RobotPulse tracks the physical AI supply chain — actuators, precision drives, encoders, torque sensors, and the rare earth magnets that make all of it move. The parts of the humanoid nobody puts on a keynote slide.
Approximate bill-of-materials share. Directionally consistent with public teardowns and supplier commentary — treat as a framing device, not a spec sheet.
What's moving in the physical AI supply chain.
Every degree of freedom in a humanoid needs its own actuator — roughly seven per arm, forty-plus across a full body. Each one bundles a frameless motor, a precision reducer, an encoder, and a torque sensor into a package that has to survive millions of cycles at human-scale loads. That stack is where the majority of the bill of materials lives, and unlike the AI model layer, it can't be improved with a software update. Which makes the component suppliers — not the robot brands — the part of this market with pricing power.
Novanta supplies the electromechanical guts — motors, encoders, torque sensing. Timken supplies the precision motion products that turn raw motor speed into controlled, high-torque movement. Moog integrates complete actuation systems and carries a multi-billion-dollar backlog from aerospace and defense work that transfers directly. USA Rare Earth sits upstream of all of them, building domestic magnet capacity. Different risk profiles, same bottleneck.
High-torque actuators need high-performance permanent magnets, and high-performance permanent magnets need neodymium, praseodymium, and dysprosium — a supply chain overwhelmingly concentrated in China. Tesla has publicly flagged magnet availability as a limiter on Optimus production ramps. That single dependency is why a mining-and-magnets company shows up on a robotics watchlist at all, and why the domestic-capacity buildout is a policy story as much as an industrial one.
Reducers trade off backlash, stiffness, weight, efficiency, and cost, and no single design wins everywhere. Harmonic drives dominate in low-backlash arm joints. Cycloidal designs handle shock loading in hips and knees. Planetary gearsets are cheaper and better suited to lower-precision joints. As humanoid volumes scale, the pressure shifts from "which is best" to "which can be manufactured to spec at a hundred times current volume" — a question that favors incumbents with existing precision machining capacity.
Traditional industrial robots operate behind a cage precisely because they have no idea what they're touching. Humanoids meant to work near people need force and torque feedback at every joint, closing a control loop fast enough to stop before something breaks. Adding that sensing to forty-plus joints multiplies both cost and calibration complexity, and it's a meaningful driver of why per-unit actuator content is so high in this category.
The supplier thesis is genuinely attractive: component makers get paid regardless of which robot brand wins, and several of them already have profitable legacy businesses in aerospace, medical, and industrial automation. But that same diversification cuts the other way — robotics is a small revenue slice for most of these names today, so the stocks trade largely on their existing end markets. Investors buying the humanoid story need to be honest about how much of it is actually in the numbers yet.