The NVIDIA open-weight release of GR00T N1.7-3B on 7 July and Figure's delivery of F.03 to BMW on 30 June arrived within the same week and pull in opposite directions: the first is the most concrete commoditisation move the capability layer has yet seen; the second is the most concrete deployment-scale signal the physical-embodiment layer has produced. Together they sharpen the thesis that durable rent in humanoid robotics lives below the model, not above it.
The capability floor just became a commodity
NVIDIA published GR00T N1.7 — a 3-billion-parameter vision-language-action foundation model with a Cosmos-Reason2-2B backbone, trained on 21.6 million data points across 13 datasets (roughly 32,000 hours of real demonstration and 8,000 hours of simulated rollouts) — as an open-weight release under the NVIDIA Open Model License on Hugging Face (model card; developer blog). The benchmark improvements are real: DROID-F6 +61% over the prior generation, end-to-end latency of 27.9 ms on H100 with TensorRT, deployable via Isaac ROS to Jetson Thor. Twelve humanoid OEMs are disclosed adopters — 1X, Agility, ANYBotics and Bellboy among them. This is the first dated disclosure that the "Android of robotics" thesis has cleared the model-checkpoint threshold rather than just SDK availability.
The strategic mechanism is the same NVIDIA has run at the datacentre with CUDA: give away the model, lock in the compute target. Jetson Thor is the sanctioned inference target; the NVIDIA Open Model License and the LeRobot dataset format are the emerging standards (NVIDIA news). Every disclosed non-Figure OEM in the top tier now runs on NVIDIA-standardised weights — a stronger long-tail capture story than I held in the prior June scan — but it is a rent floor on compute, not a margin ceiling. The model layer is now explicitly commodity for any OEM without a proprietary VLA, and that matters for how one sizes exposure above Layer 4 of the value chain.
Figure is the exception. It runs its own Helix VLA and, on 30 June, delivered F.03 to BMW — extending a deployment relationship already validated at Spartanburg: 1,250 hours of continuous industrial duty cycle and more than 30,000 X3 vehicles over roughly ten months (figure.ai/news). Figure's BotQ is running at one unit per hour with more than 350 units delivered and over 80% first-pass yield. That is the industry's only disclosed, at-scale payback-relevant dataset, and the F.03 delivery to the same customer is same-customer-continues, not new-customer-adds — a meaningful distinction at this stage of the ramp.
The private-capital signal is equally sharp
Figure's 8 July notice of unauthorised secondary-market share sales — naming Hiive, Forge Global, EquityZen, Unicorns Exchange and Audeo Ventures (figure.ai/news) — is an acute private-market valuation signal even without a disclosed price. Separately, 1X's organisational cadence in a 35-day window deserves attention: World Model Lab launch on 4 June, VP Engineering hired 25 June, CFO hired 2 July, 25-DoF NEO Hands announced 9 July (1x.tech/discover). A CFO hire at this moment is a capital-raise and commercial-scale prep signal, not a product-cycle signal. The Brookfield supplier-and-buyer pattern flagged in the two prior scans — alt-manager capital sponsoring both the humanoid OEM (Figure Series C) and the demand-sink (Catalyst Brands apparel-DC, deal signed 26 May) within the same theme — may now have an analogue forming around 1X if it takes a mixed-capital round in Q3–Q4 2026. That round window is almost certainly already open.
1X's 25-DoF tendon-driven NEO hand also matters independently of the capital story. More degrees of freedom per hand means more tactile, torque and position sensors per unit, which lifts the per-robot BOM regardless of who wins the model race. TE Connectivity (TEL), sitting at roughly 16x forward, captures that content across OEMs; Harmonic Drive (6324.T) captures actuator volume as hand joints proliferate. Neither depends on the model-layer outcome. Advanced-packaging plays BESI.AS and Disco (6146.T) retain their role in the Jetson Thor CoWoS-S/L packaging queue, and Nabtesco and private Suzhou Green Harmonic sit alongside 6324.T in the reducer layer — all unchanged operationally.
MP reprices the upstream entry point
MP Materials (MP) has traded from a 52-week high of $100.25 to $51.71 as of 9 July — a roughly 48% drawdown — against a backdrop of USA Rare Earth litigation over magnet technology and engineer hiring practices, plus talent-crunch execution risk (Yahoo Finance). Deutsche Bank lowered its price target from $70 to $61 while maintaining Buy. The DoD floor at approximately $110/kg NdPr Oxide and the Apple offtake structure are structurally unchanged. The prior scan framed MP as richly priced at a sovereign-floored chokepoint; at $52 that sovereign support is an actionable margin-of-safety rather than a background note. MOFCOM Decree 2025-18 — governing heavy-rare-earth-separation licences for Dy, Tb and Sm — remains operative and binds humanoid motor magnets and Jetson Thor's compound-semi adjacents simultaneously, which is why this chokepoint merits the attention it does. Lynas (LYC.AX) and Iluka (ILU.AX) sit alongside MP at this layer; Lynas has published no 2026 commissioning updates through page one of its news index as of 10 July (lynasrareearths.com/news). For public-market context on the downstream: the BMW press feed contained no humanoid items between 20 June and 9 July (bmwgroup.com/press), and Tesla's news feed was rate-limited this cycle — so the Optimus production cadence remains a live Q3 earnings event, with Hexagon (HXGBY) and Symbotic (SYM, as the rigid-goods logistics analogue) providing the nearest public-market calibration in the OEM and end-buyer tiers respectively.
What would change my mind
If GR00T's successor ships under a proprietary rather than open-model licence, NVIDIA is re-tightening the model layer and the commoditisation dynamic reverses — the capability moat re-forms above Jetson Thor rather than below it, which directly supports rather than strains the AI-service gross-margin thesis. That single event would require a wholesale revision of where value accretes across Layers 4 and 5. Separately, if MP Materials misses its Q3 shipment run-rate against the Apple and DoD offtake contracts in the mid-November 10-Q, the sovereign-floor margin-of-safety thesis breaks and the upstream picks-and-shovels argument loses its most important structural support.
What I'm watching next
Figure's BotQ throughput update for multi-shift sustained performance by Q3; whether 1X announces a scaled commercial customer — apparel-logistics or auto-assembly analogue — by Q4; the Tesla Optimus unit-cost and volume print at Q3 earnings in late July; and whether a GR00T-N-successor releases with changed licence terms. On the upstream side, MP's Q3 10-Q shipment run-rate against Apple and DoD offtake is the primary re-verification event for the sovereign-floor thesis. The 1X capital round, if it materialises in Q3–Q4, will be the next alt-manager private-credit data point to track.