State of Physical AI
A US-first value-chain map of humanoids, embodied AI, and smart hands — where the margin actually pools as of June 2026
Own the bottlenecks and the brains; rent the robots. Avoid pure-play OEM hype.
Concentrate exposure where defensibility is real and pricing power is durable: the compute layer (NVIDIA Jetson Thor effectively standard for on-robot inference) and the physical bottlenecks China weaponizes — rare-earth magnets, precision reducers, planetary roller screws. Among private names, prefer the defense-autonomy software franchises with revenue and contract backlogs (ANDURIL high-conviction at ~$61B, SHIELDAI medium at ~$12.7B with ~$540M 2026 revenue) over capital-intensive humanoid OEMs racing to sub-$20k price points before unit economics close (FIGURE medium at ~$39B — extraordinary deployment momentum but a hardware P&L still subsidized by venture capital). For public access, use diversified ecosystem ETFs (KOID for humanoid-pure, BOTZ/ROBO for breadth) plus the diversified arms (NVDA, TSLA, GOOGL) rather than single-name bets; this is not financial advice.
No single layer is 'won,' but the durable margin sits in two places: (1) edge compute, where NVIDIA Jetson Thor (Blackwell, 2,070 FP4 TFLOPS, 128GB) is the de-facto inference standard with CUDA lock-in and a vertically integrated GR00T model stack; and (2) the physical bottlenecks China controls — ~69% of rare-earth mining, ~90% of magnet refining (export-controlled since Apr 2025), plus precision reducers and planetary roller screws with 26+ week lead times. The AI-brain layer (Physical Intelligence π0, Figure Helix 02, NVIDIA GR00T N2) is the highest-optionality layer but the least settled — defensibility is still forming.
Full-stack value chain
Seven layers scored on bottleneck control + margin durability, not demand. The margin pools at the top (compute, magnets) and the bottom is a trap (OEM assembly).
The board's central lesson: a layer can be a genuine bottleneck (precision reducers, with 26+ week lead times) and STILL be a margin desert when a low-cost rival scales into it (Harmonic Drive's 2.7% net margin). Real, durable value pools where control is COMPOUNDING — compute (NVIDIA's chip+model+sim vertical) and the rare-earth chokepoint China defends with export controls. Own those layers (and the defense-autonomy software franchises) as the picks; rent the robot-OEM layer through diversified ETFs, because that's the layer where exploding demand and collapsing margin are the same place.
Shift-point register
Ranked by margin-at-stake × demand-durability × evidence-strength. The flagged rows clear the bar and are promoted to a deep-dive.
VLA brain in-housing
Value migrates from the robot body to the in-house VLA/world-model brain
Once chassis become commoditized (Unitree/AgiBot drive BOM toward $20-35k), the durable margin and differentiation sits in the policy model — the VLA that turns one demonstration into generalized behavior. Whoever owns the model + the data flywheel owns the recurring software/per-task economics; the body becomes a cost center.
Opportunity board
Where to position as of June 2026 — ranked by defensibility, then investability. Own the bottlenecks and the brains; get OEM exposure only through ETFs.
Durable compounder
4Undervalued / high-potential
6Short / avoid
0No names in this bucket.
Views & the Voices
The strongest bull and the strongest bear case, the US-listed ETF expressions of the theme, and where the tracked QAI Voices roster nets out — each stance stamped trackable vs inferred.
Bull case
the longPhysical AI is the logical next S-curve after generative AI: a general-purpose humanoid amortizes one VLA/world-model stack across a multi-trillion-dollar labor base, and 2026 marks the cross-over from demo-video era to real commercial deployments, real revenue, and dedicated production lines. The same names cross-link three tailwinds — frontier embodied-AI models (Figure Helix), defense autonomy at scale (Anduril, Shield AI), and the silicon/component picks-and-shovels (NVDA, rare-earth/actuator supply chain).
- TAM is enormous and being revised UP, not down: Goldman lifted its 2035 humanoid forecast 6x (from $6B to $38B); Morgan Stanley models a ~$5T humanoid ecosystem by 2050 (~13M units by 2035 scaling toward 1B by 2050), with unit price falling from ~$200k toward ~$50k in wealthy markets — an EV-like cost curve. (GS / MS, 2025-2026 vintage)
- Deployments are now real, not staged: Figure 02 ran 11 months on BMW's Spartanburg X3 line (30,000+ vehicles, 90,000+ parts, 1,250 operating hours, completed Nov 2025); the field has ~16,000 humanoid units deployed and platforms transitioning to actual production. (Figure / BMW, Nov 2025)
- Defense autonomy is a here-and-now revenue engine with structural budget tailwinds: Shield AI projects 80%+ growth to $540M+ revenue in 2026 with a U.S. Air Force CCA mission-autonomy selection; Anduril doubled to a $61B valuation on a $5B Series H (May 2026, Thrive + a16z). These are deployed-product companies, not pre-revenue. (Fortune Mar 2026; CNBC/TechCrunch May 2026)
Bear case
the short / avoidValuations price a labor-replacement future that the technology has not delivered: humanoid revenue is tiny, most 'capability' is teleoperated or human-assisted, the data modality dexterity actually needs (touch/force) has no internet-scale corpus, and the high-torque hardware sits on a Chinese rare-earth magnet chokepoint. The single-stock picks here (FIGURE pre-revenue, defense names cyclical to budgets) and the ETFs (mostly old-line industrial automation, not pure humanoid) both carry hype-cycle re-rating risk.
- Valuations have detached from revenue: Figure's $39B post-money (Sep 2025) on essentially zero revenue already exceeds Goldman's entire $38B 2035 humanoid TAM — nine years early. Optimus targets keep slipping (Musk admitted ~zero Optimus did 'useful work' in 2026; V3 reveal pushed again), and prediction markets repeatedly doubt the timelines. (Fortune/Electrek 2026)
- The autonomy is largely illusory: nobody publishes mean-time-between-interventions (the field's most honest number); NEO ships at ~60-70% autonomy (a third of tasks done by a remote human), and Tesla's in-factory units are 'primarily for learning and data collection, not productive tasks.' A product that needs a human in another room is not finished. (independent reviews, 2026)
- A structural data gap caps dexterity: the train-on-human-video + teleop-fine-tune recipe is missing the haptic/force modality human dexterity runs on, with no equivalent corpus — and hardware confirms it (Figure's BMW report flagged the densely-packed, thermally-constrained forearm as the dominant failure point).
Constructive-but-divided: strong agreement that embodied/physical AI is a real, large, multi-decade category and that defense autonomy (Anduril, Shield AI) is already a deployed-revenue business — but sharp disagreement on whether 2026 humanoid valuations are justified by anything other than narrative.
Dispersion: Very wide. TAM estimates span an order of magnitude by horizon (GS $38B by 2035 vs MS ~$5T by 2050); private marks are inflecting up violently (Figure $39B, Anduril $61B, Skild $14B, PI $11B) while capability skeptics and rare-earth analysts argue the autonomy and supply-chain reality lags the marks by years. The bull/bear gap is timing-driven, not category-driven.
The cleanest split: defense-autonomy names have current revenue + budget tailwinds and command higher conviction (Anduril high, Shield AI medium); pure humanoid plays (Figure medium) are priced on optionality. Watch three falsifiable signals — a published autonomous MTBI on paid tasks, non-China magnet supply scaling, and pilot-to-recurring-revenue conversion. ETF wrappers give thematic beta, not humanoid purity. Not financial advice.
US-listed ETF expressions
4The QAI Voices
stance · trackable / inferredPrediction matrix
Directional calls across Sep'26 / Dec'26 / Jun'27 / Jun'28, confidence decaying high → low over the horizon. Each cell is the call; click a row for the full reasoning, leading indicator, and falsifier.
As of late June 2026, the robotics/embodied-AI value pool is migrating along eight tracked sub-themes, with the sharpest, best-evidenced shifts concentrated at two ends of the stack: the in-house VLA/world-model brain (Figure Helix 02 in-fleet; NVIDIA Cosmos 3 + GR00T as the rent-the-brain alternative across 14+ partners) and the physical chokepoints that gate every Western ramp — China's rare-earth magnet licensing (Optimus' 'magnet issue') and ~26-week harmonic-drive actuator lead times. Manipulation data, not compute, is the consensus scarce input, making fielded-fleet scale a strategic data-acquisition play; the dexterous hand is fast-commoditizing yet strategically pivotal (Tesla scrapped its 22-DoF V3 hand as 'didn't work'). On the business-model side, RaaS ($2-8k/mo vs $18-35/hr labor; Toyota/Amazon/Spanx/Mercedes pilots) shifts value to fleet operators, and defense autonomy is the highest-margin, government-funded proving ground (Anduril $61B / $4.3B-2026-guide / $20B Army ceiling; Shield AI $12.7B / >$540M / Hivemind across 26 vehicle classes). The dominant uncertainties: whether the proprietary-data moat is durable or open datasets close it, and whether Western mass-production (Figure BotQ ~1 robot/90min, Tesla Fremont line) can close China's ~80%-of-2025-units cost+volume lead. Covered private names — FIGURE (medium conviction), ANDURIL (high), SHIELDAI (medium) — plus reference arms NVDA/TSLA/GOOGL. Five regimes span brain-led winner-take-most, horizontal picks-and-shovels, supply-chain-gated, defense-led, and hype-de-rating outcomes. Not financial advice; private valuations carry a wide valuation-to-revenue gap. Sources span KraneShares/KOID, McKinsey, Figure, Anduril, Shield AI, NVIDIA, and multiple supply-chain teardowns, June 2026 vintage.
Valuation scenarios
Every target is scenario-conditional with a probability; the verify produced zero outright buys. Tap a name for its full bull / base / bear ladder.
Regime calls
The four cross-cutting forces and when each bites across the Sep'26 → Jun'28 horizon. Tap any force or modulating risk to read the full call.
Regime A
The VLA/world-model + proprietary fleet-data flywheel proves a durable moat. Value concentrates in a few vertically-integrated owners (Figure, Tesla) who own brain+body+data; the body commoditizes. Bull for first-movers with real fleets; bear for body-only integrators. Triggered by VLAs showing real task-generalization in production (matrix falsifier NOT hit) and proprietary data beating open baselines.
Regime B
NVIDIA Cosmos/GR00T + open datasets commoditize the brain; many integrators field competent robots on a shared stack. Margin pools to the horizontal compute/model/sim platform (NVDA) and to the scarce physical layers — rare-earth magnets, harmonic-drive actuators. Bull for NVIDIA and component chokepoints; bear for the proprietary-VLA-as-moat thesis (cross-cutting risk #3 realized).
Regime C
China's rare-earth licensing + cost+volume lead (75k+ units/yr) caps Western ramps regardless of AI progress; the binding constraint is physical, not algorithmic. Value accrues to ex-China material/magnet supply, rare-earth-free motor topologies, and vertically-integrated actuator makers. Western valuations de-rate on slipped ramps; Chinese OEMs dominate unit volume. Triggered by sustained license friction + ex-China supply staying pre-volume.
Regime D
Government-funded defense autonomy (Anduril Lattice, Shield AI Hivemind) scales the embodied-AI + sim-to-real stack profitably years before commercial humanoid RaaS reaches durable revenue. The highest-margin, best-capitalized embodied-AI business is defense; commercial humanoids remain in the pilot valley. Bull for the defense-software names; bear for near-term commercial-humanoid revenue timing. Triggered by Anduril/Shield AI meeting guides while RaaS pilots fail to convert to multi-site contracts.
Regime E
Private valuations (Figure $39B pre-revenue) compress on a pilot-to-platform stall, a high-profile deployment/safety failure, or a capital-markets rotation. KOID/BOTZ NAVs fall, down-rounds cascade, capex-heavy ramps burn cash against distant revenue. The whole complex re-rates to revenue, not narrative. Triggered by deployment stalls + the valuation-vs-revenue gap (cross-cutting risk #2) closing the wrong way.
How they modulate the book
China supply-chain concentration and weaponization
China shipped ~80% of 2025 humanoids and controls ~60% of rare earths, ~26% of actuators, and the cheapest motors/reducers (Unitree/AgiBot vertically integrated). The April 2025 dual-use rare-earth licensing regime can throttle Western ramps at will (Optimus 'magnet issue'). A bilateral escalation cuts both ways — Western OEMs lose inputs; Chinese OEMs lose access to leading-edge compute/VLA tooling. This single dependency sits underneath shifts #2, #3, #8, and #10.
Private-valuation-to-revenue gap and hype-cycle de-rating
Figure at $39B post-money is pre-meaningful-revenue; even Anduril ($61B) and Shield AI ($12.7B) trade at high multiples of $2-4B and ~$540M revenue respectively. Morgan Stanley's $5T-by-2050 TAM frames a long-duration bet. A pilot-to-platform stall, a single high-profile deployment failure, or a funding-winter rotation could compress the whole private complex — KOID/BOTZ NAVs and down-rounds are the transmission channel. The gap between demo virality and durable per-task revenue is the core bear thesis.
Data-flywheel moat may not be defensible
If open datasets (AgiBot World, Open X-Embodiment), egocentric-video pretraining (EgoDex), and sim-to-real (Cosmos/Isaac) close the gap, the proprietary-fleet data advantage erodes and VLA capability commoditizes — collapsing the shift-#1 and shift-#5 margin thesis. Cross-embodiment transfer working well is bullish for fast-following integrators but bearish for the first-mover data moat. The unsettled question: is manipulation data a compounding moat or a temporary lead?
Safety, liability, and labor-displacement regulation
Untethered high-torque humanoids working alongside humans carry physical-injury liability with no settled regulatory/insurance framework; a serious workplace incident could impose certification regimes that slow deployment industry-wide. Labor-displacement politics (warehouse jobs, 150% turnover notwithstanding) invite policy friction. Defense autonomy (Anduril/Shield AI) carries parallel ROE/export-control/lethal-autonomy scrutiny. Regulation is the cross-cutting brake on the RaaS deployment curve (shift #6).
Capital intensity and manufacturing-execution risk
Industrializing humanoids is brutally capex-heavy (Anduril's >$900M Arsenal-1 campus; Figure/Agility/Tesla purpose-built lines) and execution-fragile — precision actuator integration, 26-week gearset lead times, and qualification cycles resist the 'just add capacity' playbook. A ramp that slips (Tesla repeatedly pushed Optimus timing; scrapped its V3 hand) burns capital against a still-distant revenue base. The pilot-to-volume valley is where the most credible names can still stall.
Premise pressure-test
The six named 2026-Q3 catalysts the thesis rests on, probability-weighted. Click any premise for the if-true / if-false split.
Humanoid demand inflects, not fizzles
SupportedNVIDIA holds the edge-inference standard
SupportedThe rare-earth chokepoint persists
SupportedThe brain doesn't fully commoditize
ContestedOEM assembly stays margin-poor
SupportedDefense-autonomy revenue is durable
SupportedPrecision-reducer pricing keeps eroding
SupportedSub-$20k consumer-humanoid timelines are marketing
Contested