Most decentralized GPU networks advertise huge machine counts. Verified supply tells another story: verification, not price, is the real moat.
The GPU shortage is real. So is a pile of idle hardware nobody can trust.
Disclosure: we're building a venture in this category. The critique below sticks to public numbers, and we name our own stake up front.
A new "decentralized GPU network" seems to launch every few weeks, each promising cheap AI compute from hundreds of thousands of machines. The pitch lands. Then you try to run a real job, and the gap between the homepage and the network shows up fast.
The largest decentralized GPU network, by its own January 2026 report, showed roughly 2,752 GPUs that passed verification against a registered base it has put north of 327,000. Even taken at face value, that's under 1% verified.
Passing verification isn't a perfect proxy for "usable." But a roughly 100-to-1 gap between registered and verified tells you the headline number isn't the one that matters.
A network counts the machines that signed up. By default it doesn't know whether a GPU is idle or already busy, whether it's the GPU the owner claims, or whether the result it returned was actually computed, or fabricated to collect the reward.
That last one, fake work paid as real, is the Sybil problem, and it isn't theoretical. Networks that pay for participation tend to attract participants who farm the payout rather than do the work. The headline supply number climbs while usable supply stays flat.
A registration count measures how many machines signed up. It says almost nothing about how many can actually do the job.
Someone fine-tuning a model or serving inference doesn't care about a 327,000 figure on a homepage. They care about three things, in roughly this order. Isolation, so the machine's owner can't see their data or their weights. Proof that the job actually ran, on the hardware that was claimed. And availability, so the GPU is there when the work is.
Price matters too, but it's largely a solved problem: decentralized compute already runs 60-80% below the hyperscalers on comparable hardware. The harder, less crowded question is trust.
The networks pulling real, paid workloads today curate supply by hand: whitelists, closed supplier lists, manual vetting. It works well enough to win the jobs, but hand-vetting every machine doesn't scale to hundreds of thousands of them. The open problem is making verification automatic and cryptographic at open-network scale, so a buyer can trust a machine nobody hand-approved.
That's the bet we're making with the AI venture inside our studio: compute that's verifiable, isolated, and available. The goal was never to count the most GPUs. It was to stand behind the ones we list.
Nothing here is an offer to sell a security or investment advice.
Field notes on venture building, AI, and capital. No spam, unsubscribe anytime.
By subscribing you agree to receive AOS Insights e-mails. We use your address only for this newsletter - see our Privacy Policy.
