Cooling a data center: free cooling, adiabatic, liquid and immersion
Cooling decides two things that matter to your bill: how much electricity the data center spends removing heat, and how dense it can let you pack. Here, without jargon, are the main cooling families and what they change for you.
Estimate my cost →Why cooling weighs so much #
A server turns almost all the electricity it draws into heat, which must be removed continuously. The energy spent on that cooling does no computing: it is extra electricity, passed on to your bill. Cooling is also what sets the maximum density a site can host.
PUE, plainly #
PUE (Power Usage Effectiveness) compares the data center's total energy to the energy actually useful to the servers. A PUE of 1.2 means 20% extra for the facility; 1.5 to 2 is common on older sites. The closer to 1, the better. Compare it with care: it depends on climate, load and measurement method.
Classic mechanical cooling #
Compressor-based chillers produce cold air or chilled water. Simple and proven, but power-hungry, hence a high PUE. Still widespread, especially on sites designed fifteen years ago.
Free cooling #
The idea: use cold outside air (or water) instead of compressors. Indirect free cooling has outside air cool a water loop through an exchanger; direct free cooling brings outside air into the room. The PUE gain is clear, but the result depends on the local climate.
Adiabatic cooling #
Air is cooled by evaporating water on a wet medium: evaporation absorbs the heat. Very effective on dry air, less so on very humid air. Its trade-off is water use, hence the recurring water-versus-electricity arbitration.
Air, water, or straight onto the chip #
Air suffices up to around 20 to 30 kW per rack. Beyond that comes liquid cooling: rear-door heat exchangers, then direct-to-chip close to the components. For the extreme densities of AI, immersion submerges servers in a dielectric fluid.
The real trade-off: water or energy #
Free cooling and adiabatic save electricity but can consume water. Immersion can cut both, at the cost of specific infrastructure. The right choice is local: climate, water availability, regional water stress.
What it changes for you #
Three direct consequences: the density the site accepts (kW per rack), the PUE and thus the energy passed on to your bill, and resilience (is cooling redundant, N+1 or N+2?). The right questions to ask: real PUE, maximum density per rack, cooling type, redundancy, and water use.
FAQ #
The closer to 1, the better: a PUE of 1.2 is excellent, 1.5 to 2 is still common on older sites. Compare with care, because PUE depends on climate, load and measurement method.
Beyond around 20 to 30 kW per rack, air no longer suffices: direct-to-chip or immersion become necessary. Most GPU clusters are in this case.
It depends on the location. In a water-stressed area, the water use of adiabatic cooling becomes a real issue; elsewhere, it can be a good trade-off against electricity.
Written on 1 September 2026.
From reading to comparing: relevant data centers
Estimate my cost → Compare data centers