A market, not a price list

Every few seconds the scheduler runs a clearing auction across the candidate set for each pending job. Operators have standing asks; jobs carry a bid ceiling. The clearing price is what the marginal accepted operator asked for, so a job frequently pays less than its ceiling — the average job clears at 71% of the price it was willing to pay.

The spread between the cheapest and most expensive hour in a week is wide, and it is wide in a way that is highly predictable. This is what makes ceilings and schedules such effective instruments: most of the volatility is diurnal, not random.

Nobody wants a variable bill for a production system. The instruments below all do the same economic thing — move price risk from the customer to the operator or to ROBO — and they are priced accordingly.

The interesting finding is how little most teams buy. Reserved capacity is 11% of spend even though almost every customer has a production workload. Teams turn out to be far more tolerant of price variance on batch and training work than they expect to be, once they can see the diurnal pattern.

Paying strangers for compute is a trust problem in both directions. The buyer does not want to pay for work that never ran; the operator does not want to run twelve hours of somebody's training job and then chase an invoice. ROBO settles on an EVM chain specifically to remove both halves of that problem.

Submitting a job escrows its maximum budget in a contract. The scheduler will not place work that is not funded. When the job finishes, the operator submits the signed execution record — image hash, accelerator-seconds, attestation — and the contract releases $ROBO against it, refunding whatever the job did not spend. Nobody has to trust anybody; they have to agree on a record.

Settlements are batched. Writing a transaction per job would make small inference calls uneconomic, so a rollup aggregates a settlement epoch and posts one proof. That is what keeps the on-chain cost per job in the fractions of a cent rather than dominating the bill.

There are three ways to be on the earning side of this network, and they are frequently conflated. They carry different work, different capital and very different risk.

Operating is a hardware business: you supply accelerators and get paid per verified accelerator-second, and your bond is slashed if you vanish mid-job. Staking is underwriting: you back an operator's bond with $ROBO and share their fee stream, which means you also share their slashing. Verifying is labour: you re-run sampled work and get paid from the settlement fee, more for catching a mismatch than for confirming one.

The honest framing is that only the first is a compute business. The other two are ways of pricing somebody else's reliability, and they pay accordingly.

An operator's decision to attach hardware is a straightforward comparison against the alternative, which is usually leaving it idle. What makes it work is not the peak price — it is that the network finds paying work for hours that would otherwise have earned nothing.

The median operator earns most of their revenue in hours when the clearing price is below the weekly median, because that is when their hardware would otherwise be dark. This is the single most counter-intuitive number in ROBO's economics, and it is why render queues matter so much.

Bootstrapping a two-sided compute market means paying operators to show up before there is enough demand to pay them properly. ROBO did this, it worked, and it is the part of the model most likely to be misread as durable economics.

Subsidy peaked at 34% of operator revenue in early 2025 and is now under 8%. The schedule is public and continues to zero. Any assessment of whether this network works should be made against the unsubsidised number, which is why we publish both.

Every job settles against a signed record of what ran: the image hash, the accelerator-seconds consumed, the attestation, and the clearing price at match time. Either side can dispute within 72 hours.

Disputes affect 0.4% of settled jobs. Seventy-one percent resolve without a human, because the disagreement is almost always about billable duration — a node that hung, a drain that was slow — and the telemetry answers it directly. The remainder go to an operator council, and ROBO's own losses are paid from the reliability insurance pool rather than from the operator.