How the spot market works

Spot is a two-sided market for AI inference. Buyers do not subscribe to a reserved GPU. They bid a maximum price per 1,000 tokens on each POST /v1/chat/completions request. Sellers rest asks on a per-model book with a limit price and a token capacity.

When a bid crosses an ask, the gateway dispatches the buyer's chat request to the winning seller and returns an OpenAI-shaped completion. The fill price is the resting ask (maker) price, never above the bid. Unmatched bids wait until the bid timeout, then fail with bid_timeout / no_match.

Price-time priority decides who fills: cheapest ask first, then earliest rest time. Sellers refresh capacity and TTL independently of the book; capacity reports are informational and do not move resting asks.