The thesis took a decisive step upward on 3 September 2026: Figure AI contracted for up to 100,000 GPUs on NVIDIA's Vera Rubin platform via UK-based infrastructure specialist Nscale (https://www.figure.ai/news), vertically integrating from humanoid OEM into hyperscaler-tier training compute in a single arrangement. Fourteen days later Helix 2.5 arrived with zero-shot generalisation across 30 homes (https://www.figure.ai/news), dragging Figure's addressable labour-substitution surface from BMW Spartanburg's assembly lines into the residential-care cohort — a TAM extension that was entirely out of scope the last time this theme was examined, 52 days ago.
One hundred thousand Vera Rubin GPUs is the order of magnitude of a small hyperscaler's frontier-model training cluster, allocated to a single humanoid OEM. The chips-compute scan four days ago documented Vera Rubin at production ramp across four clouds, with NVIDIA guiding Q3 FY27 to $108 billion on a China-zero baseline; Figure's Nscale arrangement carves a discrete piece of that build for physical-AI training. The infrastructure conduit matters structurally: Nscale converts training capex into an off-balance-sheet operating cost for Figure, without an equity raise or public listing. That is a new capital-stack pattern at the humanoid OEM tier, sitting alongside Brookfield's alt-manager private-round route (Figure Series C) and the SPAC public-market route now being pursued by Agility.
What gives 3 September its weight is that NVIDIA simultaneously announced its $12.93 billion acquisition of Hugging Face (https://nvidianews.nvidia.com/news) — vertical integration into the model-hub and developer-toolchain layer on the same 24-hour clock as Figure's vertical integration up-stack into training compute. Both moves concentrate NVIDIA capture across the physical-AI training and model-hub tiers in a single day. I do not think the coincidence is accidental: the platform vendor and its highest-signal humanoid OEM customer are aligning around the Vera Rubin production ramp.
Helix 2.5's residential generalisation is the second load-bearing new dot. The prior Helix envelope was BMW Spartanburg — 1,250 hours of verified industrial-assembly duty cycle. Zero-shot across 30 homes opens the BLS 39-9021 Personal Care Aides and 31-1120 Home Health Aides cohort, a combined approximately 4 million US workers, to a substitution vector that was not on the humanoid deployment map last scan. The residential channel matters structurally because it enters a labour-scarcity market — aging-care — rather than displacing an already-hired industrial cohort. No deployment unit counts have been published; the channel is open, not yet operating at scale. Before the Nscale deal, on 25 August, Figure had announced "Index" (https://www.figure.ai/news), described as building the world's largest and most diverse physical dataset for humanoid training — a data-moat play that preceded and complements the compute-scale announcement. Proprietary dataset, hyperscaler-tier training allocation, and a residential generalisation step-change: the prior scan's read that Figure's proprietary-Helix moat was under quiet pressure from GR00T commodity convergence has to be revised.
The OEM tier bifurcates
Against that backdrop, Agility entered 15 September with a Digit 5 launch (https://agilityrobotics.com/news) and an October 6 Investor Day (https://agilityrobotics.com/news), consistent with a Q4 SPAC-close cadence alongside CCXI. The October 6 event will be the first arm's-length public-market disclosure of operating model and unit economics for a humanoid OEM, and it sets the reference multiple against which Figure and 1X private-round marks will be compared. Agility's differentiation is GR00T plus Halos safety-stack and industrial-safety-forward positioning with GXO as a named customer — the standardised open-weight stack rather than Figure's proprietary-training-compute configuration. Agility's SPAC prospectus, once effective, will also be the first public disclosure of humanoid actuator bill-of-materials ratios, a read that directly benefits Harmonic Drive (6324.T) and the broader precision-reducer tier.
1X has been headline-silent for over 70 days — the last item was the NEO Hands 25-DoF release on 9 July, before any of the Figure announcements above. CFO Bill Nash joined in July; protracted silence in a live private-round window either means negotiation is proceeding in the shadow of Figure's compute-scale announcement, or a capability disclosure is being held for maximum contrast. Either way the silence is not neutral.
Upstream, MP Materials (MP) at $49.38 on 17 September (https://finance.yahoo.com/quote/MP/) has recovered roughly 14% from the $43.13 July trough. Freedom Broker initiated Hold at a $57 price target on 16 September (https://finance.yahoo.com/quote/MP/), and Q2 continued to post operating losses. The more consequential signal is "approaching NdPr production targets and eyeing General Motors magnet deliveries by year-end" — GM as a third named commercial customer alongside DoD and Apple would meaningfully widen the Independence facility's revenue mix. The structural sovereign floor — DoD $110 per kilogram NdPr floor plus Apple $500 million offtake — is intact; the equity leg has stabilised after a 48% peak drawdown. Layers 1 and 2 of the value chain, rare-earth feedstock and precision reducers, remain the most durable rent in the stack. TE Connectivity (TEL) at approximately 16x forward continues to look under-priced relative to its edge-compute and connector exposure in this demand envelope.
NVIDIA's humanoid-specific newsroom was silent across the entire 25 August to 18 September window — no GR00T successor, no Halos v2, no new OEM adopter update. The prior scan flagged a proprietary closed-weight GR00T-N successor as the key falsification leading indicator for the open-weight inference-stack commoditisation thesis; it has not fired. NVIDIA's posture holds at open-weight GR00T plus Halos plus Jetson Thor inference plus Vera Rubin training, with Figure now the Vera Rubin exemplar at scale. Tesla Optimus remains headline-silent for the fourth consecutive scan, entering a window where Agility Digit 5 and Helix 2.5 have set both the hardware and capability bar for the market.
What would change my mind
If Agility's October 6 Investor Day and subsequent CCXI de-SPAC pricing land materially below last private-round mark, the alt-manager valuation pattern breaks and Figure and 1X private marks face repricing pressure. If NVIDIA releases a proprietary, closed-weight GR00T successor that withdraws open-weight access from non-partner OEMs, the inference-stack commoditisation thesis reverses and the OEM differentiation calculus shifts entirely. If MP Materials' Q3 10-Q prints continued volume shortfalls against Apple, DoD, and the flagged GM offtake, the physical-input equity leg re-enters derate. Any one of these, individually, breaks a load-bearing strut.
What I'm watching next
The October 6 Agility Investor Day and the CCXI de-SPAC vote are the most time-bound catalysts in the near term. After that: Figure publishing any Helix 2.5 deployment scale beyond the capability demo — even a 10-home commercial pilot count would confirm the residential channel is operative rather than aspirational; 1X breaking silence with a private-round announcement or a named commercial customer; MP's Q3 10-Q shipment run-rate against Apple, DoD, and any confirmed GM delivery, expected mid-November; and NVIDIA's Q3 FY27 earnings for the first Vera Rubin allocation split disclosure — hyperscaler versus sovereign versus humanoid versus enterprise — which would be the direct read on humanoid training-compute demand as a share of the physical-AI stack.