PCE: the metric that measures the MW that never reach your racks

If your provider quotes you 40 MW and you accept them, once your racks are full you will discover from the single-line diagram that only 28 MW reach compute. The other 12 went to cooling, electrical conversion losses and auxiliary loads: load you never saw on the invoice because it was never measured. For two decades the industry called that gap “bad PUE”. The Open Compute Project has just given it its own name: PCE, the fraction of your provisioned power that actually reaches compute. It is not another efficiency metric, it is a design metric, and it changes the question you ask your provider.

In Mexico, where 72% of the installed capacity is already in a single state and asking the grid for MW takes years, that question stopped being academic.

One clarification before going further: the framework is new, accepted by the OCP steering committee on October 5, 2026. It has methodology, defined variables and two real cases calculated: that is enough to make design decisions, even though there is not yet a validated industry average.

The question PUE never answered

PUE measures operational efficiency: total energy consumed divided by that of the IT equipment, with the meter running. It is useful, but it answers about the past: your PUE of 1.3 describes how you consume today, not how much of what you already paid for reaches the rack.

PCE moves the focus to the start of the design chain. It is defined as the provisioned compute power, structurally available within a declared boundary, divided by the total provisioned power under a declared redundancy basis. In simple formula: PCE = power for compute ÷ total provisioned power. It is a dimensionless value between 0 and 1.

Three items are subtracted from the total provisioned, and it is worth naming them precisely because each one requires a different intervention:

Before calculating, the framework requires two explicit declarations: the electrical boundary of the installation (which sources and which loads enter the calculation) and the redundancy basis (N, N+1 or 2N). All variables are expressed under that basis. Without them the number is not comparable to anything.

The two numbers the document calculates

The OCP document resolves two real cases:

The first leaves 14.1 MW available out of 20 MW; the second, 79 MW out of 100. Neither is an industry average: they are specific installations, and for that reason they do not serve as a reference.

The case that gets cited most is design against design. Two 50 MW plants, one with chilled water that takes 10 MW of cooling and another with a warm liquid loop and dry coolers that takes 5 MW. With the same electrical losses and the same building load, the first leaves 37 MW for compute and the second 42 MW. The difference between a PCE of 0.74 and one of 0.84 is five megawatts decided on paper, before pouring concrete. The question is no longer “is my PUE good”, but “does my design deliver compute or cooling”.

Why this weighs more in Mexico than in an unconstrained market

The usefulness of PCE multiplies where the electrical grid cannot grow at the pace of demand, and that is the Mexican case.

The MEXDC 2026 Mexico Data Center Market Report brings together the following data: Querétaro adds 202 MW in commercial operation, 72% of the 279 MW installed nationwide. The state records 181 MW under construction and 1,192 MW announced through 2031. ODATA operates more than 100 MW there; alongside it are Microsoft, Triara, KIO, Amazon, Equinix and Ascenty. Direct investment announced for 2026-2031 amounts to 20,600 million dollars.

When 72% of your national market is in a single state and the project pipeline through 2031 exceeds a thousand megawatts, the question “how much power do you have?” depends on the grid. The question “of the power you already paid for, how much reaches the server?” depends on you and on your retrofit. It is the only one an operator can move.

And there is the other one that almost nobody measures. A 30 MW installation built 15 years ago can have a PCE of 0.60: 18 MW for compute and 12 MW in support, most of it in cooling. If an engineering analysis shows that a retrofit brings that line down from 20 MW to 12 MW, 8 MW become available for compute. For a customer that needs 6 MW more for AI, the conventional answer is to ask the grid for more power; the question PCE does ask is whether the site can raise its value to 0.85 and release those 6 MW without new interconnection.

The order matters: a site with the lowest PCE is not necessarily the best investment, and another with a barely better value may have much more recoverable. The metric serves less to compare and more to direct capital.

The three declarations before asking for the figure

If you are going to ask your provider or designer for the PCE, the order matters: almost nobody has the number today.

A document that answers those three with verifiable numbers is a design; one that answers with a PUE and a percentage saving is a quote.

What PCE still is not

It is worth saying with the same clarity with which the metric is announced, because the market is overexposed to numbers that look precise: there is still no validated industry average. The ranges that circulate —old installations with air cooling between 0.50 and 0.65, sites with liquid cooling in the high 0.80 range— are directional, declared as such by the authors, and must be cited as “Airsys and the OCP suggest”, never as a market reference.

PCE also does not free megawatts on its own: it makes the existing allocation visible so that designers can evaluate whether an infrastructure change releases more power for IT. And it does not replace PUE: it complements it. The framework summarizes it in one sentence: PUE looks through the rear-view mirror of operational efficiency, and PCE is meant to be the windshield. Its acceptance depends on whether it helps decide better with consistent assumptions about cooling, redundancy and boundary; it is expected to resonate with design teams and with the operators who resolve retrofits under power constraint.

What to do in the next 30 days

Request the single-line diagram of your site and make it count. The five PCE numbers come out of it: total provisioned power, cooling, losses, auxiliary loads and the boundary. It is an afternoon’s exercise with your electrical team. And define the boundary and the redundancy basis before comparing anything: without those two declarations any figure you receive is uninterpretable, including the provider proposals that offered “energy efficiency” as an argument.

Calculate the PCE of your current installation with your own figures and compare it with that of a hypothetical retrofit. You do not need any tool: it is a subtraction and a division over variables you already documented. It is the same inventory as a data center energy audit, with a different variable at the center.

If your project asks for new MW, first ask yourself how many of the MW you already have are trapped. The answer does not come from the grid provider or the regulator, but from your design documentation. And in the Bajío, where the grid is already at its limit, the constraint documented in our Querétaro electricity consumption profile gives it market value.

Sources