Mining alone has to beat the cost of the card out of block rewards, and on rented hardware we measured that it does not. Serving AI inference is paid per request, at rates set by demand rather than by a network's difficulty — so the card earns from paid work and mines only in the seconds nobody is using it.
That is the whole thesis of this pool: inference is the revenue, mining is what keeps the card from idling.
Mining to this pool is fully live — everything in the other guides works right now and pays in PRL.
Serving is not open self-service yet. The node software, the earnings accounting and the 75% operator share all exist and run, but we are onboarding operators a few at a time rather than opening a form, and the USDC payout rail is not armed: serving earnings accrue against your address, and no automatic send has been switched on. We would rather you read that here than discover it after wiring up a rig.
If you operate GPUs — one rig or a data centre — and want to be in the next group, start at the connect wizard, which carries the current request-access path and the honest state of each tier.
In the meantime, the cards you have can mine today. That part needs no permission from anyone.
Which cards can do both at once. It is a question of spare memory, not raw speed: a 32 GB card can hold a model and still have room to mine in the gaps, while a 24 GB card running the same model has nothing left over and simply serves. Smaller or older cards cannot serve at all and keep mining as they do today.
The split: an operator keeps 75% of what their card bills for inference. That rate is in the code that does the accounting, not just in a document.
Why access is gated rather than open. A serving node answers real customer traffic, so a node that is wrong is a customer's problem, not just ours. We would rather bring operators on slowly enough to verify each one than publish a one-click installer and hope.
What we will not claim. We have published revenue-per-card figures for serving before that rested on an assumed utilisation rate we had never measured. We are not repeating that here: there is no $/day figure on this page, and there will not be one until we can source it to measured demand.
The 75% operator share and the accrual behaviour are in the live earnings and payout code on this pool; the payout engine's own state is dry-run, which is why no sending is claimed above. Tier availability mirrors the connect wizard and the private-beta label on the home page, read 2026-11-03.