Design an inference serving system for a large language model (e.g., Claude) that minimizes latency and maximizes throughput across a GPU cluster. The system sits between client-facing API services and a pool of GPU inference backends. Key components to address include: (1) a router/load balancer between the API service and…