PUE (Power Usage Effectiveness), defined by The Green Grid, is Total Facility Power ÷ IT Equipment Power. A PUE of 1.0 means every watt drawn by the facility reaches the IT/GPU load with zero overhead — in practice, cooling, power distribution losses, and UPS conversion losses push real facilities above 1.0.
Power Usage Effectiveness (PUE) is the data center industry's standard metric — defined by The Green Grid in 2007 — for how much of a facility's total power draw actually reaches the IT/GPU equipment versus being consumed by cooling, power distribution losses, and other overhead.
PUE = Total Facility Power ÷ IT Equipment Power. Total Facility Power includes everything: the IT/GPU load itself, plus cooling (CRAH/CRAC units, chillers, cooling distribution units for liquid cooling, pumps, cooling towers), power distribution losses (UPS conversion inefficiency, transformer losses, switchgear), and auxiliary loads (lighting, office space, life safety systems). A PUE of 1.0 is the theoretical minimum — every watt goes to IT with zero overhead, which is physically unreachable since cooling and power distribution are never lossless.
Legacy air-cooled enterprise data centers commonly run PUE 1.8–2.5. Modern enterprise/colocation facilities with economizer-based free cooling typically achieve 1.4–1.6. Hyperscale operators (Google, Microsoft, Meta) report annual fleet-wide PUE averages around 1.1–1.2 using free cooling, high-voltage direct power distribution, and, increasingly, liquid cooling for GPU racks. AI/GPU clusters push cooling load per rack far higher than traditional IT racks, which is why liquid cooling — direct-to-chip or immersion — has become central to keeping PUE low as rack densities climb past 50-100+ kW.
GPU training clusters draw 5-10x the power density per rack of traditional enterprise IT, which pushes cooling systems much harder relative to the IT load. Air cooling becomes both energy-inefficient and physically constrained at these densities, which is why hyperscalers building AI infrastructure are moving to direct-to-chip liquid cooling and immersion cooling — both of which meaningfully lower the cooling-related overhead in the PUE equation compared to CRAH/CRAC air cooling at the same IT load.
Facilities purpose-built for GPU clusters with liquid cooling are increasingly targeting PUE in the 1.1-1.3 range, similar to hyperscale fleet averages, because liquid cooling removes heat far more efficiently than air at high rack densities. A retrofitted legacy air-cooled facility running GPU racks will often show a materially worse PUE simply because its cooling infrastructure was not designed for that density.
No — PUE only measures electrical power overhead. Water-cooled and evaporative cooling systems can achieve excellent PUE while consuming significant water, which is why Water Usage Effectiveness (WUE) is tracked as a separate, complementary metric. See the PUE vs. WUE concept explainer in this studio.
Not for reporting purposes. PUE varies with outdoor temperature (free-cooling hours), IT load level, and time of day, so The Green Grid methodology calls for averaging over a full trailing 12-month period for a representative annual figure. This calculator is useful for quick design-stage estimates, not compliance reporting.
Liquid has a much higher heat capacity than air, so direct-to-chip and immersion cooling can remove the same heat load with far less pumping/fan energy and can operate at higher supply temperatures that enable more free-cooling hours — both directly reduce the cooling kW in the PUE numerator relative to the IT load in the denominator.
Try our AI Data Centers Studio
More calculators, simulators, and guides for this discipline.
AI Data Center Electrical Illustrations (Full Access)
Included with All-AccessA zoomable interactive reader — free preview, then unlock the full set.