Why Are Companies Still Buying GPUs If So Many Sit Idle?

A new consortium says most AI data centers use under 15% of their chip capacity. It wants to connect idle machines instead of building more.

A consortium called the National Compute Network launched on October 7, 2026, with a specific complaint: the AI industry has spent two years treating compute like it's scarce, when a meaningful share of what's already been bought may just be sitting there.

The number behind that complaint is blunt. Independent, single-tenant data centers average less than 15% net utilization, according to a document the consortium itself prepared. Read that twice. Companies are paying for tens of thousands of GPUs, and the group says most of that capacity goes unused most of the time.

The network isn't proposing another data center. It wants to connect the ones that already exist. One lab needs compute. A cloud provider somewhere else has idle chips. A shared scheduling system is supposed to match the two, showing available capacity, chip type, location and price across every participant.

What the National Compute Network actually proposes

The group has roughly 760 megawatts of compute connected or scheduled to connect today, with a stated goal of 2 gigawatts by 2030. Anjney Midha, a former Andreessen Horowitz general partner, is named as one of its organizers. Sam Singh, who leads AI at the startup 1X, is quoted supporting the effort. Access is meant to extend beyond private labs to public-sector employees, universities and national labs.

How the matching system is supposed to work

In practice, this isn't as simple as flipping a switch. Different workloads need different chip types, cluster sizes and network layouts. A thousand GPUs spread across ten buildings don't behave like a thousand GPUs wired together in one room. The consortium's own framing treats this as a scheduling problem more than a hardware one, which is a narrower claim than "we've solved the GPU shortage."

Can You Trust the 15% Utilization Number?

Not fully. The 15% figure comes from the consortium itself, not an independent audit, so treat it as a claim rather than settled fact.

Not yet, and not blindly. The 15% figure is doing a lot of work in this story, and it comes from the organization with the clearest incentive to make compute sound underused. That's not an accusation. It's just how the arithmetic of credibility works: the group selling the grid wrote the number that justifies the grid.

Who produced the claim, and why that matters

No independent audit of that 15% figure has surfaced yet. So treat it as a claim under review rather than an industry fact. Here's why the story still holds up even if the real number is 25% or 40% instead of 15%. At any of those levels, a company spending heavily on scarce chips while a large pool of existing hardware sits unreachable outside its own cluster has a real problem. Owning compute and being able to use it when you need it are not the same thing.

ClaimSourceStatus
760 MW connected, 2 GW target by 2030National Compute Network consortiumReported, attributed to the group
<15% average utilization at single-tenant data centersConsortium's own documentClaimed, not independently audited
$500B+ in AI infrastructure financing MOUsNvidia press release, August 2026Confirmed by Nvidia, non-binding

Why an Idle GPU Costs More Than Idle Warehouse Inventory

A GPU, the building, the cooling and the power hookup around it are sunk costs. Every idle hour is money nobody gets back.

Unused steel in a warehouse is cheap to wait on. An idle GPU isn't, and here's the part worth sitting with. The chip is expensive. The building around it is expensive. The cooling system, the networking and the power connection it took months to secure are all expensive too. None of that spending pauses while the machine sits empty.

Then a newer chip generation ships, and the clock resets. Every hour a GPU sits idle is an hour of depreciation nobody gets back. That's the logic that makes the National Compute Network's pitch land even before you accept its exact percentage: a routing failure on hardware this expensive is a different kind of waste than a routing failure on steel.

The electricity analogy, and where it breaks

The easiest comparison is the power grid. Nobody building a factory owns the specific plant generating their electricity. Generation sits in one place, demand sits somewhere else and a transmission system connects them. If compute worked the same way, ownership and access would split apart the way they already have with electricity.

The analogy only goes so far. Electrons move across any wire built for them. AI workloads don't move across any GPU cluster built for them. Training runs often need thousands of chips wired tightly together, and a company isn't likely to lend out hardware it might need for its own run next week. The National Compute Network is a real attempt at a hard coordination problem, not a settled solution to one.