House Research
Physical AI — The Supply-Chain Value-Capture Heat Map
The Physical-AI value chain, six layers from raw inputs to end applications, each
tested with one question: does this layer capture value or commoditize — and on what
mechanism? Value flees modular, multiply-sourced layers and concentrates where someone
holds a scarce input, sets the standard, or owns the customer.
As of 28 July 2026 · verdicts revisited as evidence changes
Captures scarce input · standard · customer lock
Contested capture possible, mechanism unproven or migrating
Commoditizes modular and multi-sourced — margin compresses
▼ End applicationsRaw inputs ▼
L6Apps
Integration / Applications
Deployers embed robots into real operations — logistics, auto plants, eventually the home.
GXO · BMW · DHL · Toyota Canada · Schaeffler · Mercado Libre
Captures — unpriced
Customer lock plus the deployment-experience loop, at near-zero valuation premium.
Why it captures: deployers accumulate operational truth — utilization, failure
modes, integration cost — that no lab can buy, while paying warehouse-robot benchmark prices for the
option on labor substitution. What would change the verdict: humanoid work-rates failing to close the
gap against proven warehouse robots by ~2028.
L5Data
Data & Simulation
Teleoperated demonstration data, egocentric video, and the simulation substrates that train the models.
Scale AI · XDOF · state-backed collection centers · Nvidia Isaac + Cosmos · MuJoCo
Contested
Raw collection is commoditizing fast; capture migrates to fleet-captive experience.
The refinement: a frontier-scale training corpus now costs low-single-digit millions
to assemble — the raw-data moat is dollars-shallow. What cannot be bought at any price: corrective
experience from robot fleets working under real economics. What would change the verdict:
breadth-generalization achieved mainly on purchased or synthetic data.
L4Brains
Foundation Models (VLA “brains”)
Vision-language-action models that turn pixels plus instructions into motor commands.
Physical Intelligence · Skild · Field · Nvidia GR00T · Gemini Robotics · Figure Helix
Contested → commoditizes
Architectures diffuse as open weights; platform giants price the model layer at roughly zero.
The evidence: $20B+ of private marks against one unaudited $30M revenue claim in the
entire category; the only realized exit went at roughly 0.6x its last mark. The master crux: does the
imitation-learning scaling law bend toward breadth before those marks reprice?
L3OEMs
Robot Hardware / OEMs
Humanoid and general-purpose robot assembly — the visible layer the funding chases.
Unitree · Figure · Tesla Optimus · Agility · UBTech · Apptronik · 1X
Commoditizes
Assembly margin is already compressing at the volume leader; escape requires owning the customer or the data loop.
The 2026 tell: the layer's only profitable maker grew revenue 68% in Q1 while
recurring profit fell 52%, and whole-robot prices are down ~72% in two years — the EV-assembler pattern,
not the Apple one. What would change the verdict: any OEM sustaining 40%+ gross margin at 10k+ units a
year on third-party components.
L2Compute
Compute / Silicon
Datacenter training compute and on-robot inference chips, plus the toolchain between them.
Nvidia (Thor, CUDA / Isaac) · Qualcomm Dragonwing · Tesla captive silicon
Captures — at training
Scarce training supply plus the CUDA/Isaac standard; edge silicon itself commoditizes.
The strategy tell: the chip leader gives the robot brain away free to commoditize the
layers above and below its silicon — capping every model licensor's margins in the process.
What would change the verdict: a credible non-CUDA robotics training stack winning a top-five lab.
L1Inputs
Components / Actuation
Precision reducers, roller screws, motors, magnets, sensors, batteries — roughly 40–60% of a humanoid's bill of materials.
Harmonic Drive · Nabtesco · Leaderdrive · MP Materials · LG Energy Solution
◆ Rare-earth magnets — the one strengthening moat. China controls
~90% of separation and magnet-making; prices have roughly doubled off the 2025 average; the export-control
suspension expires 10 November 2026. About 1% of the bill of materials — gating 100% of production.
Captures — repricing
Mechanical-precision moats are melting under Chinese price deflation; the magnet chokepoint is strengthening at the state level.
The divergence inside the layer: reducer incumbents trade at ~125x earnings on falling
profit — a decaying moat priced as a durable one — while the magnet node carries a US policy floor
under its downside. What would change the verdict: a major OEM shipping at volume on all-domestic-Chinese
actuation without quality regression.
Hover, tap, or focus a layer for the mechanism and what would change the verdict.
How to read this map
Every layer is scored with the same test: a layer captures value only if it controls a scarce,
hard-to-replicate input, sets a standard others must conform to, or owns the customer relationship. A layer
that is modular and multiply-sourced commoditizes — margins compress toward the cost of capital
regardless of how fast units grow.
Capture is not static. The interesting question is never “who captures value now” but
“where is it heading, and what would move it” — which is why each layer carries an explicit
falsifier. Verdicts are revisited as the evidence changes.