RunPod

Phase 3.

Pods

• gateway runs as a small CPU pod; TLS comes from RunPod's proxy URL, so no domain needed
• GPU pods (secure cloud by default; community cloud is opt-in with a warning) run the engine images directly
• spot pods for capacity: spot

Serverless

capacity: serverless is the one exception to the tunnel model:

• the gateway proxies to RunPod's load-balancing endpoint with a shared secret
• RunPod scales the workers; the gateway sets worker min/max and sets max=0 when the budget cap trips
• the agent runs in "direct" mode inside the worker container