A dozen layers stand between the foundation model and the factory floor. The naïve view treats them as a stack — each layer earning a fair cut of the robot. The truth is more asymmetric.
Four forces crush some layers toward zero. Four forces make others permanent. What emerges is not a stack. It is a silhouette.
The naïve view of the Physical AI value chain. Foundation models at the top. Distribution at the bottom. Everything in between, weighted equally — as if margin were fair.
This is the last time it will look like this.
A robot is not sold by a model. It is sold by whoever owns the buyer — the factory, the warehouse, the hospital system. Distribution meets the customer before any other layer does.
Every dollar upstream is a pass-through.
Open source flattens simulation. Global labor arbitrage — and then the automation of the arbitrageurs — guts teleoperation. Environmental fragmentation prevents scale in integration and maintenance.
The layers that looked like obvious businesses turn out to be obvious wages.
Training capex and proprietary robot data protect foundation models. A supplier monopoly — the Harmonic Drive pattern — protects specialty actuation. Network effects from fleet data protect insurance. Customer ownership protects distribution.
Everything else is flat.
The silhouette resembles the smile curve from semiconductors — fat at the ends, thin in the middle — with a single bump for precision actuation.
Everything else in the middle is engineering work that pays engineering wages.
Build at the ends of the stack, or inside a physics-constrained hardware pocket. Do not build in the middle unless you intend to work for wages.
The forces are not an opinion. The silhouette is a forecast. One integration platform company, one regulation that neutralizes insurance pricing, one OEM that captures distribution — two of those three are already in motion.
Hover a layer to read the forcesPercentages are margin estimates. * speculative · † modeled · ‡ sourced · ~ qualitative. Current chart values are speculative estimates.
If you are building in Physical AI, the most important question is not which layer but what is the force that shapes that layer, and what does it look like in ten years.
A layer with open-source pressure will be compressed. A layer with network effects from its own data will stay fat. A layer with a supplier monopoly will stay fat. A layer with customer ownership will stay fat. A layer that varies environment-by-environment is fragmented today — until someone writes the abstraction. That company doesn’t exist yet in Physical AI.
The rest of it is wages, dressed up as a company.
· · ·