The Cheapest New AI Capacity Is the Hardware Already Paid For

Power is the constraint everyone now agrees on. Use is the constraint nobody publishes, and it is the larger number. The pitch for the grid starts there, with the claim that independent single tenant data centers average under 15 percent net compute utilization.

Axios reported the launch on 7 October. The membership list is not the story. The thing being sold is.

The grid is a market design before it is a product

A grid like this does not build anything. National Compute describes it as pooled research infrastructure, a shared network over machines that are currently isolated. Members contribute idle hardware, reserve larger clusters for planned runs, and pay only for the power their jobs draw. The system displays available capacity, chip type, location, price and utilization, then matches work to it.

The framing in the group’s own papers is deliberate. It compares the arrangement to the power grid and to the interstate highway system, two things that turned many private assets into one public utility. Access opened first to government, education and national laboratory teams. Crusoe, an infrastructure provider, is among the backers, alongside chipmakers, universities and institutional investors.

The supply estimate is where the claim becomes testable. Roughly 750 megawatts are connected or in sight, against 2 gigawatts by 2030. That starting figure is smaller than a single campus announced this year and larger than the entire footprint most enterprises will ever own.

Utilization is the number the buildout does not publish

The gap the grid is aiming at shows up in measurement. The 2026 State of Kubernetes Optimization Report is built on direct reads from tens of thousands of clusters across AWS, Azure and Google Cloud. Average CPU utilization fell from 10 percent to 8 percent in 2025. Memory utilization fell from 23 percent to 20 percent. GPU utilization across the whole sample averaged 5 percent, and on AKS it was 2 percent.

Overprovisioning moved the other way. CPU overprovisioning climbed from 40 percent to 69 percent, and memory overprovisioning reached 79 percent. Requests get written when a workload is deployed, and they get written conservatively, because the penalty for asking too little is an outage while the penalty for asking too much is invisible. The safe number becomes the permanent number, and the scheduler then treats it as a requirement.

A grid turns idle hardware into a scheduling contract

What has to work is not the fiber. It is the scheduler and the price. A job that runs for two hours needs to find the same silicon, the same interconnect and the same storage tier wherever it lands, and the operator holding the machine needs a reason to hand it over when a paying customer may want it tomorrow.

That is why the utilization numbers deserve one caveat. A dedicated training cluster at 5 percent is a different asset from a shared fleet at 5 percent. The first is idle between runs by design. The second is idle because nobody sized it correctly, and that is the pool a grid can actually harvest.

Three questions decide whether it does. Which provider gives up control of scheduling first. What happens to a running job when the owner wants the machine back. And whether a reserved block of capacity ends up cheaper than the same block bought directly.

Related reading. The data center backlash has not moved the capacity number, and Firmus announced 1.6 gigawatts and handed over 42 megawatts.