Power usage effectiveness (PUE): How data centers measure the cost of cooling

Power usage effectiveness, or PUE, is a ratio that shows how much of a data center’s electricity actually reaches the computers. The rest goes to overhead such as cooling and power delivery. A PUE of 1.0 would mean no overhead at all, and every point above 1.0 is energy spent on something other than computing.

It is the number operators quote when they say a facility is efficient, and it sits behind the current rush into cooling equipment.

One electricity bill, split into two parts

Picture a restaurant kitchen. The cooks do the useful work, but the building also needs ventilation, lighting and a walk-in fridge. If the whole building uses 150 units of power and the cooks’ stoves and ovens use 100, the ratio is 1.5. The extra 50 is the cost of keeping the kitchen running.

A data center works the same way. The servers, storage and network gear are the cooks. Cooling, lighting and power conversion are the building around them. PUE puts one number on that split.

The formula and what goes into it

The formula is simple: PUE equals total facility energy divided by IT equipment energy. According to Vertiv’s explainer, the top figure is measured at the data center’s meter and includes IT equipment, cooling, lighting and power delivery. The bottom figure is the energy delivered to compute, storage and networking gear, measured after power conversion.

Everything that is not IT load counts as overhead. In most facilities cooling is a major part of it, which is why a better cooling system can move the number directly.

The ratio is easy to write down and harder to measure well. Vertiv notes that shared utility meters may force operators to estimate and subtract loads that do not belong to the data center. Teams also differ on whether a given subsystem counts as IT or infrastructure.

A metric from 2007 that was meant for tracking one site

The Uptime Institute, a data center standards body, says the members of The Green Grid, an industry consortium, agreed on PUE in 2007. The aim was to let an operator track one site’s improvement over time. Uptime stresses it was not intended as a metric for comparing different sites.

That caveat matters, because PUE is now often used as a league table. Climate is one reason it misleads. In that same 2020 analysis, Uptime says hot or humid regions, or those short of water, can use free cooling less, which raises PUE. Sites in unstable grid regions may also build in extra redundancy, which uses more power.

The same article reports an average annual PUE of 1.59 among nearly 450 survey respondents in 2020, and little or no improvement since 2013. About one in six respondents did not know their largest site’s PUE.

Hyperscalers report near 1.1, and the average is higher

The gap between the best and the typical operator is large. Google’s efficiency page reports a 2025 fleet-wide average PUE of 1.09. For the second quarter of 2026 it lists a quarterly PUE of 1.10 and a trailing twelve-month figure of 1.09. Google describes PUE as a standard industry ratio comparing non-computing overhead energy with the energy used to power IT equipment.

Put next to Uptime’s survey figure of 1.59 for 2020, that is a different league. The two numbers are not strictly comparable, since one is a single company’s fleet and the other is a survey of mixed operators. Still, they show that purpose-built sites run by large operators carry far less overhead than the typical facility.

Why AI makes the number matter again

AI chips pack a lot of power into each rack, and all of that power ends up as heat. Moving it out takes cooling equipment, and that equipment draws power too. A cooling design that removes heat with less energy lowers the overhead in the numerator, and in this way directly lowers PUE.

This is the background to our report that Samsung is reportedly in talks to buy Modine, a maker of data center cooling systems. The talks are unconfirmed and rest on anonymous sources. The report does not say anything about PUE. The link is our own reading: cooling is the main piece of overhead that a supplier can shrink, so it is where equipment makers and their buyers compete.

What PUE leaves out

PUE says nothing about how much useful work the computers do per unit of energy. Two sites can share a PUE of 1.2 while one runs older chips that do far less per watt. A site can also lower its PUE without lowering its total consumption, since the ratio cares only about the split, not the size of the bill.

Other gaps come from what the ratio counts. Uptime notes that PUE does not capture all aspects of efficiency. Water use in cooling is a separate matter that has its own metric, called water usage effectiveness (WUE). Vertiv lists WUE and carbon usage effectiveness (CUE) as related measures.

One quirk is often cited by engineers. Fans inside servers count as IT load rather than overhead. Shifting cooling work from room-level equipment into the servers can therefore flatter the ratio without saving energy overall.

The practical advice follows from this. Read a PUE as a trend for one facility, check which period it covers (a quarter or a trailing year), and ask whether total energy and water use are reported alongside it.

What to watch next

Two things are worth following. The first is whether operators of AI-heavy sites keep reporting low ratios as rack power rises, or whether the average stalls as it did in the Uptime surveys. The second is whether buyers start asking for energy and water figures per unit of computing work rather than the ratio alone. For any cooling deal, including a confirmed Samsung one, the useful test is whether it shows up in disclosed efficiency numbers from the sites that use the equipment.