Jun 2026
Why every decentralized compute project died
Nobody killed decentralized compute. The latency budget did.
Every few years someone raises a round to build the "AWS killer": a global mesh of idle GPUs and laptops that rents out compute with no datacenter, no Bezos, no rent extraction. Golem launched in 2016. iExec, SONM, Dfinity's early pitch, Akash, dozens of token-funded clones. The technology mostly worked. The marketplaces mostly cleared. And almost none of them captured a workload that anyone would have otherwise paid AWS for.
The reason is boring and physical, which is exactly why it keeps winning. Compute is not a commodity you ship around. It is a thing that has to sit very close to your data and very close to your other compute. Decentralization moves it far away. That trade is fatal for most of the workloads worth money.
The 10-gigabit cable is the whole product
Open a rack in any datacenter and the interesting part is not the chips. It is the wiring. Inside a single training cluster, GPUs talk over NVLink at 900 GB/s and across nodes over InfiniBand at 400 Gb/s, with switches tuned so that 1,000 GPUs behave like one big GPU. A modern training run is a tight loop: every step, every GPU exchanges gradients with every other GPU, then waits. The slowest link sets the pace for all of them.
Now spread those GPUs across home internet connections. A good residential uplink is maybe 20 Mbps. That is roughly 20,000 times slower than the InfiniBand link the workload was designed for, with latency measured in tens of milliseconds instead of microseconds. The GPUs spend their lives idle, waiting on the network. You have not built a supercomputer out of scattered parts. You have built the world's most expensive way to keep graphics cards warm.
This is why the decentralized projects that survived quietly stopped competing for the hard workloads. Akash rents whole machines, not slices of a training cluster. The honest version of the pitch is "spare VMs in cheap datacenters," which is a real business and a much smaller one than "replace the hyperscalers."
Verification is a tax nobody priced in
Centralized compute has a property you never think about: you trust the result. When AWS returns a number, you do not re-run the job on three other clouds to check that AWS did not lie. In a permissionless mesh, the node running your job is an anonymous stranger with a direct financial incentive to return return 0 instead of doing the work and pocket the fee.
So every serious design bolts on verification. Redundant execution means running each job two or three times on different nodes and comparing. That is a 2x to 3x cost multiplier before you have computed anything useful. The clever alternatives are worse in practice. Zero-knowledge proofs of computation can verify a result without re-running it, but generating the proof can cost 1000x the compute of the original job. You are paying a thousand dollars to prove you did a dollar of work honestly. Trusted hardware enclaves like SGX shrink the tax, but then your "decentralized" system trusts Intel's attestation servers, which is just centralization wearing a fake mustache.
The market that was already efficient
The pitch always assumes a giant pool of idle compute sitting around mispriced. There isn't one. Cloud spot markets already arbitrage idle datacenter capacity down to pennies, with the fast interconnect included for free. Crypto mining already vacuumed up the genuinely idle consumer GPUs and pointed them at the one workload that tolerates zero coordination between nodes: hashing, where every machine works alone and never talks to its neighbor. That is the tell. The only decentralized compute that ever scaled was the kind that needs no communication at all.
And the customers these projects wanted (AI labs, render farms, scientific computing) are exactly the customers who need tight coupling most. The workloads that could tolerate a loose, high-latency mesh are low-value batch jobs that spot instances already serve for cheap. So the addressable market is the intersection of "valuable enough to pay for" and "loose enough to decentralize," and that intersection is nearly empty.
None of this means the idea is fake forever. There are real niches: privacy-preserving inference where you genuinely cannot send data to a US cloud, censorship-resistant hosting, embarrassingly parallel science like Folding@home that ran for two decades precisely because protein folds need no cross-node chatter. Those are good and they are small. What dies, every single time, is the version that promises to undercut AWS on the workloads AWS actually makes money on. Physics set the price of moving a gradient across the internet, and no token can subsidize its way around the speed of light.
What we kept, what we cut
Verse does not argue with the speed of light. We do not split a training run across home uplinks, that grave is full. We cut the workloads physics forbids and kept the one nobody else nailed: verifiable single-capsule compute. The piece Golem and friends missed was verification, and replay is it. A deterministic capsule can be re-run by anyone, so an anonymous stranger’s result is checkable for the price of a sample, backed by a bond. Add content-addressed storage that pays nodes to actually serve, and a heat field that turns popular data into a CDN for free. Not an AWS killer. A verifiable computer for the jobs that do not need a 900 GB/s cable.