Skip to main content

The operator network

ZeroSignal doesn't run the AI models you talk to. They're served by operators — independent parties who run inference hardware and sell its output on an open network. This page explains who they are, how they get paid, and why the system works without you having to trust them.

If you're here to run a node rather than understand the network as a user, start at What is an operator.

What an operator is​

An operator runs the node software in front of an AI backend and registers on-chain: their payout address, the public key their node signs with, the endpoint to reach it, and optionally a human-readable name. The models they serve and the prices they charge are advertised by their node and shown live in the model picker.

There's no application process and no gatekeeper — the registry is a smart contract. Anyone can register an operator, and any user can transact with any operator. That's the "open rails" property described in Why use crypto?.

How a request is paid​

Payments are designed so neither side has to trust the other:

Quote​

Before sending, your client asks the operator for a signed ticket stating the maximum price for the request. The operator can't raise the price after the fact.

Escrow​

That maximum is locked in the escrow contract on-chain — the operator can see the funds are real, but can't take them yet.

Receipt​

After the response streams back, the operator signs a receipt stating the actual tokens used and the actual amount charged.

Verification​

Your client checks the receipt before settlement: the signature must match the operator's on-chain key, the amount must not exceed the ticket's maximum, and a hash in the receipt must match the response you actually received.

Settlement​

The escrow pays the operator their charge in full, sends the protocol fee (which you paid on top, not skimmed from the operator) to the treasury, and refunds the rest to you. The result is final on-chain.

This is why the per-message cost in the app isn't an estimate — it's a verified, operator-signed amount with a hard ceiling you saw up front.

What you trust, and what you verify​

Verified by your client, automatically:

  • The price ceiling — an operator can never charge more than the ticket said.
  • The receipt signature — charges come from the operator's registered key.
  • Response integrity — the receipt is bound to the bytes you received.
  • The encryption keys — the key your prompt is sealed to is signed by the operator's on-chain identity, so a relay can't substitute its own.

Trusted, with skin in the game:

  • The operator sees your decrypted prompt while running the model — that's inherent to inference. What it doesn't see is who you are: requests arrive through a relay (below) with no account or identity attached.
  • Advertised prices and latency are self-reported, but the receipt mechanism means dishonest pricing can't actually charge you more than quoted.

Relays​

Every operator node also acts as a relay for other operators. Your request travels to its target through one relay hop:

  • The relay can't read anything — the envelope is sealed to the target operator's key.
  • The target operator never sees your IP address — it sees the relay's.
  • A fresh relay is picked for every request, and a relay owned by the same party as the target is never used — it's a hard rule, so no single operator ever sees both who you are and what you asked. (The relay must also speak a compatible protocol version and, best-effort, sit on a different network — a different /16 subnet — than the target. Among the eligible relays the system prefers ones it has measured as reachable, but falls back to the full eligible pool when none are known-reachable yet, so a cold relay still gets tried.)
info

If no eligible relay is available for a target, the app refuses to send directly — it fails closed rather than expose your address.

Health and compatibility​

The app continuously probes every registered operator: whether it's reachable, what it serves, and whether it speaks a compatible protocol version. Unreachable or incompatible operators are flagged in the picker and skipped by Auto mode, so requests only go to nodes that can actually answer them.

Running an operator​

The network is open: if you run AI hardware, you can sell inference on it by running a node and registering on-chain. See What is an operator to get started.