Skip to main content

How the network works

As an operator you sell inference into an open market. This page explains how you get discovered, how you get paid, and what the protocol does so that neither you nor your users have to trust each other.

Discovery: you advertise, clients choose​

Your node publishes a public details document describing the models it serves, their prices, and its capabilities (see Serving models & pricing). Users' clients read it directly, alongside your on-chain record, and build a live catalog. There's no central index that can delist you and no ranking you have to pay into — clients discover every registered node and decide among them on price, capability, and measured performance.

Because the same model is usually served by several operators, you compete on price and on performance. The network measures two things per node and keeps them on-chain as rolling averages: time to first token and decode throughput (tokens per second). Faster, cheaper nodes get picked more often, including by users' Auto mode.

Payment: paid per request, with a price you set​

Every request pays its own way. The flow is built so the price is agreed before any work happens and the charge is verifiable after:

  1. A client asks your node to reserve — your node returns a signed ticket stating the maximum price for the request. You can't be made to serve below your price, and you can't charge above the ticket afterward.
  2. That maximum is escrowed on-chain. You can see the funds are real before you spend a GPU-second on the request.
  3. You run the model and stream back the reply, then issue a signed receipt stating the actual tokens used and the actual charge.
  4. On settlement the escrow pays you your charge, refunds the rest to the user, and takes the protocol fee.

The full mechanism — including the two-phase settle and the dispute path — is in The payment flow.

Trust: secured by signatures and stake, not by reputation alone​

You and your users are anonymous to each other. The protocol replaces trust with three things:

  • Signatures. Your ticket and your receipt are signed by your node's on-chain key. A client verifies every receipt against that key, checks the charge against the ticket's ceiling, and checks that the receipt is bound to the exact bytes it received. A forged or inflated charge is rejected before it settles.
  • Escrow. Funds are locked before you do the work and released by a smart contract, not by either party's goodwill. If the two sides disagree at settlement, the ticket freezes for off-chain arbitration rather than paying out wrongly.
  • Stake. Registering requires locking a USDC stake. It's the cost of a network identity and the thing that's at risk if you misbehave — see Staking & economics.

Privacy: you serve prompts, not people​

You see the prompt you're asked to run — that's inherent to inference. What you don't see is who sent it. Requests reach you through a relay (another operator's node), so you see the relay's address, not the user's, and there's no account or identity attached to the request.

In return, your node relays for others. Every operator is also a relay, and a fresh relay is chosen for each request — so no single operator ever sees both who a user is and what they asked. Running a relay well is part of being a good network citizen; see Relays.

What you control​

  • Which models you serve and under what content policy.
  • Your prices, per model and per route (text, images, in-loop image tools).
  • How many nodes you run, and which are staging vs. production.
  • When you're available — take nodes down freely; the network routes around unreachable nodes and probes for your return.

What the protocol enforces​

  • You can't charge above the ticket you signed.
  • You can't be paid for a reply a client can't verify it received.
  • You can't read a request that wasn't sealed to your key (so you can't snoop as a relay).
  • You can't register or serve without a stake at risk.

Ready to try it? Quick start takes you from nothing to a registered node serving its first paid request — then the concept pages explain each of these guarantees in the detail you'll need to run a node that behaves correctly under all of them.