How AI-Trader works
A continuous limit order book where the traded good is a million output tokens from a named model at a named quantisation.
Put your GPU on the book. Requirements, the bench step, and how to set a floor.
An OpenAI-compatible endpoint with one extra field: your price ceiling.
How we decide a node really produced the tokens it billed for.
What the venue takes, when escrow moves, and how payouts clear.
The unit of trade
A contract on AI-Trader is one million output tokens from a specific model at a specific quantisation, for example llama-3.3-70b-instruct at Q4_K_M. Quant is part of the contract because it changes the output. A buyer who wanted Q8 and silently got Q4 did not get what they paid for, so the two never share a book.
Why an order book and not a price list
Supply here is unusually lumpy. A single operator can add a 4090 at midnight and pull it at 8am when they need it for work. A fixed price list either sits above the clearing price and leaves GPUs idle, or sits below it and leaves buyers unserved. A book lets both sides express a limit and lets the clearing price move on its own.
Matching
Strict price priority, then time priority. The cheapest eligible ask fills first; equal asks are filled oldest-first. Eligibility is physical rather than reputational — a node only sees the queue for models its VRAM can hold and its memory bandwidth can serve at the speed it quoted.
Session uptime is deliberately not a routing input. Ranking by uptime would route every job to whichever node had been connected longest, starve new supply, and reward operators for never patching their machines. It survives only as a tie-break between identical asks.
Settlement
- A buyer's bid escrows
size × limitat post time. Escrow is a hold, not a charge. - When the bid crosses, it fills at the ask. The spread goes to the buyer, not the venue.
- Tokens are metered as they stream. The venue re-tokenises the output and bills its own count.
- On completion the provider is credited and the unused remainder of the escrow is released. A job that dies mid-stream settles only for what was delivered.
Verification runs between steps three and four, and can hold or reverse a credit.