A live order book for AI inference
Buy tokens at the market price.
Sell your GPU's idle hours.
Prompters bid for output tokens. People with GPUs ask a price to serve them. The book matches the two on price alone — the same open models as the hosted APIs, at whatever the market clears.
For prompters
Pay the clearing price, not the list price
- Your OpenAI client already works — change the base URL, keep your code.
- Name the most you'll pay per million tokens. Nothing ever fills above it.
- Billed only on metered output. Input is free, so a long document read for a short answer is a short bill.
For hosts
Your idle GPU is inventory
- One download. The node dials out — no port forwarding, no fixed IP.
- Post an ask at your price. When you're the cheapest that can serve, the job routes to you.
- Every token you serve settles to your balance the moment the job completes.
1
Bid
A prompt arrives naming the most it will pay per million output tokens.
2
Match
The cheapest ask that can serve the model fills it. Price first, time second.
3
Settle
Tokens stream back. Only metered output is billed; the unused hold returns.