As AI workloads push data center power demand higher, many operators are stuck working within fixed electrical envelopes. Pulling more power from the grid is often not an option in the near term, and building new on-site capacity is both slow and expensive. A newly proposed metric called power compute effectiveness (PCE) is drawing attention as a way to help operators squeeze more usable compute capacity out of the power they already have provisioned.
What PCE Measures
PCE is defined in a paper from the Open Compute Project, authored by Airsys Cooling Technologies, as the share of a facility’s provisioned power that is structurally available for compute rather than tied up in cooling, electrical conversion losses, and auxiliary systems like lighting and security. If a site has 100 MW of provisioned power and 70 MW of that is actually available to run servers, its PCE score is 0.70.
Paul Quigley, Chief Strategic Relations Officer at Airsys and a co-author of the paper, described the core question PCE answers: of all the power provisioned for a data center, how much of it is actually provisioned for compute.
How It Differs From PUE
Power usage effectiveness (PUE) has been the industry’s standard efficiency metric for more than two decades, and according to Uptime Institute, average PUE has improved from roughly 2.5 to about 1.6 over that period, with some sites now reporting 1.1 or better. But PUE measures operational efficiency once a facility is already running, and it can be inconsistent in how it classifies loads like fan power or shared utility metering.
PCE instead looks upstream at how provisioned power is allocated before servers are even switched on. Bill Kleyman, executive chair of Data Center Programs at AFCOM, said the metric is most useful for operators evaluating retrofits and trying to determine how much provisioned power can actually support IT equipment, framing PCE as a complement to PUE rather than a replacement.
Why It Matters Now
The metric’s value becomes clearest for operators trying to add compute capacity without expanding their total power envelope. Quigley offered an example of a facility with 20 MW allocated to an older cooling architecture, 7 MW to electrical losses, and 3 MW to auxiliary loads. An engineering study might find that a cooling retrofit could cut the cooling allocation to 12 MW, freeing up 8 MW of structurally available compute capacity, assuming the investment is technically and financially justified.
Because PCE is new, there is no validated industry average yet. The Airsys/OCP paper cites two real-world examples: a 20 MW colocation site scored 0.705 and a 100 MW hyperscale campus scored 0.79. Quigley said many older air-cooled facilities may fall between 0.50 and 0.65, while well-designed liquid-cooled sites could reach the high 0.80s, though these figures should be treated as directional rather than established benchmarks.
Where It Fits Best
PCE appears most useful in design work, retrofit planning, and capital allocation decisions, where it can highlight how much of a facility’s provisioned power is going to overhead rather than compute and help operators prioritize which upgrades offer the best return.