Private GPU private silicon

A private GPU is an on-demand, dedicated LLM host that is yours for the duration of a session (a “burst”). You start it through the API, infer on it exactly like any OpenAI-compatible model, and stop it when you’re done — paying only for the time it runs.

How it relates to your org & user (console)

Private GPUs are per user, billed to the user’s org. Configuration happens in the console before you can start anything:

• Offerings — staff enable GPU offerings for your org (hardware tier, hourly rate, optional max duration). You can only start an active offering your org has. Staff manage this on the console GPU page.
• Entitlements — your org grants the offering to a group or a specific user. If you’re not entitled, POST /v1/gpu/bursts returns 403.
• One live burst per user — you can’t start a second burst until the first is stopped (409 conflict).

Your API key does the rest — the same rk-… key you use for public models authenticates all private-GPU endpoints and ties the burst to you + your org.

Endpoints

1. Start a burst

POST /v1/gpu/bursts with an offering id (a UUID your org is entitled to) and a duration in minutes. You will already have the offering id from staff; on the console it is shown on the GPU page.

curl -s -X POST "$SELIGO_BASE/gpu/bursts" \
  -H "authorization: Bearer rk-…" \
  -H "content-type: application/json" \
  -d '{"offering_id":"<offering-uuid>","duration_min":60}'

A successful start returns 201 with your new burst (including its id and the session slug):

{"burst":{"id":"<burst-uuid>","slug":"burst-…","status":"pending",
            "auto_end_at":"2026-09-15T18:00:00Z", …}}

2. Wait for it to be ready

A cold GPU takes several minutes to boot (a warm boot is faster). Poll your bursts:

curl -s "$SELIGO_BASE/gpu/bursts" -H "authorization: Bearer rk-…"

The status moves through pending / provisioning / booting_warm → ready (then draining / ended after stop). Once ready, GET /v1/models lists the stable seligo/gpu alias for your key:

curl -s "$SELIGO_BASE/models" -H "authorization: Bearer rk-…" \
  | grep -i "seligo/gpu"

3. Infer on the GPU

Use the stable alias seligo/gpu as the model. It always resolves to your live, ready burst — so your client code never needs to know the per-session endpoint id.

curl -s "$SELIGO_BASE/chat/completions" \
  -H "authorization: Bearer rk-…" \
  -H "content-type: application/json" \
  -d '{"model":"seligo/gpu",
       "messages":[{"role":"user","content":"Analyse this data residency question"}]}'

While the burst is still booting, calls to seligo/gpu return 409 with code gpu_booting. If no burst is live for your key they return 404 gpu_not_started — start one first (step 1).

4. Stop the burst

Stop when done so you stop paying:

curl -s -X POST "$SELIGO_BASE/gpu/bursts/<burst-uuid>/stop" \
  -H "authorization: Bearer rk-…"

The burst drains (finishes in-flight work) then the GPU is torn down and deregistered. A refund is settled for the unused reserved time.

Chat UI equivalent

If you use the Seligo chat interface in the console, it drives the same machinery: the picker shows a purple Private GPU row when you have a live burst, and chat turns are relayed to your GPU through the same router. The console GPU page shows offerings, entitlements, your live burst (uptime + auto-stop), and lets you start and stop it without writing any curl.