Data Center Cooling: How It Actually Works
Almost every watt a server uses becomes heat — and AI racks generate more of it than anything the industry has ever built. Here's how data centers get rid of it: from classic air systems to the liquid revolution, and who builds the equipment.
The four ways data centers stay cool
1. Air cooling (CRAC/CRAH)
The classic: chilled air is pushed through raised floors and cold aisles, absorbs server heat, and returns through hot aisles. Hot/cold aisle containment keeps the streams separate. Still the default for standard racks.
2. Free cooling / economisation
When outside air (or evaporatively cooled air) is cold enough, facilities skip mechanical chilling entirely. This is why so many campuses sit in Oregon, Iowa and Quebec — the climate does the work.
3. Liquid cooling (direct-to-chip)
Coolant flows through cold plates bolted onto CPUs and GPUs, carrying heat away far more efficiently than air. The standard for dense AI clusters — full guide here.
4. Immersion cooling
Whole servers are submerged in dielectric (non-conductive) fluid. Maximum density, no fans at all — the frontier of AI-era cooling.
Why AI changed everything
A traditional rack draws a few kilowatts; dense GPU racks draw many tens of kilowatts. Air simply can't move that much heat out fast enough, which is why nearly every new AI campus is engineered for liquid from day one, and why retrofits are sweeping through existing facilities. Cooling is typically the largest chunk of a facility's non-IT energy — the difference between an efficient and wasteful cooling design shows up directly in the PUE number.
Who builds data center cooling
The equipment market is dominated by specialist manufacturers — names you'll see on machinery plates in almost any facility:
- Vertiv — thermal management across CRAC/CRAH units, chillers and liquid systems.
- Schneider Electric — cooling, power and monitoring under one roof (including the APC brand).
- Stulz — precision air-conditioning specialists.
- Munters — evaporative and free-cooling systems.
- CoolIT Systems — direct-to-chip liquid cooling for dense compute.
- LiquidStack & Submer — immersion cooling specialists.
We have no commercial relationship with any vendor listed — this is an editorial overview.
The water question
Evaporative cooling trades electricity for water — which is why "how much water does AI use" became a headline question. We break down the real, sourced numbers here: How much water does ChatGPT use?
Liquid cooling deep-dive →PUE & efficiency →Power →
Frequently asked questions
How are data centers cooled?
Four main ways: chilled-air systems with hot/cold aisle containment (the classic), free cooling using outside air in cold climates, direct-to-chip liquid cooling for dense AI racks, and full immersion in dielectric fluid. Most large facilities combine several.
Why do AI data centers need liquid cooling?
Dense GPU racks draw many times the power of traditional racks, and air physically can't carry heat away fast enough at those densities. Liquid absorbs heat far more efficiently, so new AI campuses are engineered for direct-to-chip liquid from day one.
What companies make data center cooling systems?
The major equipment makers include Vertiv, Schneider Electric, Stulz and Munters for air and evaporative systems, and specialists like CoolIT Systems, LiquidStack and Submer for liquid and immersion cooling.
Does data center cooling use water?
Evaporative designs do — they trade electricity savings for water consumption. Other designs (closed-loop liquid, dry coolers) use little to none. See our sourced breakdown of AI's real water numbers.