A coalition of AI startups, cloud providers, researchers and investors launched the National Compute Grid on Tuesday, an effort to pool AI hardware across competing labs and rent it out through one shared scheduling system. The pitch rests on an uncomfortable number from the consortium’s own paper: independent single-tenant data centers average less than 15 percent net computing utilization, meaning some of the most expensive chips ever built sit idle most of the day while smaller labs call around trying to find capacity for a training run.
The planning behind this one is not new. Data center supply has tightened for two years and hyperscalers have responded by buying more. The consortium’s draft paper, reviewed by Axios, argues the cheaper first move is to actually use what exists. Capacity, chip type, location, price and utilization would show up in one place, and the scheduler would match workloads to hardware automatically.
Members can contribute unused machines and reserve larger clusters ahead of planned training runs. Access is open to commercial users as well as to government, education and national laboratory teams. The group says roughly 760 megawatts of compute is connected or within reach, and it targets 2 gigawatts by 2030. Initial access during the preview phase runs through the company’s website for people with .gov or other public institution emails.
Who is behind it
Anjney Midha, a leader of the effort, told Axios the project exists to share computing for commercial research and public-sector use. “The best way to scale AI in America efficiently, and stay at the frontier and stay competitive with China, is to be” coordinated around an open standard, he said. Sam Singh, head of AI at robotics startup 1X, framed the access problem from the buyer side: larger players like OpenAI and Anthropic can pay far more and lock up capacity on long-term contracts that smaller operators cannot touch. “We need to encourage a healthy AI ecosystem, and have more than two companies to own all the compute,” he said.
The public-sector side already carries a dollar figure. POLITICO reports that National Compute, the company behind the grid, plans to donate $100 million in computing credits to the White House, expected to be announced Thursday at an event with Office of Science and Technology Policy director Michael Kratsios. The credits would go toward the Genesis Mission, the administration’s government-wide push to use AI to accelerate scientific discovery. The donation represents a private-sector contribution to the administration’s science agenda rather than a new government appropriation, and it gives the grid a federal anchor tenant from day one.
Why utilization is so low
The 15 percent figure needs context. Training runs are bursty. A lab may fill a cluster to capacity for weeks, then finish the run and leave the machines sitting while researchers write up results or wait for the next dataset. Cloud contracts rented on long-term commitments reward that pattern, since the lab pays for the cluster whether it uses it or not. A shared scheduler attacks the in-between hours, offering idle time to a second tenant at rates the first tenant’s contract can absorb. Airlines fill seats this way. Most AI infrastructure until now has not worked like that.
| Grid parameter | Figure |
|---|---|
| Capacity connected or in sight | About 760 megawatts |
| 2030 target | 2 gigawatts |
| Single-tenant DC average utilization | Under 15 percent |
| Federal compute credits pledged | $100 million |
The backstory on timing
The launch landed the same week several big-AI money stories broke. Lambda, the Nvidia-backed AI cloud provider, is raising up to $4 billion at a $14.5 billion pre-money valuation, with its contracted backlog reportedly jumping from about $15 billion in June to $50 billion in September. Marvell Technology set itself a target of $70 billion to $90 billion in annual revenue by fiscal 2031, up from roughly $8.2 billion this year. SpaceX is in early talks to borrow $40 billion to buy Nvidia GPUs. Every one of those deals assumes new building. The grid assumes reuse. If both the building and the reuse happen, the supply picture for AI compute looks materially different from the one in current analyst models, which mostly extrapolate from construction pipelines alone.
The hard parts
Pulling it off means solving problems the industry has usually dodged. Cross-platform scheduling requires a common way to describe jobs across Nvidia, AMD and custom accelerator hardware, which touches the portability problem nobody has fully cracked. Workload isolation across competing tenants raises security questions whose answers are not cheap, especially for frontier-model training data that companies treat as confidential. And the pilot phase limits access to public institutions with .gov emails, so commercial availability, where the utilization promise actually pays off, remains a promise for now.
The economics work only if two assumptions hold. First, that idle capacity exists at the scale the consortium claims, which its utilization data supports at least for the independent segment. Second, that owners of that capacity will contribute it rather than hoard it, since a lab that rents out its slack clusters is arming a potential rival with compute. Singh’s argument covers the first assumption. Nobody has yet answered the second in public. The $100 million in federal credits solves it for the government’s side by making the grid a recipient of state-backed demand rather than a neutral marketplace, which may be the reason it exists at all.
The contract language is worth reading when it lands. Whether the scheduler serves public research at low cost, or simply becomes another bilateral deal layer, will tell you whether this is a genuine common carrier for compute or a marketing wrap on the same closed arrangements that run AI today.