Pay-per-second AI inference

How a model becomes a paid endpoint

The reference provider (@arl/provider) turns any command into a paid job, for example a container that loads your model and reads the prompt from standard input. It lists the jobs with their price per second and the longest run, and it never lets a client pass its own command.

Why it fits inference

• Short requests are billed fairly: a two-second generation costs two seconds.
• Payment is authorized before the GPU starts, and settlement happens before the output is released.
• A signed usage receipt with a hash of the output lets the client prove what it paid for.
• Clients rate the provider on the ERC-8004 registry after each paid job, so good service builds a public record.

How it works

1. Define a job that runs your model, for example docker run --gpus all --network none <image>.
2. Set the price per second and the longest run, and list the service in the ARL app.
3. Clients call POST /jobs/<name>; your node answers 402 until a signed ceiling arrives.
4. The model runs for the paid time; the seconds used are settled in ARL and returned with the output.

Questions