Design a shared, external-facing platform for serving large language model inference. The system should support multi-tenancy with per-tenant rate limiting and cost attribution, ensure low-latency responses under high concurrency, and maintain high availability with graceful degradation.