Decentralised compute networks have moved past the pilot stage with real inference workloads and paying customers. That is genuine progress, and it also removes the excuse for vague reporting.
The number that matters
Paid utilisation. A network can show enormous registered capacity while renting almost none of it, and can show impressive apparent revenue that is actually its own token emissions circulating back through the system. Separate customer payments in stable value from protocol rewards, and most of the sector's headline figures shrink considerably.
Where the model genuinely fits
Batch inference and fine-tuning workloads that tolerate heterogeneous hardware and variable latency. Frontier training does not fit and probably never will; the interconnect requirements are not negotiable.
The honest version of this sector is a spot market for spare accelerator capacity. That is a decent business. It is not the one the louder projects describe.
Utility demand versus token demand
Networks selling real compute still see most token activity driven by speculation rather than usage. Separating the two in reported metrics is rare.
Published utilisation rates and paying-customer counts would settle the argument quickly, which is presumably why so few projects publish them.
