The test of useful work is brutal and simple: would anyone buy the work at the price the network pays for it? For most proof-of-useful-work systems that spread — call it the usefulness gap — is wide, measurable from public data, and quietly does all the economic work the whitepapers attribute to usefulness itself.
- Usefulness claims split into classes — inference, training, rendering, storage, scientific compute — and each class carries a different verification burden, which is where the economics hide.
- The verification tax is the design constraint: checking that useful work happened costs real compute, and every design cuts that cost by pushing the work back toward lottery-like uniformity.
- The arithmetic on the largest example: after Bittensor's December 2025 halving, emissions run 3,600 TAO/day with 41% to miners. At $300/TAO that is ≈$443k/day of miner pay — enough to rent roughly 7,400 H100s around the clock at open-market rates.
- The gap question is whether miner output would fetch that $443k from actual buyers. Emission-funded pay is a subsidy until demand-side revenue matches it — for most PoUW networks it does not yet.
- The gap closes only through buyer-priced demand: metered inference, task markets, fee burn. Emissions bootstrap supply; they cannot manufacture demand.
Grayscale — Bittensor on the eve of the first halving
Zpoken decentralized-AI practice
A taxonomy of usefulness claims
“Useful work” is five different engineering problems wearing one name. Inference: outputs are cheap to produce once, expensive to verify without re-running. Training: months-long, stateful, and the hardest to verify of all — you are attesting a process, not an output. Rendering: verifiable by sampling, tolerant of spot checks. Storage: the solved case — proofs of replication and spacetime are mature precisely because storage is uniform. Scientific compute: verifiable when the problem has a checkable certificate, unverifiable folklore when it does not. The pattern: the more heterogeneous and market-valuable the work, the more expensive it is to verify — and verification cost is paid by the network, in-band.
The verification tax
Every PoUW design pays a tax to confirm the work happened: redundant re-execution (pay 2–3× for every unit of work), statistical spot-checking (cheaper, probabilistic, gameable at the tails), or cryptographic proof (strongest, and for ML workloads still orders of magnitude more expensive than the work itself). The tax creates gravitational pull back toward uniformity: work that is identical across miners is cheap to score, so scoring functions reward it — and the “useful” work drifts toward benchmark-shaped tasks that exist to be scored rather than used. Holding a scoring function against that pull is the actual engineering problem in this category; we have felt it directly in the inference-mining infrastructure we engineer, where keeping scored work aligned with buyer-shaped work is a design budget line of its own.
The arithmetic
Put the largest live example under the model. Bittensor's first halving (December 2025) cut daily issuance from 7,200 to 3,600 TAO, with the emission split unchanged: 41% miners, 41% validators, 18% subnet owners.
| Quantity | Value | Basis |
|---|---|---|
| Daily emissions | 3,600 TAO | post-halving issuance, public docs |
| Miner share (41%) | ≈1,476 TAO/day | protocol split |
| Miner pay at $300/TAO | ≈$443k/day | price assumption — rerun at spot |
| Open-market equivalent | ≈7,400 H100s, 24/7 | at $2.50/hr rental assumption |
The usefulness question, made concrete: does the network's daily miner output equal what roughly 7,400 continuously-running H100s produce for paying customers? If yes, the emission is a fair wage. If not — and no PoUW network today publishes demand-side revenue anywhere near its emission line — the difference is the usefulness gap, funded by token holders as dilution. That is not an accusation; it is what bootstrapping looks like. It becomes a problem only when the whitepaper calls the subsidy a market.
When the gap closes
Three levers, in ascending order of honesty. Tighten scoring so emission pay tracks work quality — necessary, insufficient; you can pay precisely for work nobody wants. Add demand-side metering — buyer-paid inference, task markets, fee burn — so the network earns external revenue against its emission spend; the ratio of the two is the single most informative health metric a PoUW network can publish, and almost none do. Or accept the subsidy explicitly as a bootstrap with a sunset, and engineer the transition. The networks that survive their halvings will be the ones whose demand leg was real before the emission leg halved under them.
Limits
The $300 TAO price and $2.50/hr H100 rate are stated assumptions — rerun the table at spot before quoting it. Subnet-level economics vary enormously and the aggregate hides winners. And the analysis says nothing about whether decentralized compute is strategically worth a subsidy — reasonable people fund gaps on purpose. The claim is narrower: measure the gap, publish it, and stop calling emissions demand.
If you're working on something similar, we'd like to hear about it. Talk to a founder — 30 minutes.