Why Data Centers Fight Heat

Every bit of electricity a server uses ends up as heat. A rack of AI servers can draw as much power as dozens of homes, all packed into a cabinet the size of a fridge. Without cooling, chips would throttle or fail within minutes.

So cooling isn't a side detail. It's one of the biggest costs in running the internet, and one of the main limits on how fast AI data centers can grow. This post explains how it works, using one simple number: PUE.

🎯 Quick Answer (30-Second Read)

  • Almost all power used by servers becomes heat that must be removed.
  • PUE (Power Usage Effectiveness) = total facility power ÷ IT equipment power. 1.0 would be perfect.
  • Industry average: about 1.56 (Uptime Institute, 2024 survey). Google's fleet: 1.09 (2024).
  • Air cooling works for normal racks. Liquid cooling is taking over for dense AI racks.
  • Cooling also uses water, which is becoming as important a constraint as power.

The Math: PUE

PUE = Total facility energy ÷ IT equipment energy

The "extra" above 1.0 is overhead: cooling, power conversion losses, lighting.

PUE For every 1 kW of computing... Overhead
2.0 2 kW total 1 kW wasted on overhead
1.56 (industry avg.) 1.56 kW total 0.56 kW
1.09 (Google fleet) 1.09 kW total 0.09 kW
xychart-beta title "Overhead energy per 1 kW of IT load" x-axis ["PUE 2.0", "PUE 1.56 (avg)", "PUE 1.09 (Google)"] y-axis "kW of overhead" 0 --> 1.1 bar [1.0, 0.56, 0.09]

Going from 1.56 to 1.09 cuts overhead from 0.56 to 0.09 kW per kW, about 84% less overhead, which is the comparison Google itself makes.

Worked example: a 100 MW AI campus (IT load).

  • At PUE 1.56 → 156 MW total, 56 MW of overhead.
  • At PUE 1.09 → 109 MW total, 9 MW of overhead.

The difference, 47 MW, is roughly the power of a small town, used purely to keep machines cool.

Watch out: PUE numbers aren't always comparable. Google reports a fleet-wide, year-round average including all overhead. Uptime's survey asks operators for the PUE of their largest site. Some companies quote their best facility on its best day.

How Air Cooling Works

flowchart LR CH["Chiller / cooling tower
removes heat outside"] -->|"cold water"| CRAH["Air handlers
(CRAH units)"] CRAH -->|"cold air"| COLD["❄️ Cold aisle
server fronts face here"] COLD --> SRV["Server racks
fans pull air through"] SRV --> HOT["🔥 Hot aisle
exhaust air"] HOT -->|"warm air returns"| CRAH CRAH -->|"warm water"| CH

The classic layout is hot aisle / cold aisle: racks face each other so cold air comes in the front and hot air goes out the back into a separate aisle. Keeping hot and cold air from mixing is the cheapest efficiency win there is.

Ways to cut cooling energy:

  • Run warmer. Modern servers handle warmer inlet air than older guidance suggested.
  • Free cooling: use cold outside air or water when the climate allows, instead of running chillers.
  • Evaporative cooling: evaporating water removes heat cheaply, but consumes water.
  • Smarter control: Google has described using machine learning (from DeepMind) to tune cooling systems.

Why AI Changes Everything: Liquid Cooling

AI training chips draw far more power per chip than typical servers, and they're packed densely. Air can't carry heat away fast enough at those densities. Liquids hold much more heat than air for the same volume, so the industry is moving to:

Direct-to-chip
Cold plates sit on the processors and coolant flows through them. Now common for high-end AI servers.
Immersion
Entire servers sit in a tank of non-conductive fluid. Very efficient but harder to maintain.
Rear-door heat exchangers
Water-cooled doors on racks catch heat as air leaves. A bridge between air and liquid.

The Water Question

Evaporative cooling saves electricity but uses water. Operators track WUE (Water Usage Effectiveness, litres per kWh of IT energy) alongside PUE. In dry or water-stressed regions, including many Indian cities, there's a real trade-off: lower PUE using water, or higher PUE using closed-loop systems that consume little water.

Frequently Asked Questions

What is a good PUE for a data center?

The industry average has stayed around 1.5 to 1.6 for several years, at 1.56 in Uptime Institute's 2024 survey. Large modern facilities often report 1.1 to 1.2, and Google reported a fleet-wide average of 1.09 for 2024. Lower is better, with 1.0 the theoretical ideal.

Why do AI data centers need liquid cooling?

AI accelerators use much more power per chip and are packed densely, producing more heat per rack than air can remove efficiently. Liquid carries heat far better than air, so direct-to-chip and immersion cooling let operators run dense AI racks safely and efficiently.

How much of a data center's energy goes to cooling?

It depends on the facility. At a PUE of 1.56, about 36% of total energy is overhead, much of it cooling. At a PUE of 1.09, overhead falls to about 8%. Efficient design, warmer operating temperatures and free cooling make most of that difference.

Conclusion

Data center cooling is a physics problem with a simple scorecard: PUE. Every watt of computing becomes heat, and every watt spent removing it is overhead. The best operators push PUE close to 1.0 with careful airflow, free cooling and smart controls, while AI's dense racks are forcing the industry from air to liquid cooling, with water use as the next big trade-off.

Sources

Last verified: 14 October 2026.

Related reads: Why Free Apps Cost Billions to Run · Kubernetes Explained · Claude API Pricing

I explain the tech behind everyday products. Follow on LinkedIn or X, or invite me to speak.