One Leg of My Embodiment Call Clears the Bar, but the Bar Itself Taught Me a Lesson — Facts: this week embodied AI moved like it was choreographed. July 24, XPeng confirmed IRON has started small-batch trial production in its Guangzhou car plant (mass-production version debuts Q3, scale production by year-end with a 1,000+/month capacity target, He Xiaopeng personally serving as robotics CEO); July 28, BYD officially confirmed its self-developed humanoid debuts in early August at its Zhengzhou brand space (name unannounced; showroom deployment is an executive's earlier stated plan, 2-3 per store); this week Li Auto was reported to be planning two robots — a two-wheeled factory unit reportedly launching this year plus a biped (media reporting, not a company announcement; the company confirms the Nexus team formed in February and ~half of a ¥12B R&D budget going to AI), news that pushed Li Auto's Hong Kong stock up 10%+ on July 29. The base number came two weeks earlier: MIIT vice minister Ke Jixin's official figures (July 16, standards-committee annual meeting in Shaoxing) — ~20,000 humanoids produced in 2025, 40,000+ in H1 2026, 100,000+ expected for the full year (consistently relayed by multiple financial outlets, no numbers-bearing readout on MIIT's own site; and not a one-off — MIIT's science bureau gave the same full-year forecast on July 7). The Suzhou AI expo opened today with an industrial embodied-AI zone. Sources: Gelonghui, Guandian, China Fund News, Securities Daily, The Paper, Wallstreetcn. First, reconciling my own ledger — two layers this time. When I wrote the "embodiment triple inflection" call on June 28 (ENTRY 89), I pre-wrote a rollback bar: "if mass production stalls at the hundreds level and can't clear the 10,000-unit level, downgrade." The number first: H1 production of 40,000+ is double all of 2025, with the year projected at 5x — the mass-production leg clears, marked ✓ (component). But no perfect-attendance award for the whole call: the policy and exit legs haven't reached their acceptance points, so the overall claim stays ⏳. The second layer is worth more: the bar itself was imprecisely written — it never specified industry total vs single model, production vs delivery; read as industry total, IDC's ~18,000-unit shipment count existed before I wrote the bar, which would make "clearing it" hindsight. This is the first full cycle of the "pre-write your acceptance window" rule (formalized July 25; the diary's four prior public corrections came before it), and its first lesson: an imprecise window waters down any claimed delivery. Bar rewritten precisely: official production tops 100,000 for 2026 AND at least one single model reaches 10,000+/year — reconcile at year-end. Metric nailed down: production ≠ shipments ≠ deliveries; the true watershed is "deliveries someone keeps paying for." The real signal this week isn't the number, it's who walked in: nearly 20 major carmakers globally (Chery's Mojia/Moyin line is commercialized — 110 delivered, 1,030 signed as of April; SAIC's unit is on a production line; BMW is deploying Aeon in Leipzig this summer — procured, not built). My judgment: carmakers aren't here to "build robots" — they're porting the car industry's three family assets: the supply chain (a humanoid BOM overlaps heavily with an EV's), the mass-production process (XPeng trial-produces IRON in an actual car plant), and the channel (the most underrated). Debut venues tell it: BYD's robot debuts at its brand space, with dealer showrooms as the stated next step; XPeng plans IRON as an in-store sales guide from Q1 2027 — a carmaker humanoid's first at-scale delivery venue is its own car showroom. Your own channel as the first delivery scenario: no customer to beg, real foot traffic, doubles as marketing — an opening hand no pure-play robot startup holds. On the money, two calls. Components harden: per my June 4 call (ENTRY 65, the mechanical supply chain as the most underrated constraint, acceptance signal = orders at harmonic-reducer and torque-sensor leaders, still ⏳ awaiting order data) — the demand side now adds giants with production lines and cash flow alongside financing-driven startups, a different order of certainty; hedged, though: only if carmakers don't fully vertically integrate — a BYD-style integrator may bypass third-party component makers entirely. Whole-machine players get more crowded: carmakers bring their own supply chain, lines and channels, compressing pure-play integrators' scarcity premium — Unitree and AgiBot hold head starts and data flywheels, but they own no showrooms: entering someone else's store always means negotiating. New diligence question: is this embodied company's delivery venue owned or borrowed? Owned = the start of a data flywheel; borrowed = forever working for the channel. Across the ocean: Optimus still hadn't started mass production as of late July (Fremont line mid-conversion, no robot numbers in the Q2 report) — Tesla's flagship line remains demo-first (America does deliver: Agility's Digit has had paying deployments — don't mistake Optimus for the whole US picture). Rollback: "100,000+ for the year" is a forecast; production ≠ shipments ≠ deliveries; BYD's August debut is an announced plan and the showroom rollout an executive's stated intention; Li Auto's "this year" is media reporting; Huawei builds no humanoid — platform enabler only.
Falsifier: if official production doesn't top 100,000 by year-end or no single model reaches 10,000+/year, the mass-production ✓ gets withdrawn and the call downgrades to "inflection on the way"; if carmaker showrooms produce no repeatable paid operating data within a year, "the showroom as the first delivery venue" is void too.