NodalZen

Blog · Explainer

PUE and cooling without the jargon

· 6 min read · pue, cooling, efficiency

What power usage effectiveness (PUE) means, how air and liquid cooling differ, and why the number matters for AI data centers.

Every data center pays for two kinds of electricity. The first kind runs the computers. The second kind runs everything that keeps those computers alive: fans, pumps, chillers, lights, and the losses in the power equipment. A single number tells you how big that second bill is. It is called PUE.

What PUE means

PUE stands for power usage effectiveness. It is a simple ratio:

PUE = all the electricity the building uses ÷ the electricity the computers use

If the computers use 100 units and the whole building uses 150, the PUE is 1.5. That means for every unit of useful computing, the building spends another half unit on overhead.

A perfect score would be 1.0, where every bit of power goes to the computers and nothing to overhead. No real building reaches that, because cooling and power conversion always cost something.

An everyday comparison: think of a delivery van. The parcels are the useful load. The van's own weight, fuel and driver are overhead. PUE asks how much "van" you are paying for to move each parcel.

Where the industry stands

The Uptime Institute, an industry research group, surveys data center operators every year. In its 2026 survey, the average PUE across facilities was 1.52. When the results are weighted by size, so big sites count for more, the figure drops to 1.36, because larger and newer sites tend to be more efficient (Uptime Institute, August 2026). The previous year's survey reported an average of 1.54 (cited by Google, data center efficiency page), and Uptime noted that average PUE had barely changed for six years in a row (Uptime Institute, July 2025).

Looking ahead, the International Energy Agency's base case expects the global average PUE to fall from 1.41 to 1.29 by 2030, saving about 90 terawatt-hours of electricity, according to a summary of its April 2025 Energy and AI report (International Institute of Refrigeration, May 2025). The IEA and Uptime figures are calculated differently, so they are not directly comparable.

Why cooling is the big lever

Chips turn almost all the electricity they use into heat. If that heat is not removed, they slow down or fail. So cooling is usually the largest part of the overhead, and the place where a better design saves the most.

Air cooling

Traditional data centers cool with air. Cold air is pushed through the front of the racks, picks up heat from the chips, and hot air leaves out the back. Machines called chillers or cooling towers then remove the heat from that air or from water loops connected to it.

Air works well when each rack uses a modest amount of power. But air is a poor carrier of heat. As racks get hotter you need more fans moving more air, and at some point it stops being practical. Uptime Institute analysts wrote in 2025 that very high rack densities would be difficult, expensive or even impossible with air alone, and that AI hardware roadmaps promise 200 kilowatts per rack or more within a few years (Uptime Institute Journal, April 2025).

Liquid cooling

Liquid can carry far more heat than the same volume of air. That is why car engines use a radiator full of coolant instead of just a fan.

The most common form in data centers is direct-to-chip cooling: a thin metal plate, called a cold plate, sits on top of each hot chip, and liquid flows through it, carrying the heat away to the building's cooling system. Another approach, immersion cooling, dunks whole servers in a special fluid that does not conduct electricity.

Liquid cooling is still the minority. In Uptime's 2024 cooling survey, 22% of respondents said their organization used some direct liquid cooling, and 61% said they did not yet but would consider it (Uptime Institute, May 2024).

Why PUE matters to you

PUE turns into money and grid capacity. Suppose your chips need 10 megawatts. At a PUE of 1.5 the building needs 15 megawatts from the grid. At 1.2 it needs 12. That 3 megawatt gap means a smaller grid connection, smaller transformers and backup systems, and a lower power bill every hour of every year.

What PUE does not tell you

PUE has limits. It says nothing about how useful the computing is, and it does not count water. Evaporating water in cooling towers saves electricity, so a design that stops using water can see its PUE rise. Microsoft said exactly this when it described its closed-loop, no-evaporation cooling design in 2024 (Microsoft, December 2024). Water use is worth checking separately.

Try it in NodalZen

In Size, switch your cluster between air, direct-to-chip and immersion cooling and watch how its PUE and the total power the building needs change. The Learn lessons include a beginner walkthrough of cooling and PUE.

Sources

Learn the basics Open the planner