Economic Model

What the subnet pays, what a miner spends, and — stated openly — what is not yet resolved.

Scope. Emissions, the reign, and the miner cost model are implemented and specified below. Value capture is not designed: there is no gateway, no pricing, and no revenue path. Token and staking mechanics are Bittensor's, not ours.

Emissions: the reign

The reign is a king + equal-share chain: the current king plus the last 4 miners who held the crown and are still registered — up to 5 paid slots total. Whoever is currently paid splits the emissions equally:

Why equal share rather than graduated slots. The reward for climbing isn't a bigger slice of a graduated ladder — it's entering the chain at all, and the only door in is taking the crown, so the gradient lives at the coronation rather than in the slot weights. A dethroned king keeps earning while it decays out of the chain (bumped off the back, oldest first, as newer ex-kings arrive) instead of dropping straight to zero — the anti-hoarding pension tail that gives a beaten miner a reason to keep submitting rather than a cliff to camp against.

Incumbency, and why it decays

Only the king's slot is contested — the ex-king chain fills purely by being dethroned (FIFO) and empties by aging out, not by being out-scored. To take the crown, a challenger must beat the king by a margin ε, not merely tie it — otherwise measurement noise alone would churn the crown every epoch and nobody's revenue would be predictable.

But a permanent margin protects a stale king forever. So ε decays with the king's artifact age:

eps(age) = 0.002 + (0.02 - 0.002) * exp(-age / 8)

A fresh artifact is protected from noise (ε ≈ 0.02); an old one is nearly free to unseat (ε → 0.002). Age is measured per artifact, not per miner — resubmitting the same bundle does not reset the clock.

Exact ties go to the earliest commit block. Copying late never pays.

Burn

Emission is burned to uid 0 only when nobody in the chain is still registered — e.g. a fresh subnet with no king yet, or every current king/ex-king has deregistered. A deregistered ex-king does not burn its own share: the remaining registered members simply split the full emission among themselves at a bigger equal share.

Burn is a feature, not a leak: it prevents the subnet from paying anyone when there is nothing worth paying for.

What a miner spends

Mining is CPU-only. The pool models are remote, so there is no GPU to rent.

The inference bill has a hard ceiling built into the mechanism. Eligibility requires total_cost ≤ budget (default $0.50 per epoch's slice), and an epoch is ~20 minutes, so:

72 epochs/day × $0.50  =  $36/day        <- upper bound, if you spend the entire budget

That is a ceiling, not a target. Under it, spending less buys you nothing.

Cost is a ceiling, not a divisor

An earlier design scored miners on quality / cost. That is broken, and the reason is worth stating: a ratio rewards a miner who answers everything with the cheapest model and scores badly, as long as it scores badly cheaply. It also makes the score unbounded and unstable near zero cost.

So cost became a gate:

• Under the budget, cost does not enter the score at all. Quality wins.
• Over the budget, you are excluded from the reign entirely — not scored zero, excluded (a zero would still occupy a paid slot when fewer than five miners are eligible).
• Separately, a challenger cannot dethrone the king while costing more than king_cost × 1.10.

The result: miners are pushed to maximise quality inside an owner-declared willingness-to-pay, which is what a buyer of routing actually wants.

Per-benchmark floors

acc_b ≥ f_min (default 0.10) on every benchmark. Without this, the cheapest way to satisfy a cost budget is to abandon the hardest benchmark entirely. The floor makes benchmark-abandonment a disqualification (below_floor:<b>), not a strategy.

What is not designed

Publishing an "Economic Model" page that omits this would be dishonest.

• There is no revenue path. The gateway does not exist; there is no pricing, no payments, and no answer to who operates it. See Using the Gateway.
• Token mechanics are Bittensor's. Alpha, staking, and emission schedules are subnet-level Bittensor concerns. We have not designed anything on top of them.
• The routing moat is absent on general traffic. Our own measurements — five independent reads — find that an achievable router does not beat the best single model on accuracy, and the best single model is often the cheapest. The defensible value is cost at intermediate quality targets, not beating the frontier. Read Findings.
• The subnet-specific case is thinner than the product case. Even where routing value is real, a tiny trained router captured most of it on held-out data — which cuts both ways. Value that is easy to capture is value that open competition saturates quickly, weakening the argument for an ongoing incentive mechanism.

We are documenting a mechanism that works, not asserting a market that does. If the economics gate passes on a real stream of target traffic, this page gets a revenue section. Until then it does not.

See also

• The KOTH Protocol — scoring, eligibility, the dethrone guard
• Findings — the measurements behind the caveats above
• Bittensor Integration — weights, burn, and the reign on-chain