AI Data Centers Are Cleared for Takeoff, but How Much Payload Can They Carry?

NVIDIA and the hyperscalers are standardizing the ‘airframe’ for AI data centers. How much payload each facility can carry is a different question, and that’s where metrics like Power Compute Effectiveness (PCE) come in.

In brief: At the 2026 OCP Global Summit, NVIDIA, AMD, Google, Meta and Microsoft released open specifications for AI data center design, grid resilience and liquid-cooled racks supporting up to 250 kW. These standards define the interfaces, but not how much of a facility’s provisioned power is left for compute once cooling and other infrastructure take their share. That share is measured by Power Compute Effectiveness (PCE): Provisioned Power for Compute divided by Provisioned Total Data Center Power. It depends largely on thermal architecture.

Anyone who flew into San Jose for last week’s OCP Global Summit may have caught themselves asking the question every air traveler asks eventually: how does something weighing more than 70 tons get off the ground, and stay there?

Most OCP attendees will likely understand the physics of flight to some degree. But knowing how a wing works doesn’t stop any of us marveling when we see it in action.

For aircraft designers and airlines, though, the more interesting question isn’t how the plane flies. It’s how much it can carry when it does. And that turns out to be a useful lens on what came out of OCP.

Every aircraft has a maximum takeoff weight. Within that limit, every pound of airframe, systems and fuel is a pound that can’t be used for what actually earns the airline revenue: passengers and cargo.

Data centers are arriving at the same discipline. For most of the industry’s history, power was relatively easy to secure and floor space was the scarce resource. AI has turned that around. The grid connection is the AI data center’s maximum takeoff weight. Every megawatt claimed by the power and cooling infrastructure that supports the IT is a megawatt that can’t be allocated to compute.

The industry just published the airframe

The OCP Summit saw NVIDIA and the hyperscalers publish the latest blueprint for the AI data center “airframe”—the equivalent of the fuselage, wings, tail and landing gear that hold everything together.

• The Open Data Center: Next-Generation ML Infrastructure Design Principles (Revision 0.7.0), authored by AMD, Google, Meta and NVIDIA, sets common guidelines for the structural, power and cooling infrastructure that AI systems arriving in 2027–2028 will need.
• Alongside it, Google, Meta, Microsoft and NVIDIA published the Interface Specification Between Hyperscalers and Utility Grid for Disturbance Performance, which defines how large AI facilities should stay connected through grid disturbances.
• And NVIDIA contributed its MGX 1.2 rack design for Vera Rubin, which the company says is built around fanless liquid-cooled compute trays and standardized power and cooling interfaces, and supports racks of up to 250 kW.

The message was that with these standards, equipment makers, colocation providers and operators finally have a common target, and no longer have to wait on proprietary specifications from individual customers. It also confirms where the industry is focused: power, cooling, compute and the building itself should be engineered as one system.

The airframe is standard. The payload isn’t.

The specifications published at OCP define interfaces and envelopes. But what they don’t decide is how much of a facility’s provisioned power is left for compute once the supporting infrastructure has taken its share.

That is largely a thermal architecture question, and it is the question behind Power Compute Effectiveness (PCE), the metric Airsys introduced in a white paper accepted  through the Open Compute Project at the start of this month, following review by the OCP steering committee. PCE is a simple ratio: Provisioned Power for Compute divided by Provisioned Total Data Center Power. Start with the facility’s provisioned capacity, subtract what is provisioned for cooling and heat rejection, electrical losses and auxiliary loads, and express what remains as a share of the whole. In aviation terms, it measures payload: how much of the takeoff weight is left to carry passengers and cargo.

As the paper explains: “As compute density increases and utility interconnection constraints tighten across the industry, the Open Compute Project community is positioned to drive adoption of PCE as a standard reporting metric enabling operators, owners, and planners to evaluate infrastructure capacity allocation at the design level and make more informed decisions about how provisioned electrical capacity is put to work.”

The white paper’s worked examples show what is at stake. In an illustrative 20 MW colocation facility, 5.9 MW of provisioned capacity is allocated to infrastructure before it reaches IT, giving a PCE of 0.705. In an illustrative 100 MW hyperscale campus, the figure is 21 MW and the PCE is 0.79. In both, cooling and heat rejection make up by far the largest share: 4.8 MW and 16 MW respectively. That is power the operator has contracted, connected and paid for, but which can’t carry compute under that design.

At the 250 kW rack densities NVIDIA says MGX 1.2 supports, every megawatt the thermal plant claims is the provisioned power for four racks. Two operators can build to the same open specification, install the same racks and connect to a grid supply of the same size, and still end up with very different amounts of provisioned power left for compute. The difference is thermal architecture.

The standards flow downstream. So does the question.

The more consequential journey is the one these specifications are about to make. The hyperscalers wrote them around their own builds, but open specifications are designed to travel: through OEMs and system builders, into reference designs, and onto the floors of colocation providers and enterprises whose facilities were planned for a very different era of density.

As that happens, every operator will face the question the hyperscalers have just answered for themselves: of the power I have provisioned, how much can I actually allocate to compute? For operators with a fixed grid connection and a long queue for more, that decides how much of the next generation of AI hardware they can host.

Aviation learned long ago that payload is won or lost on the drawing board. The encouraging news for data centers is that the numbers needed to measure it already exist. PCE draws only on design and commissioning documents operators already hold, and it is open and vendor-neutral by design. The hyperscalers have published the airframe. The next step for everyone else is knowing how much payload their own facilities can carry, and what is weighing them down.

Read the white paper: Introducing Power Compute Effectiveness (PCE), a Provisioned Capacity Allocation Metric for the New Data Center Era