Edge silicon runs roughly 30‑50× below datacenter memory bandwidth, and device‑RAM ceilings cap what models fit, so frontier inference stays in metro datacenters. Verified framing: latency, not cost, sets the split — real‑time needs low‑round‑trip placement, meaning a datacenter closer to users, not a chip in your pocket. The 'train in cloud, infer at edge' thesis failed three tests. Industry reporting calls the AI‑PC/phone supercycle dead on arrival; the edge‑silicon leader's own filings show core handset chips with datacenter ambitions early and unproven. Edge AI is real. Edge relief is not.