Decentralised GPU supply finally met real demand — unevenly

Decentralised compute networks spent their first years with a supply problem in reverse: plenty of registered hardware, almost no paying work. Utilisation figures published this quarter show that changing on three networks and not on the others.

The networks where utilisation rose share a characteristic that has nothing to do with token design. They solved verification — the customer can establish that the job ran on the hardware claimed, in a way that survives a dispute.

What buyers actually asked for

• Attestation that the advertised device served the job, not a slower substitute.
• Predictable eviction terms, so a long training run cannot be displaced mid-epoch.
• Data residency the buyer can verify rather than a claim in documentation.
• Settlement in a unit their finance team already accounts for.

Where the price advantage is real

For interruptible batch work — rendering, fine-tuning, evaluation sweeps — the discount against hyperscaler on-demand pricing is substantial and holds up under scrutiny. For latency-sensitive inference it largely disappears once you account for cold starts and routing overhead.

We moved our evaluation sweeps and kept serving where it was. That split was obvious in hindsight and took us a year to reach.

The networks that understood this positioned themselves as batch markets rather than cloud replacements. The ones still marketing themselves as a general alternative are the ones whose utilisation did not move.

What to watch

Whether attestation becomes portable across networks. Today each has its own scheme, which means a buyer's verification work does not transfer — and that is the main thing keeping this market fragmented.