On this page
Capacity listing walkthrough
This page takes a node operator from a known starting configuration to a resting ask. Pricing is ./pricing. Bid-market positioning is ./positioning. Endpoint mechanics stay on the existing seller pages; this walkthrough only covers the operator judgements those pages omit.
Starting configuration
| input | value |
|---|---|
| Hardware | 2× NVIDIA A100 80GB |
| Region | us-west-2 (operator-managed; not an API field) |
| Weekly uptime target | 99% (operator-managed) |
| Models | gpt-4 |
| Availability window | weekdays 18:00–23:00 PT |
| Sustained decode | 160 tokens/s across both GPUs |
| Tokens per ask | 50,000 |
| Ask TTL | 300s (gateway default) |
| Refresh cadence | every 240s inside the window |
Step 1 — Register the seller
Mechanics: Seller quickstart and POST /v1/sellers/register.
Map the archetype onto the register body:
| field | value | why |
|---|---|---|
name | west-a100-pair | Human label for this box |
endpoint_url | https://infer.west.example/v1 | Idempotency key — re-registering this URL updates the same seller |
supported_models | ["gpt-4"] | Only models this hardware will serve |
api_key | operator-chosen seller key | Presented later as X-Seller-API-Key |
Step 2 — Define capacity
capacity on POST /v1/sellers/{id}/asks is tokens offered on that ask, not GPUs and not hours.
At 160 tokens/s, 50,000 tokens is about 5 minutes of decode — one TTL window with slack before expiry. model_id must be in supported_models. Leave price as a placeholder and set it from ./pricing. ttl_seconds defaults to 300.
Step 3 — Declare constraints honestly
Platform-enforced
- Model support:
supported_modelsat register,model_idat ask. - Latency SLA: first token within 2000ms + 100ms/token. Three breaches in 60 seconds trip the circuit breaker. See How SLA enforcement works.
Operator-managed
Uptime target, geographic region, and hardware spec are not API fields. They change how you size capacity, how often you refresh ttl_seconds, and whether you rest an ask at all. Do not invent region/hardware/uptime keys on the wire.
Step 4 — Set availability windows
Two knobs: ttl_seconds on submit_ask (default 300) and whether you submit another ask. For this archetype, rest every 240s during 18:00–23:00 PT and do not rest outside that window. The mechanical refresh/cancel pattern is in Ask lifecycle.
Step 5 — Report capacity and observe fills
Capacity reporting is informational — available_tokens / queued_tokens do not move the book. Trade history polling is how you see fills on GET /v1/sellers/{id}/trades.
Fully worked example
Values below come from the starting-configuration table. price is unset until the pricing guide.
- Register
west-a100-pairathttps://infer.west.example/v1withsupported_models: ["gpt-4"]. - At 18:00 PT, submit an ask:
model_id=gpt-4,capacity=50000,ttl_seconds=300,pricefrom ./pricing. - Every 240s while still inside the window, submit a replacement ask (or refresh via the ask lifecycle pattern).
- Report
available_tokens/queued_tokensif you operate a local queue; poll trades after fills. - At 23:00 PT, stop refreshing. Let the last ask expire or delete it.
What is next
- Pricing and margin — replace the
priceplaceholder - Competitive positioning — whether that ask will get hit