The thesis on humanoid robotics moved materially today: Figure AI has published the first factory-grade yield curve for a humanoid OEM, and signed its first apparel-logistics customer — developments that shift the debate from "can these things be built?" to "are they being built at industrial scale?" That shift matters now because it gives the market a primary-disclosure run-rate to mark against, rather than management projections.
From demos to yield curves
The manufacturing disclosure is the more significant of the two. Figure's BotQ line increased throughput 24-fold over 120 days — from one unit per day to one unit per hour — with more than 350 third-generation units delivered and an end-of-line first-pass yield above 80%, the battery line running at 99.3% (https://www.figure.ai/news/ramping-figure-03-production). For context, automotive end-of-line FPY benchmarks typically sit between 85% and 95%; Figure 03 is converging on that range inside four months, beginning from what was effectively artisan assembly. The metric is primary-disclosed and therefore markable; the annualised run-rate is mine: a 2,000-hour single-shift year implies roughly 2,000 units annually, 24/7 operation implies roughly 8,800 — which brackets the company's previously cited 12,000-unit scaling target but makes clear that target requires multi-shift operation.
This matters for the broader OEM bucket. Tesla carries a ~176x forward multiple and maximum analyst coverage; Hexagon AB (HXGBY), whose AEON robot has an industrial customer at BMW Leipzig, has neither a published yield disclosure nor a throughput metric. 1X, Agility, and Unitree are further back still. Figure's yield curve draws a line between a nascent "demonstrated-manufacturable" sub-bucket and the larger "demo-only" sub-bucket. Only the former can begin underwriting supplier contracts.
The second commercial vertical
The Catalyst Brands agreement (https://www.figure.ai/news/figure-signs-agreement-with-catalyst-brands) adds something the BMW Spartanburg pilot — 1,250 operating hours over roughly ten months — did not: a second demand vertical, and a harder one. Apparel distribution centres involve variable SKUs, soft goods, and sortation — dexterity-heavy work historically considered more demanding than rigid-product warehouse flows. That Figure chose this as its second public customer rather than a safer rigid-goods 3PL signals confidence in Helix-02's generalisation and aligns with its May capability demonstrations of autonomous household tidying tasks. Catalyst's Reno, Nevada facility serves JCPenney, Aéropostale, and Brooks Brothers, with potential expansion across the portfolio.
Volume and price terms are undisclosed. What is disclosed — and what I find more structurally interesting — is the investor overlap: Brookfield is named as a shared investor in both Figure (Series C, September 2025) and Catalyst Brands. An alternative-asset balance sheet is sponsoring both the robot supplier and the deployment demand-sink in the same theme. This is not an incidental coincidence; it is a textbook capital-flow mechanism where the marginal funder of the deployment pipeline is decoupled from arms-length price discovery. The effect is that theme-internal valuations can detach from underwritten economics longer than a purely public-market theme would allow — a point that compounds the valuation strain already visible at Tesla (~176x forward) and Harmonic Drive Systems (6324.T, ~100x forward).
The value chain: what changes and what doesn't
The enduring-economics read holds. Layers one and two — rare-earth magnets and precision reducers — remain the most durably profitable over a three-to-five-year horizon because substitution requires multi-year capex and metallurgical qualification. MP Materials (MP, +199% over fifty-two weeks) and Harmonic Drive Systems (6324.T, +134%) are the right names at the binding bottleneck; the problem is that both are already richly priced. Today's Catalyst deal adds a third demand vector into US distribution centres for MP specifically: the Reno facility is domestic, deepening the case for Independence TX magnet capacity alongside the existing EV and defence vectors.
At the capability layer, NVIDIA's Jetson Thor and Isaac GR00T stack (https://developer.nvidia.com/isaac/gr00t) remains the "Android of robotics" for the long tail of OEMs — but Figure runs Helix, its own vision-language-action model, not GR00T. The Android frame applies cleanly to 1X, Agility, Hexagon, and Unitree; it is contested at the one OEM that has so far demonstrated process-industrial repeatability. Whether Helix-style proprietary stacks are an outlier or a template is a question the next chips-compute scan needs to answer, because the durable-capture case for NVIDIA turns on it.
The picks-and-shovels case for Tokyo Electron (TEL) at roughly 16x forward is unchanged: Catalyst's apparel-DC deployment pulls force-torque sensing and connector content into a new vertical without changing the supplier list. On the downstream-beneficiary side, Symbotic (SYM) remains the cleanest public proxy for rigid-goods warehouse automation; the apparel-distribution analogue is more fragmented, and if Catalyst's Reno deployment scales the public exposure may route through material-handling suppliers such as Honeywell Intelligrated or Dematic (via Kion) rather than a pure-play OEM.
What would change my mind
The yield curve is the new load-bearing fact, and it has a clear falsification: if Figure BotQ throughput stalls below roughly one unit per two hours by Q3 2026, the process-industrial-repeatability thesis is broken. Separately, if the Catalyst Reno deployment fails to enter sustained operation by Q4 2026, the second-vertical story collapses back to a single auto-industry pilot. On Tesla, if Optimus production prints below ~1,000 units or unit cost stays above $50,000 by Q4 2026, the scale-optionality premium built into a ~176x multiple becomes very hard to defend. Any of these, individually, would force a material downgrade of the 2026 base case.
What I'm watching next
The NVIDIA GR00T N2 / DreamZero release date remains unverified — the prior framing of "previewed for end of 2026" should be treated as speculating until NVIDIA dates it. MP Materials' magnetics segment ramp and Suzhou Green Harmonic capacity and pricing disclosures are still unverified because SEC EDGAR direct browse is returning 403 errors; those need to resolve in the next scan. The Brookfield capital-stack pattern is worth testing across the autonomous-mobility theme — Waymo and Cruise capital-stack mapping would confirm whether the supplier-and-buyer sponsorship template is idiosyncratic to Figure or systemic across the physical-AI universe. And I need the BLS OES 53-7000 material-moving series to anchor the labour-substitution demand model for apparel-DC work.