A humanoid robot is mostly joints, magnets and gears. China makes almost all of them.

The machine that drew the crowd in Beijing this week could walk, pick up a bottle without crushing it, and — to the delight of a phone-camera thicket several rows deep — fold itself into a backflip and land upright. It is easy, watching that, to think you are looking at the future of artificial intelligence given a body. You are not, quite. The part of the robot that decides whether it can exist at all is not the model running somewhere in its chest. It is a component the size of a hockey puck, buried at each shoulder, hip, knee and wrist, and there are dozens of them. It is called an actuator, and almost nobody photographs it.

The week's news, though, was written in a different language entirely — the language of markets. On August 19, Unitree Robotics, the Hangzhou firm whose dancing and backflipping machines have become the public face of Chinese robotics, listed on Shanghai's STAR Market. It priced its shares at 150.8 yuan, about $22.34, valuing the company at roughly $9 billion. Then the market got hold of it. The stock rose by as much as a reported 500 to 600 percent on its first day of trading, and by the close the company was worth somewhere around $50 billion, with some tallies later stretching toward $66 billion — ahead, on paper, of Baidu and JD.com. DeepSeek, the AI lab, sat among the strategic investors. It was the first pure humanoid-robot maker to list on the mainland, and the number everyone quoted was the valuation.

The valuation is the narration. The bill of materials is the story. And the bill of materials, read line by line, keeps arriving at the same place on the map.

Count the joints

A humanoid robot is not one machine. It is forty to seventy small machines bolted to a frame and taught to cooperate. Each degree of freedom — each place the robot bends — needs an actuator, and an actuator is not a single part but a stack of them working in a space smaller than a fist: an electric motor, a gearbox that trades speed for torque, a sensor that tells the controller where the joint actually is, and the controller itself. Industry supply-chain analyses put the actuators at somewhere between 30 and 50 percent of the total material cost of a humanoid. The intelligence gets the keynote. The joints get the invoice, and the invoice is most of the machine.

This is the unglamorous truth that a backflip is very good at hiding. What makes a humanoid hard is not teaching it to think. It is building, at scale and at a price anyone can afford, the several dozen precise, powerful, compact joints that let it move without tearing itself apart. And every one of those joints is a small anthology of upstream dependencies — the motor needs one thing, the gearbox needs another, and each of those things is made somewhere, by someone, usually far from the company whose logo is on the chest plate.

• A frameless torque motor — the muscle — which needs powerful permanent magnets to hit the power density a walking robot demands.
• A precision reducer — the gearbox — usually a harmonic (strain-wave) or planetary unit, which converts a fast, weak motor into a slow, strong joint.
• A sensor stack — position and often torque — so the controller knows what the joint is doing to the millimetre and the newton-metre.
• A controller that closes the loop hundreds of times a second, because a humanoid that reacts slowly is a humanoid that falls over.

Take that list apart and follow each link one step further upstream than the press release does, and the story stops being about robots and starts being about materials, geography and concentration. That is where the real constraints always live.

The magnet problem

Start with the muscle. A high-torque actuator wants the strongest permanent magnets it can get, which in practice means neodymium-iron-boron — NdFeB, the same rare-earth magnet family that turns an electric car's motor. A humanoid, by some estimates, consumes more than twice the magnetic material of a single EV motor, because it has not one motor but dozens. Multiply that across a production run and the humanoid boom becomes, quietly, a rare-earth story.

And rare earths are the most concentrated chokepoint in the entire assembly. China accounts for something like 69 percent of the world's rare-earth mining and roughly 90 percent of the processing and refining that turns raw ore into finished magnets. That second number is the one that matters. Mining can, with years and capital, be opened elsewhere; the refining and magnet-making capacity is where the true bottleneck sits, and it has been deliberately built and held in one country. When Beijing adjusted rare-earth export licensing earlier in this cycle, the volatility rippled straight into every company trying to plan a magnet supply. None of this was designed as a weapon. It accreted, the way chokepoints do — one rational, cost-driven decision at a time — until the muscle of the world's humanoid robots depended on a material stream that runs through a single set of hands.

The country that threw Unitree a party on Wednesday is the same country that controls the parts every other company's robot needs. — On the geography of the humanoid bill of materials

The gear nobody photographs

Now the gearbox. The joints of a humanoid mostly turn on harmonic reducers — strain-wave gears, an elegant piece of engineering in which a thin, flexible steel cup is deformed by an elliptical bearing so that its teeth mesh with an outer ring at only two points, and each rotation of the input advances the output by a whisper. It is compact, it delivers enormous reduction ratios in a thin package, and it is exactly what you want in a wrist or an ankle. It is also fussy — prone to failure under backdriving loads, sensitive to manufacturing tolerance — which is precisely why it was, for years, a business almost nobody could break into.

For a long time the precision-reducer market was effectively a Japanese monopoly, dominated by a small number of specialists who had spent decades perfecting the metallurgy and the machining. That is the sort of quiet chokepoint that decides who can build a robot and at what cost — not the software, not the sensors, a gear. What has shifted the humanoid economics is that Chinese suppliers, Leaderdrive prominent among them, broke into that market and, by industry accounts, cut the cost of a precision reducer by something close to 40 percent. Cheaper reducers do not make the news. They make the difference between a robot that costs as much as a luxury car and one that a factory might actually buy by the hundred.

Why the price target runs through China

Put the pieces together and you can see why the entire industry's stated ambition — a capable humanoid at something like a $20,000 price — is, today, a Chinese supply-chain proposition. The cheapest path to the magnets runs through China. The cheapest path to the reducers runs through China. The cheapest actuator assembly runs through China. This is not a claim about Chinese robots; it is a claim about all of them.

Consider the flagship that is supposed to represent the American answer. By supply-chain analyses of Tesla's third-generation Optimus, roughly 70 percent of the component value is sourced from Chinese suppliers — with firms such as Sanhua and Tuopu named as actuator-assembly partners and Leaderdrive among the reducer sources. You can design a humanoid in California and write its behaviour in a data centre in Texas, and still find that most of what the machine physically is was made on the other side of the Pacific. The body has a nationality, and it is not necessarily the one on the badge.

That is the context the Shanghai listing sits inside, and it is why the valuation, however frothy, is not merely a meme. Unitree is not just a company that makes a charismatic robot. It sits close to the top of a supply chain that its own country dominates several layers down — the magnets, the reducers, the actuator assembly. When investors bid a robot maker above Baidu in an afternoon, part of what they are pricing, consciously or not, is home-field advantage over the physical guts of an entire emerging industry.

What the backflip didn't escape

None of this means the outcome is settled. Concentration cuts both ways, and it invites its own undoing. The same volatility that Beijing's export controls introduced is precisely the signal that pushes every other country to fund magnet refining, to underwrite a domestic reducer maker, to treat the humanoid bill of materials as a strategic problem rather than a procurement one. Chokepoints are strongest right up until the moment they frighten everyone downstream into building around them. That process is slow — measured in the years it takes to stand up a refinery or qualify a new gear supplier — but it has started, and the export-licence tremors are what started it.

And there are chokepoints even the Chinese supply chain does not escape. The controller in each joint, and the compute that runs the robot's perception, still traces back to the same leading-edge fabs in Taiwan that everything else in artificial intelligence depends on. The force-torque sensors that give a robot a delicate grip are their own small oligopoly. A humanoid is a stack of dependencies, and no single country holds every layer — which is exactly why the layers one country does hold matter so much.

So watch the valuation if you like; it will move, probably violently, and it will tell you mostly about the mood of the market. The number I would watch instead is duller and harder to find: the price of a precision reducer, the availability of a magnet, the share of a robot's parts that can only be bought from one place. A machine that can wave, walk and land a backflip is a marvel. It is also a supply chain standing upright, and most of that supply chain, for now, runs through a single country that has shown it understands exactly what it holds. The spectacle can be cloned by Friday. The magnet cannot.