The compute shortage is real. The trust shortage is the one that pays.

There's a genuine scarcity of GPUs and, at the same time, a pile of idle hardware nobody will touch. Both are true at once, and the distance between them is trust, not supply.

Disclosure: we're building a venture in this category. Every figure is attributed to a named source, and we name our own stake where it bears on the argument.

Two facts that seem to contradict each other, both true at once.

Fact one: there is a real shortage of AI compute. GPU-ready data center capacity sits below 1% vacancy in prime hubs, with power availability, not space, as the number-one bottleneck (QuoteColo, 2026). The good clouds are full and pricey.

Fact two: there is a vast amount of GPU hardware sitting idle right now. Tens of millions of consumer cards, dark most of the day. Registered supply on decentralized networks that dwarfs what's actually in use.

Hold both. The shortage and the glut are real simultaneously. Which means the problem isn't quite what it looks like.

The shortage you can measure

Start with the part that's straightforwardly true. Capacity is genuinely tight. Vacancy under 1% in the prime markets, top-tier operators gating single racks behind 100 kW-plus minimums, smaller teams pushed to regional providers because they can't get power at the front of the line (QuoteColo, 2026). Pricing reflects it: global weighted-average colo costs rose to roughly $217 per kW per month and climbed from there, with Singapore running $310 to $470 (CBRE, "Global Data Center Trends 2025," Jun 24 2025).

If that were the whole story, the fix would be obvious: build more. And the industry is. But it doesn't explain the other fact.

The glut hiding in plain sight

Because at the very same moment, an enormous amount of GPU capacity sits unused. On one large network, roughly 2,752 GPUs were verified against a registered base put above 327,000, under 1% in use (a leading decentralized GPU network, 2025 year-in-review, Jan 2026). Read that as the opposite of a machine shortage: a mountain of registered hardware almost none of which is doing paid work. Add the consumer pool: a four-card gaming rig can earn its owner $700 to $1,400 a month at high utilization (a major GPU marketplace, May 2026), and most of those cards earn nothing because their owners never put them to work.

So we have scarcity and abundance in the same market. The GPUs exist. They're just not the GPUs a serious buyer is willing to use.

There's no shortage of graphics cards. There's a shortage of graphics cards a buyer is willing to trust. Close that gap and the glut becomes supply.

What actually separates the two

Walk up to that idle hardware as a buyer and you immediately understand why it's idle. You can't trust it. You can't confirm an anonymous machine ran your job honestly rather than faking a result to collect a fee, a real failure mode on networks that pay for participation. You can't be sure the operator can't see your data, and on consumer rentals the honest answer is sometimes that they can: "a host can snoop on your workloads," as one widely-cited marketplace thread put it (Hacker News, discussion 36026101). You can't get a record an auditor would accept.

That, not silicon, is the wall. The glut is untrusted, and untrusted compute is, for any buyer who cares about correctness or confidentiality, the same as no compute at all. The scarcity is one of trust before it is ever one of GPUs.

Why this is the part that pays

This is the commercially interesting part. Adding raw GPUs to the world is capital-intensive and slow, and everyone with money is already doing it. Adding trust to GPUs that already exist is an engineering problem, and it converts a near-free mountain of idle hardware into usable supply.

Solve trust, and the glut stops being inert. Verifiable work turns an anonymous machine into one a buyer can rely on, the mechanism we cover in Proof you can't fake. Structural isolation turns a snoopable host into a safe one, which we cover in What "isolated compute" actually means. The verified-versus-registered supply gap that defines the whole opportunity is the flagship, The sub-1% problem in decentralized compute.

So the interesting work isn't manufacturing more GPUs, which is slow, capital-heavy, and already crowded. It's making the idle ones trustworthy: verifiable inference so a buyer can trust the work, per-job isolation so they can trust the privacy, reputation scoring so trust compounds over time. The compute shortage gets solved with capital. The trust shortage gets solved with architecture. Only one of those is a business you can start today.

Nothing here is an offer to sell a security or investment advice.