Run metadata:

• ran_at: 2026-09-12
• job_id: physical_ai_model_to_matter
• task: Task 1 — Physical AI / Embodiment Scan + Task 2 — Ethan Portfolio Engine Report
• scope: physical AI / embodiment / robotics / bionics / AI materials / BCI / defense robotics / capital-flow signals
• source_policy: web-current; current claims below are based on sources fetched in this run.

Executive read

The strongest signal in this run is not another humanoid demo. It is the tightening stack around data → simulation/world models → deployment → manufacturing scale → defense procurement. Figure is explicitly framing its constraint as data + compute, signing a claimed $3.5B initial compute commitment with Nscale for up to 100,000 NVIDIA Vera Rubin GPUs, while Skild claims it crossed $100M ARR within ten months of first commercial deployment and is deploying its robot brain with NVIDIA and Foxconn for Blackwell assembly. These two signals shift the sector from “can humanoids move?” toward “who owns the robotic data/compute/deployment flywheel?” Sources: https://www.figure.ai/news/figu…REDACTED, https://skild.ai/blogs/skil…REDACTED. (figure.ai)

Deployment quality is now the filter. Figure’s BMW write-up claims Figure 02 ran 10-hour weekday shifts, loaded 90,000+ parts, logged 1,250+ runtime hours, and contributed to 30,000+ BMW X3 vehicles before fleet retirement into Figure 03 learnings; that is materially better signal than household chore videos because it includes hours, parts, shift cadence, and failure learnings. Source: https://www.figure.ai/news/production-at-bmw. (figure.ai)

Defense robotics/autonomy remains the most immediately monetizable physical-AI channel. The U.S. Army moved TITAN into production with two delivery orders totaling $192M: $127M to Palantir and $65M to Anduril for eight initial production systems, joining nine retained prototypes, with further production expected in FY27. Source: https://cpeisw.army.mil/2026/09/01/army…REDACTED/. (cpeisw.army.mil)

SIGNAL_LAYER

Robot intelligence stack: NVIDIA + Google + private “robot-brain” companies

NVIDIA continues to position itself as the physical-AI infrastructure layer: Cosmos world models, Isaac simulation, and Isaac GR00T models are being packaged as the development/training/deployment stack for generalized robots, with NVIDIA naming FieldAI, Skild AI, World Labs, 1X, Agility, Boston Dynamics, Figure, NEURA, and others as ecosystem users. Source: https://investor.nvidia.com/news/press-release-details/2026/NVID…REDACTED/. (investor.nvidia.com)

Google DeepMind’s Gemini Robotics 2 is relevant because it targets cross-embodiment control: the same model checkpoint is shown across Apptronik Apollo 2 variants and Franka Duo, while Google explicitly states that multi-finger dexterous manipulation remains challenging. Source: https://deepmind.google/blog/gemi…REDACTED/. (deepmind.google)

Skild AI’s September 2026 post is a high-signal company claim because it gives commercial language instead of demo language: $100M ARR within ten months of first deployment, 60+ paying customers, deployments across goods movement, delivery, inspection, security, food preparation, warehouses, factories, and data centers, plus NVIDIA/Foxconn Blackwell assembly work. This is company-reported, not independently audited in this run. Source: https://skild.ai/blogs/skil…REDACTED. (skild.ai)

Humanoid embodiment: manufacturing scale is becoming the dividing line

Figure says BotQ has moved from prototype to production phase and has demonstrated the “one robot per hour” cycle time needed for Figure 03 production targets, using custom manufacturing execution software across 150+ networked workstations. Source: https://www.figure.ai/news/ramp…REDACTED. (figure.ai)

Figure also claims Figure 03 returned to BMW Spartanburg Hall 52 in June 2026 for a logistics sequencing workflow, moving beyond prior sheet-metal loading toward whole-body manipulation and cart movement under Helix 02. Source: https://www.figure.ai/news/f-03-at-bmw. (figure.ai)

Unitree’s January 2026 clarification states 2025 humanoid shipments exceeded 5,500 units delivered to end customers, with mass-production output exceeding 6,500 units, excluding wheeled dual-arm robots and other robot forms. Source: https://shop.unitree.com/blogs/news/clar…REDACTED. (shop.unitree.com)

Tesla’s 2026 filing says it planned to unveil Optimus Gen 3 in Q1 2026, describes Gen 3 as its first mass-production design, and says start of production is planned before the end of 2026 with eventual planned capacity of 1M robots/year. This is high-upside but still mostly “planned capacity / preparation” until external deployment metrics appear. Source: https://ir.tesla.com/_flysystem/s3/sec/000162828026003837/tsla-20260128-gen.pdf. (ir.tesla.com)

Hyundai/Boston Dynamics is a slower but industrially credible lane: Hyundai says RMAC opened in the U.S. in June 2026, will expand ten-fold by year-end, trains/verifies manufacturing AI robots before production-line use, and targets Atlas deployment at Hyundai Motor Group Metaplant America from 2028. Source: https://org1.hyundai.com/content/hyundai/worldwide/en/newsroom/detail/0000001260.html. (org1.hyundai.com)

Dexterity / bionics bridge

PSYONIC’s Ability Hand integration into NVIDIA Isaac Lab matters because it connects prosthetics-grade sensorized hands with robotics policy training and “real-to-real” transfer. The company says the Ability Hand is a native Isaac Lab asset and demonstrates transfer of a pipetting manipulation task from a human prosthetic user to multiple robots. Source: https://www.psyonic.io/news/pres…REDACTED. (psyonic.io)

Google DeepMind’s own Gemini Robotics 2 blog states multi-finger dexterous manipulation remains challenging, confirming the standing thesis that useful hands are still the economic frontier, even as locomotion and gross whole-body control improve. Source: https://deepmind.google/blog/gemi…REDACTED/. (deepmind.google)

BCI / human-machine boundary

Neuralink’s current trials page lists active trials for controlling a computer and robotic arm through thought, plus an active trial for decoding words from thought, with a future visual-perception trial. Source: https://neuralink.com/trials/. (neuralink.com)

Synchron’s current trial materials show two active Stentrode studies, INTENT and FOCUS, focused on restoring digital-device control for people with severe bilateral upper-limb motor weakness; Synchron also emphasizes that its system remains investigational and is not commercially approved. Sources: https://synchron.com/join, https://synchron.com/. (synchron.com)

BCI remains portfolio-watch rather than direct public-equity action in this scan: the technical direction is clear, but the investable leaders are mostly private and still clinical/regulatory.

AI materials / autonomous labs

DOE/ARPA-E announced $34M for 12 CATALCHEM-E projects pairing AI with self-driving labs to accelerate industrial catalyst development for fuels and chemicals production. Source: https://arpa-e.energy.gov/news-and-events/news-and-insights/us-d…REDACTED. (arpa-e.energy.gov)

NIST’s Autonomous Formulation Lab links AI-driven experimental planning with robotic platforms and in-situ measurement in a closed loop for formulation optimization; the page was updated September 9, 2026. Source: https://www.nist.gov/programs-projects/autonomous-formulation-lab. (nist.gov)

Lawrence Berkeley Lab’s Genesis Mission project page describes specialized AI agents combining literature, simulations, and experimental data to propose, test, and refine materials hypotheses in Berkeley Lab’s robotic A-Lab. Source: https://www.lbl.gov/genesis-mission-projects/. (lbl.gov)

CAPITAL_FLOW_MAP

• Figure AI — private: exceeded $1B in Series C committed capital at a $39B post-money valuation in September 2025; latest September 2026 signal is a claimed $3.5B initial compute commitment with Nscale, with intent to scale above $6B. Sources: https://www.figure.ai/news/series-c, https://www.figure.ai/news/figu…REDACTED. (figure.ai)
• Skild AI — private: January 2026 Series C announcement says SoftBank led the round with NVentures, Macquarie Capital, and Bezos Expeditions participation, valuing Skild above $14B; September 2026 company claim says $100M ARR and 60+ paying customers. Sources: https://www.skild.ai/blogs/series-c, https://skild.ai/blogs/skil…REDACTED. (skild.ai)
• Apptronik — private: press page lists over $935M Series A, with repeat investors including B Capital, Google, Mercedes-Benz, PEAK6 and new investors including AT&T Ventures, John Deere, and QIA; Apptronik also lists Google DeepMind, Jabil, Mercedes-Benz, and GXO relationships. Source: https://apptronik.com/company/press-releases. (apptronik.com)
• SoftBank / ABB Robotics: ABB agreed to divest its Robotics division to SoftBank for $5.375B enterprise value, expected to close mid-to-late 2026 subject to approvals. Source: https://new.abb.com/news/detail/129776/abb-…REDACTED. (new.abb.com)
• Defense procurement: U.S. Army TITAN production awards total $192M as of Aug. 31, 2026, split between Palantir and Anduril; production order for additional systems is anticipated in FY27. Source: https://cpeisw.army.mil/2026/09/01/army…REDACTED/. (cpeisw.army.mil)
• AI materials public funding: ARPA-E’s $34M CATALCHEM-E and DOE/NIST/Genesis Mission autonomous-lab efforts indicate state-backed conversion of AI+robotics into scientific infrastructure rather than consumer product timelines. Sources: https://arpa-e.energy.gov/news-and-events/news-and-insights/us-d…REDACTED, https://www.nist.gov/programs-projects/autonomous-formulation-lab. (arpa-e.energy.gov)

CONVERSION_LAYER

Higher-conviction conversion signals

1. Figure at BMW: actual runtime, parts loaded, shift cadence, vehicle contribution, and hardware failure learnings move Figure above demo-only humanoid peers. Source: https://www.figure.ai/news/production-at-bmw. (figure.ai)
2. Skild commercial ARR claim: if verified later by customers/contracts, Skild becomes one of the clearest “robot brain” conversion examples: cross-embodiment software sold into real customers rather than vertically integrated humanoid capex alone. Source: https://skild.ai/blogs/skil…REDACTED. (skild.ai)
3. TITAN production: defense autonomy is converting through procurement faster than domestic humanoids because the buyer accepts high capex, integration complexity, and supervised autonomy. Source: https://cpeisw.army.mil/2026/09/01/army…REDACTED/. (cpeisw.army.mil)
4. NVIDIA robotics stack: NVIDIA is converting from GPU supplier to robotics development substrate: Cosmos/Isaac/GR00T, simulation, synthetic data, validation, and robot deployment infrastructure. Source: https://investor.nvidia.com/news/press-release-details/2026/NVID…REDACTED/. (investor.nvidia.com)

Lower-conviction / watch-only signals

• Tesla Optimus remains high-optionality but needs external deployment data, reliability metrics, unit economics, and customer use before being treated as more than embedded call-option value inside TSLA. Source: https://ir.tesla.com/_flysystem/s3/sec/000162828026003837/tsla-20260128-gen.pdf. (ir.tesla.com)
• Home humanoids remain weaker than industrial/logistics deployments because domestic environments multiply edge cases and safety requirements; Figure and 1X are pushing toward home, but portfolio weight should still privilege industrial conversion until real paid home retention data exists. Sources: https://www.figure.ai/news/intr…REDACTED, https://www.1x.tech/discover/neo-factory. (figure.ai)

TIME_MAP

0–6 months

• Watch Figure’s Index/data claims, Nscale compute activation, BotQ production rate, and BMW F.03 follow-up metrics. Sources: https://www.figure.ai/news/figu…REDACTED, https://www.figure.ai/news/ramp…REDACTED, https://www.figure.ai/news/f-03-at-bmw. (figure.ai)
• Watch Skild customer verification, renewals, and whether NVIDIA/Foxconn Blackwell assembly deployment becomes public customer proof rather than company blog claim. Source: https://skild.ai/blogs/skil…REDACTED. (skild.ai)
• Track TITAN delivery execution and FY27 follow-on orders. Source: https://cpeisw.army.mil/2026/09/01/army…REDACTED/. (cpeisw.army.mil)

6–18 months

• Tesla Optimus: production start before end-2026 is the first gate; useful factory work, volume, and intervention rates are the real gates. Source: https://ir.tesla.com/_flysystem/s3/sec/000162828026003837/tsla-20260128-gen.pdf. (ir.tesla.com)
• SoftBank/ABB: if the transaction closes, SoftBank becomes a larger AI+industrial robotics aggregator, but integration risk remains high. Source: https://new.abb.com/news/detail/129776/abb-…REDACTED. (new.abb.com)
• Apptronik: watch whether Mercedes-Benz/GXO/Jabil/Google DeepMind relationships produce measurable paid fleet deployments, not just partnership language. Source: https://apptronik.com/company/press-releases. (apptronik.com)

18–36 months

• Hyundai/Boston Dynamics Atlas has a 2028 deployment target at HMGMA and a 2028 robot-production ramp; this is longer-dated but backed by manufacturing/distribution/finance infrastructure. Source: https://org1.hyundai.com/content/hyundai/worldwide/en/newsroom/detail/0000001260.html. (org1.hyundai.com)
• Autonomous labs should begin producing more industrially relevant materials/catalyst outputs if DOE/NIST/Genesis Mission workflows mature from lab infrastructure to commercialization channels. Sources: https://arpa-e.energy.gov/news-and-events/news-and-insights/us-d…REDACTED, https://www.nist.gov/programs-projects/autonomous-formulation-lab. (arpa-e.energy.gov)

BOTTLENECK_INDEX

1. Dexterous manipulation: confirmed by Google DeepMind’s own statement that multi-finger dexterous manipulation remains challenging; this remains the central physical-economy bottleneck. Source: https://deepmind.google/blog/gemi…REDACTED/. (deepmind.google)
2. Deployment reliability: Skild explicitly warns that a successful demo clip can hide low reliability and says the final 5% of performance can dominate effort. Source: https://skild.ai/blogs/skil…REDACTED. (skild.ai)
3. Data + compute: Figure now states it is largely bound by data and compute needed to train Helix, making compute access a real robotics bottleneck rather than a generic AI talking point. Source: https://www.figure.ai/news/figu…REDACTED. (figure.ai)
4. Manufacturing and supply chain: Figure’s BotQ update and Tesla’s Optimus filing both emphasize production lines, supply chain readiness, and mass-production design, meaning hardware manufacturability is now thesis-critical. Sources: https://www.figure.ai/news/ramp…REDACTED, https://ir.tesla.com/_flysystem/s3/sec/000162828026003837/tsla-20260128-gen.pdf. (figure.ai)
5. Regulatory/clinical path for BCI: Synchron and Neuralink remain investigational/clinical rather than commercial-scale product stories. Sources: https://synchron.com/, https://neuralink.com/trials/. (synchron.com)

THESIS_STATE

• Physical AI thesis: strengthened. The field is moving from demos to infrastructure: compute, robot data, simulation, deployment learning loops, and manufacturing capacity.
• Humanoid thesis: strengthened but narrowed. The investable thesis should overweight companies proving deployment economics and underweight chore-video narratives.
• Dexterity thesis: unchanged / confirmed. Hands, tactile sensing, and manipulation under time constraints remain the frontier.
• Defense robotics thesis: strengthened. Procurement conversion is happening now, not only in speculative future markets.
• BCI thesis: strategically important but still watchlist for public markets; mostly private, clinical, and regulatory-gated.
• AI materials thesis: strengthened as a long-duration infrastructure trend; autonomous labs are moving into DOE/NIST/NSF-backed platforms but are not yet obvious near-term public-equity revenue drivers.

Sources

1. Figure / Nscale GPU partnership — https://www.figure.ai/news/figu…REDACTED
2. Figure BMW deployment — https://www.figure.ai/news/production-at-bmw
3. Figure F.03 at BMW — https://www.figure.ai/news/f-03-at-bmw
4. Figure BotQ / production ramp — https://www.figure.ai/news/ramp…REDACTED
5. Figure Series C — https://www.figure.ai/news/series-c
6. Figure unauthorized stock notice — https://www.figure.ai/news/noti…REDACTED
7. Skild AI ARR / deployment post — https://skild.ai/blogs/skil…REDACTED
8. Skild AI Series C — https://www.skild.ai/blogs/series-c
9. NVIDIA physical AI ecosystem — https://investor.nvidia.com/news/press-release-details/2026/NVID…REDACTED/
10. NVIDIA Isaac GR00T technical blog — https://developer.nvidia.com/blog/deve…REDACTED/
11. Google DeepMind Gemini Robotics 2 — https://deepmind.google/blog/gemi…REDACTED/
12. Apptronik press releases — https://apptronik.com/company/press-releases
13. Unitree 2025 shipment clarification — https://shop.unitree.com/blogs/news/clar…REDACTED
14. Tesla 2026 filing / Optimus — https://ir.tesla.com/_flysystem/s3/sec/000162828026003837/tsla-20260128-gen.pdf
15. Hyundai 2026 CEO Investor Day / robotics — https://org1.hyundai.com/content/hyundai/worldwide/en/newsroom/detail/0000001260.html
16. U.S. Army TITAN production — https://cpeisw.army.mil/2026/09/01/army…REDACTED/
17. Anduril CCA autonomy — https://www.anduril.com/news/u-s-…REDACTED
18. PSYONIC / NVIDIA Isaac Lab — https://www.psyonic.io/news/pres…REDACTED
19. Neuralink trials — https://neuralink.com/trials/
20. Synchron study page — https://synchron.com/join
21. Synchron homepage / investigational notice — https://synchron.com/
22. ARPA-E CATALCHEM-E — https://arpa-e.energy.gov/news-and-events/news-and-insights/us-d…REDACTED
23. NIST Autonomous Formulation Lab — https://www.nist.gov/programs-projects/autonomous-formulation-lab
24. Berkeley Lab Genesis Mission projects — https://www.lbl.gov/genesis-mission-projects/
25. ABB Robotics sale to SoftBank — https://new.abb.com/news/detail/129776/abb-…REDACTED
26. Amazon Vulcan — https://www.aboutamazon.com/news/operations/amaz…REDACTED
27. Amazon Blue Jay / agentic AI fulfillment — https://www.aboutamazon.com/news/operations/new-…REDACTED