You already know: LLM serving (batching, KV cache), prefix caching, the gateway cases and evaluation gates. This case serves one LLM product from several regions at once. The design questions: does distance matter for LLM latency (less than people think), where may each request go (residency), how to share scarce GPUs across time zones, how to survive losing a region at peak, and how to roll out models across regions.
How to read this lesson. No code, on purpose. Numbers marked assumed are assumptions you state out loud.
1. The prompt, and the questions that change the design
A mid-level answer: "deploy the model in every region, route users to the nearest one, fail over with DNS". A staff candidate adds four observations. Latency: for LLMs, queueing and prefill dominate time to first token, so one cross-region hop is often affordable. Capacity: GPUs, not servers, are the scarce resource, and regional peaks happen at different hours, so a shared pool needs fewer GPUs. Residency: some data may not leave a jurisdiction, and that includes logs, caches and backups, so residency creates islands that cannot share. Failover: a model replica takes minutes to start, but a region fails in seconds, so failover depends on headroom and degradation, not autoscaling.
Assumptions you state: