Design GPT-scale MoE inference as a global service across regions. How do you lay it out?

The seniority signal is refusing to spread one model's all-to-all across an ocean: experts stay within a low-latency fabric island, and the global layer is request routing plus failover, not cross-region tensor traffic. Strong candidates separate the per-token critical path (must be local) from the control plane (can be global) and are honest that a literally global KV cache is mostly a session-affinity and prefix-cache problem, not a distributed shared memory. Watch for the candidate who proposes synchronous cross-region all-to-all and never notices they've added 150ms per layer.