National Compute's shared AI grid counts 760 MW as connected or in sight

"Connected or in sight" puts two different states under one number ^[1]. Of the two words, "connected" is the one a scheduler can send a job to. According to Runtimewire's reading of the launch, the wording does not say how much of the 760 MW is operational, contractually committed or available to customers now ^[3]. Against the 2030 target ^[9], the figure is 38% of the goal ^[17], and that share assumes every in-sight megawatt shows up.

National Compute calls its software layer Grid Exchange ^[4]. Members can put idle capacity into it or reserve larger clusters for planned training runs, and the scheduler routes workloads across participating sites and systems ^[4]. Vultr says the pool will span NVIDIA, AMD and TPU architectures ^[10]. Chip type is one of the fields providers list capacity under ^[2]. A given job can only use the part of the pool listed under hardware it runs on, so the headline figure overstates what any one customer can reach.

Goodput pricing ^[11] is the best design decision in the launch. Under it, a preempted job or a failed node is the provider's loss. An operator that lists unreliable spare capacity pays for the unreliability, and the cost lands on the party that controls it. I think it is the right default for a pool of independent operators who do not share an on-call rotation.

The supply case rests on idle hardware. The consortium's paper, as Axios summarized it, says independent single-tenant data centers average less than 15% net computing utilization ^[5]. Taken at face value, more than 85% of their capacity goes unused ^[16]. The figure is the group's own, not an independently audited measure ^[6]. For it to carry over to this grid, the idle hardware has to sit at sites that join. It also has to be free when a job asks for it.

Demand is easier to show. The consortium says it is opening access to public-sector teams, including government, education and national laboratory users ^[12]. Sam Sinha, head of AI at robotics company 1X, told Axios that when it comes to large, long-term contracts, smaller operators have a hard time competing with OpenAI and Anthropic ^[13]. Midha helped some founders secure computing access when he was an a16z general partner ^[15]. He told Axios that U.S. compute capacity is larger than most people assume, and that what it lacks is interconnection and coordination ^[14].

The launch materials describe an intended design, not independently demonstrated results ^[7]. If every counted megawatt does convert, the consortium still needs another 1,240 MW to reach 2 GW by 2030 ^[18].