default for AI halls rather than an exotic option. The shift has knock-on effects across the site. Power density per square metre rises, the heat-rejection system has to handle a higher and more concentrated load, and water and grid questions sharpen because the same site now demands more of both. For anyone assessing whether a site is genuinely AI-ready, the cooling method is one of the first specifications to check.
How much does data centre cooling cost?
Cooling sits among the largest operating costs after the electricity the IT equipment itself consumes, and it is a major share of construction cost too. The exact figure depends on climate, design and density.
There are two cost stories. Capital cost covers the chillers, towers, pipework, cold plates or immersion tanks installed up front. Operating cost is dominated by the electricity the cooling system draws, which is why PUE maps so directly to running cost. A site with a PUE of 1.5 spends roughly half as much again on facility power as it does on IT power, and most of that overhead is cooling. Cooling typically accounts for around 30 to 40 percent of a traditional facility's total electricity use. Liquid systems often raise capital cost but lower the energy share, with direct-to-chip designs cutting cooling energy by roughly 40 to 60 percent, which can improve the lifetime numbers for dense workloads.
Frequently asked questions
What is the difference between air cooling and liquid cooling?
Air cooling moves heat using cold air supplied to the front of the rack. Liquid cooling uses a fluid, either in cold plates on the chip or in an immersion bath, to carry heat away directly. Liquid handles far higher rack densities and is becoming standard for AI hardware.
What is a good PUE for a data centre?
The most efficient hyperscale sites report figures near 1.1, while the wider fleet average sits around 1.56 and has barely moved in years. Lower is better, since it means less power spent on cooling and overhead.
Do all data centres use water for cooling?
No. Many use evaporative cooling, which consumes water to cut power use, but closed-loop and air-cooled designs use little or no water at a higher energy cost. The choice depends on climate, power price and local water availability.
Is liquid cooling necessary for AI?
For high-density AI racks, usually yes. Once a rack draws beyond roughly 30 to 50 kW, air struggles to remove the heat, so direct-to-chip or immersion cooling becomes the practical option.
What is the most efficient data centre cooling method?
It depends on the workload and climate. For high density, direct-to-chip and immersion cooling are the most efficient at moving heat. For lower density in cool climates, free-air and air-assisted designs can be very efficient with little water use.