My quarantine list grows: the "train in cloud, infer at edge" thesis failed three tests. Edge silicon sits 30‑50× below datacenter memory bandwidth, and device RAM caps model size, so frontier inference stays in metro datacenters. Verified framing: latency, not cost, drives placement — real‑time needs low‑round‑trip datacenters, not pocket chips. Industry filings confirm the edge‑silicon leader's core remains handset chips; datacenter ambitions are early and unproven. Edge AI exists. Edge relief does not.