China controls 94% of the permanent magnets in a humanoid robot

Every motor in a humanoid robot contains a permanent magnet. Every permanent magnet in a humanoid robot almost certainly came from China.

Incredibly, that's not hyperbole. 

According to a September 2026 supply chain review by Rare Earth Exchanges, China controlled 91% of refined magnet rare earths in 2024 and 94% of sintered permanent magnets ... the specific components that go into the actuators functioning as a robot's muscles and joints. As the humanoid industry moves from prototype into production phases, that dependency has gone from an abstract concern to an operational one.

US manufacturers are now warning that actuator and rare-earth supply gaps could triple production costs. That's literally where sourcing teams are actually landing when they try to build supply chains that don't run through China.

The actuator problem is the robot problem

Jeff Cardneas, the CEO of Apptronik, flagged significant sourcing challenges around actuators publicly this September. He's not alone. Actuators — the motor-and-gear assemblies that let a robot bend a knee, grip a part, or absorb a fall — are the most mechanically demanding components in a humanoid system. They require high-performance rare earth magnets to hit the torque density and efficiency specs that make a robot practical rather than theoretical.

There's pretty much no short-term substitute.

Sourcing neodymium, dysprosium, and the other rare earths that go into these magnets from non-Chinese suppliers is possible in principle, but nearly impossible at scale in the near term. The processing infrastructure — the refineries, the magnet manufacturing facilities, the supply chains — took China decades to build. Western efforts to replicate it will be measured in years or decades, not quarters.

Tesla was in Ningbo on September 17

Tesla's supplier audit in Ningbo on September 17 is a microcosm of this story. 

The audit reportedly examined Tuopu Group, Sanhua Intelligent Controls, and Joyson Electronics for Optimus production readiness. All three are Chinese suppliers, of course.

For every Western humanoid company, the path to production runs through Chinese component suppliers. I've written about diversifying that supply chain before in Forbes, and there are some other options, including just in the United States:

Of course, it’s also true – as you see in the McKinsey report referenced above – if your actuators require rare earth magnets, China has most of the global supply. And, clearly, it is getting increasingly selective in how it exports those to the rest of the world. Here too, however, there are developing alternatives. MP Materials is building domestic magnet capacity with Department of Defense backing, Niron Magnetics is commercializing rare-earth-free magnet materials and Evolution Metals & Technologies is installing magnet-making machines targeting 10,000 metric tons of annual capacity largely from “urban mining,” AKA recycling.

But the problem might be harder than I estimated at that time. And diversifying that supply chain is a multi-year project, not a 30 days or 3 months thing. That's a massive problem, because the production ramp is either happening now or imminent.

The GPU side of the equation

While the rare earth problem constrains the hardware side, the compute side has its own scale challenge. 

Figure IA and Nscale are mobilizing $3.5 billion for up to 100,000 NVIDIA GPUs to support humanoid training infrastructure, with deployment targeted for late 2027. That's the compute required to move from teleoperation-trained models to genuinely autonomous robots at scale, apparently.

(Other companies have different ideas, of course.)

The GPU spend tells us something important: the training data and model infrastructure required to run a commercial humanoid fleet is capital-intensive in ways that go well beyond the hardware cost. A $50,000 robot might need millions in training compute behind it before it does useful work reliably.

The Unitree problem

On the Chinese side, the supply chain review notes a different kind of pressure. 

Unitree Robotics experienced significant stock losses as Chinese regulators informally slowed humanoid IPOs and began scrutinizing government-backed revenue quality. The regulatory uncertainty affects the companies that Western supply chains depend on.

It's an uncomfortable loop: Western manufacturers are dependent on Chinese component suppliers, but the Chinese component market itself is subject to regulatory interventions that Western companies can't predict or control.

And don't forget: many Chinese robots use western chips, although similar to US and European efforts to circumvent the Chinese supply chain, there are domestic efforts to use Chinese chips only.

The magnetic metric

Rare Earth Exchanges uses an interesting evaluation framework: cost per productive autonomous hour, and shipped robots multiplied by verified magnet content. Not announced factory capacity, not fundraising totals, not rendered videos.

Announced capacity is cheap. 

Verified magnet content in shipped units is not. 

The gap between those two numbers is where the supply chain problem lives. China installed 354,000 industrial robots in 2025: 59% of global volume. The Chinese robotics industry is not just large; it's deeply integrated with the upstream raw material and component supply chains that the global humanoid buildout requires. Western companies building humanoids are often building on Chinese foundations whether they want to or not.

How to change this

The honest answer is: this doesn't change much in the near term. 

MP Materials' rare earth processing facility in Mountain Pass, California is the most serious Western effort at upstream supply chain independence. The Department of Defense has funded rare earth refining initiatives. The EU Critical Raw Materials Act is trying to mandate domestic sourcing. 

But arguably none of these in coming online fast enough to matter for the 2026–2028 production ramps the industry is executing.

For the next two to three years, the humanoid companies most exposed to supply chain risk are the Western ones trying to scale without Chinese component relationships. The ones least exposed are the Chinese companies that have those relationships embedded. And the ones in the middle — companies like Tesla and Figure that are explicitly building Chinese supply chain dependencies into their cost models — are making a bet that geopolitical disruption stays manageable while they scale.

That bet might be right. But it's a bet, not a plan. And the Rare Earth Exchanges review makes clear: it's a risky bet, for sure.

Sources

• Rare Earth Exchanges — Humanoid robotics supply chain September 2026 review