Decide whether a GPT endpoint belongs on Serverless GPU, a GPU Pod, or a VM by comparing traffic shape, latency, and control needs.