Embodied AI Hits a Triple Inflection, but Inflection ≠ Winners Decided — Three things converged this fortnight: (1) tech — architecture shifts from modular hand-coded rules to end-to-end unified models (X Square's WALL-A fuses VLA + world model, Zhiping GOVLA, Lingchu's Psi series; BAAI conference June 13 framed the direction as "world models + general physical intelligence"); (2) policy — June 8 MIIT + SASAC launched a "real-scene training" special action for humanoid robots / embodied AI, with the National AI Industry Fund and local state capital already in (Galbot's March ¥2.5B round included the national team); (3) capital/exit — Unitree's STAR Market IPO cleared review in June; 324 deals / ~¥39B in half a year. Judgment: updates ENTRY 65 — physical AI moves from an "underpriced supply-chain constraint" to a triple inflection (end-to-end architecture + national program + STAR exit) arriving at once, but inflection ≠ winners decided. The keyword shifts from "can it walk" (demo) to "can it deliver" (unit economics) — the delivery-phase filter begins (same as ENTRY 72's tiering; Galbot's convenience-store / smart-pharmacy pods already at 100-unit operation). End-to-end VLA + world models move the moat from hardware to the data flywheel — whoever has real manipulation data wins (Lingchu open-sourced 1,000 hrs of human hand-manipulation data to grab the data standard), validating "value is in the data flywheel, not the thin shell." Primary-market read, two layers: leaders (Galbot/Unitree/AgiBot) race on mass production + scene landing + exit channel (Unitree's STAR pass continues ENTRY 88's fifth-set opening, embodied AI is first through the onshore exit), model-data layer (X Square/Zhiping/Lingchu/StarHaiTu) races on VLA + data flywheel. Rollback: if end-to-end VLA underperforms on real industrial generalization or volume stalls at 100-unit and can't reach 10,000-unit, downgrade "inflection arrived" to "inflection on the way."