What PUE Actually Measures
Power Usage Effectiveness divides total facility power by the power consumed by IT equipment. A perfect score is 1.0 — physically impossible, because some overhead always exists. Early data centres, with their underfloor air mazes and oversized chillers, routinely ran above 2.0. The hyperscale era drove that number down hard: Google, Meta and Microsoft have published portfolio averages in the 1.1–1.2 range, achieved through hot-aisle containment, free-air cooling in favourable climates, and rigorous elimination of wasted airflow.
That progress was real. But PUE measures a ratio, not an absolute. A facility drawing 100 MW of IT power at 1.15 PUE still dumps 15 MW into cooling and ancillaries — roughly the output of a small power station, doing nothing useful.
Where AI Pressure Lands
Traditional server racks drew 5–10 kW. A dense GPU cluster rack pulls 40–130 kW. Pack enough of those into a building and the cooling system — designed for a fraction of that heat flux — becomes the binding constraint, not the servers.
Air cooling hits a wall around 40–50 kW per rack. Above that, forced air simply cannot carry heat away fast enough without heroic — and energy-intensive — measures. Paradoxically, this is where AI could improve PUE rather than worsen it: direct-to-chip liquid cooling carries heat far more efficiently than air, reducing the energy spent moving that heat out of the building. A well-engineered liquid-cooled deployment can push cooling overhead lower than an equivalent air-cooled build, even at higher absolute power.
Warm-water loops go further still. When cooling fluid returns from the chips at 40–50 °C rather than the frigid temperatures chillers traditionally target, that warmth is recoverable — district heating networks in Europe have already begun absorbing waste heat from data centres. The energy is used twice, which no PUE calculation captures.
The energy is used twice, which no PUE calculation captures.
The Metric's Real Limits
PUE says nothing about what the IT power actually does — a badly configured cluster burning energy on idle GPUs scores identically to a perfectly utilised one. It also ignores water consumption, embodied carbon in hardware, and the carbon intensity of the electricity supply.
For operators, PUE remains a useful, fast-readable dial on cooling efficiency. For investors and policymakers, it should be one instrument in a broader panel — alongside utilisation rates, water usage effectiveness (WUE), and grid carbon intensity. AI's density shock is, at minimum, forcing the industry to revisit every assumption about how heat leaves a building. That is not a crisis; it is an engineering problem with real solutions already in deployment.