Measuring the Overhead

In 2007, an industry consortium called The Green Grid proposed a simple ratio for judging how efficiently a data center uses energy. It called the ratio power usage effectiveness, or PUE. You take all the energy entering the facility and divide it by the energy delivered to the information technology equipment, meaning the servers, storage and network gear. A facility with a PUE of 2.0 spends as much on overhead, chiefly cooling and power conversion losses, as it does on computing. A facility at 1.0 would spend nothing on overhead, which no real building achieves. The metric was later standardized internationally by the International Organization for Standardization and the International Electrotechnical Commission as the standard numbered 30134-2, published in 2016.

The reason PUE caught on is that it is easy to explain and hard to argue with. The reason it is still debated is that it measures exactly one thing. It says nothing about how useful the computing is, how clean the electricity is, or how much water the building evaporates. A data center can lower its PUE by switching from electric chillers to evaporative cooling and, in doing so, raise its water use considerably. Hold that trade-off in mind, because it returns in the final module.

What does the industry actually achieve? The Uptime Institute, which runs an annual global survey of data center operators, has tracked average PUE for years. In its historical series, the average sat around 2.5 in 2007 and fell steeply over the following decade as operators contained their aisles, raised their temperatures and bought better equipment. Then the improvement stalled. Uptime's 2024 survey reported an average of 1.56, and its 2025 survey reported 1.54, which Uptime described as little changed for about six years. The survey is self-reported and skews toward the operators who answer surveys, so treat it as a reasonable portrait rather than a census.

At the other end of the range, Google publishes a trailing twelve-month PUE for its large data centers, measured across all seasons. It reported a fleet average of 1.09 for 2024. The comparison with the survey average is instructive. Google builds very large, purpose-designed facilities, controls the server hardware inside them, and has spent years tuning the controls. An older enterprise building with mixed equipment and a landlord's air handlers cannot copy that simply by wanting to. The gap between 1.09 and 1.54 is mostly a gap in design freedom, not in effort.

Much of the overhead reduction that did happen came from a change in what operators believed servers could tolerate. The American Society of Heating, Refrigerating and Air-Conditioning Engineers, or ASHRAE, runs a technical committee known as TC 9.9 that publishes the Thermal Guidelines for Data Processing Environments. Its recommended envelope for server inlet air, which ASHRAE widened to 18 to 27 degrees Celsius in 2008, is warmer than the chilly halls of the 1990s. It also defines allowable classes, labeled A1 through A4, that let equipment run warmer still for limited periods. The fifth edition, published in 2021, is the current reference.

Warmer inlet air matters because it lets a facility use outside air or cool outside water for more hours of the year instead of running refrigeration. Engineers call this economization, and it is where most of the savings came from. A building in a mild climate can spend much of the year with its chillers idle.

There is a subtlety buried in PUE that trips up even experienced readers. The denominator, the energy delivered to the information technology equipment, includes the fans inside the servers. If an operator raises the room temperature, the server fans spin faster, the denominator grows, and PUE improves on paper even if total energy barely moves. ASHRAE's guidelines discuss this interaction openly. The metric rewards shifting overhead from the building into the box.

Faraday's read is that PUE is best used as a question rather than a grade. When a developer quotes a design PUE, ask whether it is an annual figure or a peak one, whether it was measured or modeled, what climate it assumes, and whether the design relies on evaporating water to reach it. A low number with good answers to those questions is informative. A low number without them is a sales figure, and there is no shortage of those.