Everyone obsessing over the software brains inside humanoid robots may be watching the wrong part of the machine. The real bottleneck sits lower down, in the joints — in the actuators that let a robot lift, walk and grip. And at the heart of those actuators are rare earth magnets, the least glamorous and quietly most strategic component in the entire build.
The numbers make the case. Actuators account for 40% to 60% of a humanoid robot’s bill of materials. That is not a rounding error on the spec sheet — it is the single largest cost bloc in the hardware, dwarfing the cameras, sensors and compute that grab the headlines. If you want to scale humanoid production into the millions, you have to scale actuators first, and actuators mean magnets.
Here is where it gets uncomfortable for anyone hoping to build these machines outside a single supply chain. China refines roughly 90% of the rare earth magnets that go into these actuators. That is not a manufacturing preference; it is a chokepoint. A robotics company can license the smartest model in the world and still be unable to ship product if it cannot source the magnets that move the limbs.
The framing matters because it reorders what “physical AI” actually depends on. The models keep improving — that is the easy part, relatively speaking, because software scales without a foundry. Magnets do not. They require refined materials, specialized processing and a supply base that today is overwhelmingly concentrated in one country. The intelligence is abundant; the muscle is scarce.
Regulators have noticed. Europe has moved to blunt the concentration risk with the Critical Raw Materials Act, which sets a 2030 cap on how much of a critical raw material can come from any single country. In practice that is an attempt to force diversification into a supply chain that has spent years consolidating in the opposite direction. Whether that target is achievable is another question entirely, given how few alternative refining operations exist at meaningful scale.
For the consumer, none of this is visible yet — there is no shelf full of home humanoids to buy. But the economics being set now will decide whether these machines arrive as expensive novelties or as products that can actually be manufactured at volume. The companies that lock down magnet supply, or engineer their way around it, hold a structural advantage that no amount of model fine-tuning can match.
It is a useful corrective to the current hype cycle. The demos show robots folding laundry and pouring drinks, and the instinct is to credit the neural network. The harder, less telegenic truth is that a humanoid robot is fundamentally a physical object built from constrained materials — and right now, the constraint that matters most is a magnet.
