Rare earth to robot.
Where the world's robots actually come from: the single node that gates the whole chain, and how tightly everything above it is bound.
You can't build an electrically actuated robot without a motor, you can't build the motor without a magnet, and you almost certainly can't build the magnet without China. Hydraulic and pneumatic machines escape that chain; almost nothing you'd call a modern robot does, and everything anyone is racing to build (humanoids, industrial arms, servos) runs on electric actuators. That sentence is either alarmist or load bearing depending on the numbers underneath it, so we went and got the numbers. This is a note about the structure of the robotics supply chain: how concentrated it is, where it actually binds, and how a shock at the bottom travels to the top.
A robot is a stack of dependencies, and the binding one sits four layers down. Roughly 40 to 60% of a humanoid's bill of materials is actuators; an actuator is mostly a motor; the motor needs high energy NdFeB magnets; the magnets need neodymium, praseodymium, and the heavy rare earths dysprosium and terbium to stay magnetic when the motor runs hot. The magnet is a small line on the cost sheet, call it 3% of the robot, and a total dependency: every torque motor stops without it.
The striking thing isn't that China is large in rare earths; that's known. It's that the grip tightens as you move upstream, exactly where substitution is hardest. China refines around 88% of the world's rare earths and sinters over 90% of its permanent magnets; for the heavy rare earths that keep magnets stable at temperature, the share is higher still. The robot is assembled in a dozen countries. The thing it can't run without is made in one.
Is the dependency real, or is it the kind of thing that's true on paper and slack in practice? The market settles it. Each time China tightened export licensing (gallium and germanium in 2023, the seven heavy rare earths in April 2025, the sweeping MOFCOM Notice 61 that October) the rare earth complex convulsed. Realised volatility in the listed proxies doubled or quadrupled; dysprosium roughly tripled and terbium more than doubled inside a few weeks. The downstream robot makers barely moved on the day. That gap is the chain made visible: the shock is violent at the chokepoint and arrives at the assembler a quarter later, fed through depleting magnet inventory.
Which inputs are actually chokepoints, and which only look like one? A useful filter is to ask not "how concentrated is it" but "can one actor weaponise it." By that test most of the robot's materials drop out. Copper, steel and aluminium are concentrated in production yet globally fungible and freely traded; there is no licence to withhold. Run the test across the bill of materials and only two nodes survive: rare earth magnets and battery graphite, which sit behind the same Chinese export regime and the same November 2026 cliff. The rest is macro commodity exposure wearing a robotics costume.
A word on the loudest part of the sector. Humanoid programmes move on attention: reveals, demos, order announcements. Attention is easy to mistake for a ramp. The supply chain tell for whether the build out is real is mundane and physical: orders for actuators, harmonic and cycloidal reducers, and the magnets inside them. We tested whether the attention itself carried information and found it didn't, cleanly. A fade signal that looked real on three names evaporated on seven (p = 0.75). The lesson is a supply chain one, not a market one: when the narrative runs ahead of the reducer orders, it's a story, not a chain. Watch the boring layer.
What this means if you build, insure, or govern around robots: your exposure is not to "China" in the abstract but to a specific, datable licensing regime over a handful of refined materials, with an inventory buffer of one to three months standing between an announcement and a stopped line. That makes the risk legible. You can map which of your inputs trace back to the chokepoint, watch the catalysts that move it, and know roughly how long you have. The honest limits: concentration figures are estimates and move year to year, substitution (motors without magnets, ferrite, recycling) is advancing but slowly, and a suspended control is not a repealed one.
The fragility isn't in the robot. It's four layers down, in a magnet that is three percent of the cost and a hundred percent of the dependency.