Pilot, v0

An exchange for LLM inference.

Suppliers post asks for a graded model. Buyers send ordinary OpenAI requests. Every request goes to the best eligible ask.

Sell capacity Buy inference
Live simulation grade-a/deepseek-v4.1-flash-1017
–buyers pay under their max price, on average
$0.000earned by suppliers, each at its full ask
0.12 msadded by matching, p50
Each request is matched to the best eligible ask. Illustrative prices, simplified matching rules. The 0.12 ms is measured.

One grade, one spec

Every offer meets the same spec, so asks are comparable.

OpenAI-compatible

Buyers change the base URL. Nothing else.

Every trade recorded

Each request settles at the supplier's ask.

For suppliers

Sell spare capacity at your price.

You control price and volume, minute by minute.

$0.325+ fee
ask + fee
Buyerpays
Exchangekeeps the fee
Suppliergets the full ask

One payment: the supplier's ask, with the fee on top.The supplier receives $0.325. Nothing is deducted.

For buyers

Keep your client. Change the base URL.

Set your terms once. The exchange finds the supplier.

Your existing code
client = OpenAI(
  base_url="https://api.openai.com/v1","https://<exchange>/v1",
  api_key=OPENAI_KEY,IX_KEY,
)
client.chat.completions.create(
  model="your-model","grade-a/…",
  messages=[…], stream=True  unchanged
)

Matching

Every request, five checks.

  1. Live asksfor the grade
  2. Within bidafter the fee
  3. Fast enoughmeasured first token
  4. Has capacityand is healthy
  5. Best by policythen failover in order
0.12 msadded latency, p50
0.30 msadded latency, p99
2 fieldsrewritten per request: model and API key

Latency measured on localhost against a zero-latency mock supplier, 3,000 sequential requests. Real deployments add network hops.

Status

Works today

  • One grade with a live order book and quote API.
  • Chat Completions, streaming, failover.
  • Per-trade ledger. Tested against mock suppliers; live supplier tests are next.

Next

  • Settlement and invoicing.
  • Richer grade specs and quality checks.
  • Consoles, keys and rate limits.

Pilot onboarding is by invitation.

Reply in the thread this link came from. We will send a quote token or a buyer key.