New PCE Metric Could Recover Stranded Data Center Power

retained text

New PCE Metric Could Recover Stranded Data Center Power
PCE measures how much provisioned power is available for compute rather than support systems, helping operators free up capacity for AI demand without new grid power.
As AI-driven demand strains data center power availability, operators are under pressure to do more within fixed electrical envelopes. In many cases, importing more power from the grid is not immediately possible, and adding new on-site capacity is expensive or slow. That is why a proposed metric, called power compute effectiveness (PCE), is attracting interest.
PCE is designed to show how much of a facility’s provisioned power is structurally available for compute and how much is tied up in cooling, electrical losses, and other support systems. The proposed metric is not positioned as a replacement for power usage effectiveness (PUE), the industry-standard metric that compares total facility energy to the energy used by IT equipment. Instead, proponents say it answers a different question, one that is becoming more urgent in retrofit and expansion projects: within a fixed power envelope, how much capacity is really available for servers?
“I see value in PCE beyond the initial design phase, particularly for operators evaluating retrofits and figuring out how much provisioned power can actually support IT equipment,” said Bill Kleyman, executive chair of Data Center Programs at AFCOM. “The right framing is to consider PCE as a complement to PUE.”
PUE vs. PCE
PUE has been one of the data center industry’s most widely adopted metrics for more than two decades. It helped focus attention on facility overhead and pushed operators to improve energy performance. According to Uptime Institute, average PUE has improved from roughly 2.5 to about 1.6 over the last 20 years, while some facilities now report 1.1 or better.
Even so, PUE has limits. It reflects operational efficiency once a facility is running and can leave room for interpretation about what counts as IT load versus facility load. Fan power, for example, may be classified differently from one site to another, and shared utility metering can further complicate the picture.
PCE, by contrast, looks upstream. In a paper published by the Open Compute Project, Airsys Cooling Technologies defines the metric as the share of a data center’s provisioned power allocated to compute rather than to supporting infrastructure.
“PCE asks a very simple question: Of all the power provisioned for a data center, how much is actually provisioned for compute?” said Paul Quigley, Chief Strategic Relations Officer at Airsys and one of the paper’s co-authors.
If a facility has 100 MW of provisioned power and 70 MW is structurally available for compute, its PCE is 0.70. The remaining 30 MW might be allocated to cooling and heat rejection, electrical conversion, distribution losses, and auxiliary systems such as lighting, controls, and security.
That distinction is central. PUE asks how efficiently a running data center uses energy. PCE asks how a data center’s provisioned power has been allocated before the servers are even turned on.
Why the Metric Matters
The appeal of PCE becomes clearer when operators are trying to add more compute without expanding their total power envelope. If a 100 MW facility initially reserves 30 MW for cooling, losses, and auxiliary loads, it has a PCE of 0.70. If an infrastructure redesign cuts those support requirements to 20 MW, the same facility can support 80 MW of compute and raise its PCE to 0.80, all without increasing total provisioned power.
“The infrastructure changes needed to achieve this gain may require significant capital investment,” Quigley said. “What PCE uncovers is whether investment in cooling, electrical architecture, or other infrastructure can reallocate a larger portion of the existing power envelope to compute.”
That is particularly relevant today, when utility constraints are making new capacity harder to secure in many markets. PCE does not tell operators that every non-compute megawatt can or should be recovered. What it does is make the existing allocation visible, giving designers and operators a way to evaluate whether changes to the infrastructure could free up more power for IT.
Quigley offered an example: a facility may have 20 MW provisioned for an older cooling architecture, 7 MW associated with electrical conversion and distribution losses, and 3 MW dedicated to auxiliary loads. An engineering study might show that a cooling retrofit could reduce the cooling allocation from 20 MW to 12 MW. If that change is technically and financially justified, the facility could make another 8 MW structurally available for compute, subject to its remaining constraints.
Because the metric is new, there is no validated industry average yet. In the Airsys/OCP paper, two real-world examples produced PCE scores of 0.705 for a 20 MW colocation site and 0.79 for a 100 MW hyperscale campus. Quigley said many older, air-cooled facilities may land between 0.50 and 0.65, while well-designed liquid-cooled sites could reach the high 0.80s. For now, those ranges should be viewed as directional rather than as established benchmarks.
Where PCE May Be Most Useful
PCE appears to have its greatest near-term value in design work, retrofits, and capital planning.
In retrofits, PCE can highlight where infrastructure overhead is consuming an outsized share of provisioned capacity and help operators identify which upgrades are most likely to recover stranded power. A 30 MW site built 15 years ago, for example, might have a PCE of 0.60, leaving 18 MW for compute and 12 MW for support systems. If most of that support load is tied to cooling rather than to UPS losses or building loads, the analysis may point toward cooling upgrades as the better use of capital.
For design teams, PCE can also serve as a planning tool before construction begins. Quigley described a comparison between two hypothetical 50 MW designs: a conventional chilled-water plant requiring 10 MW for cooling, and a warm-liquid loop with dry coolers requiring 5 MW. If electrical losses and building load are the same in both cases, the first design leaves 37 MW for compute, while the second leaves 42 MW. That is the difference between a PCE of 0.74 and 0.84, decided on paper rather than after commissioning.
“That is a 5 MW difference decided on paper, before any concrete is poured,” Quigley said. “That’s an example where PCE differs from PUE, which can only be measured after the meter is spinning.”
The same logic can apply to capacity planning. For example, a 40 MW colocation facility with a PCE of 0.70 supports 28 MW of compute. If a customer needs an additional 6 MW for an AI deployment, the conventional answer may be to seek more utility power. A PCE analysis asks a different question: whether the site can raise its score to 0.85 through infrastructure changes. If so, the existing 40 MW envelope could support 34 MW of compute, enough to meet the demand without a new utility application.
At a portfolio level, PCE could also help operators compare sites and prioritize investment. A facility with the lowest score may not necessarily offer the best return on capital; another site with a slightly better PCE may have far more recoverable capacity. Used that way, the metric becomes less about benchmarking for its own sake and more about directing money toward the upgrades that can free the most IT load.
Will PCE Gain Traction?
That remains an open question. PUE became ubiquitous because it was simple, intuitive, and tied to a broad industry push toward energy efficiency. PCE is narrower. It is more likely to resonate with designers, infrastructure planners, and operators working through constrained-power retrofit decisions than with the wider data center market.
Kleyman said that is exactly how the industry should view it. “Our industry must constantly revisit how we build and design facilities and understand what those decisions mean for capacity and efficiency,” he said. “PCE’s traction will depend on whether it helps operators make better investment decisions, using consistent assumptions about cooling, redundancy, and facility boundaries. Another number alone won’t move the industry, but a clearer path to putting scarce power to work certainly could.”

document fetched 2026-10-24 11:21 · canonical https://datacenterknowledge.com/energy-power-supply/new-pce-metric-could-recover-stranded-data-center-power