Invented in America, built in China: the race for robot muscles

TL;DR

• The actuators (the motor and gears that move each joint) are half the cost of a humanoid. A typical one has about 29 in its body and 13 motors in its hands, together about 51% of what it costs to build (our estimate, the median of 36 robots).
• America invented them, and China makes them cheap. The precise robot gear, the direct-driveA joint with no gearbox: the motor turns the joint itself. It's smooth and easy to push back by hand, but it needs a big motor to be strong. joint and the modern magnet are all mass-produced in China now. Unitree builds a whole humanoid for $8,631.
• Magnets are the West's weak point, and the fix has a date. China separates about 99% of the heavy rare earthsDysprosium and terbium. A little of them in a neodymium magnet keeps it strong when it gets hot. China separates about 99% of them. that keep motor magnets strong, and new plants in France, Australia, Canada and the US start between now and late 2027.
• The hand is still the hardest part. The designs only just converged, and a 20-joint hand now costs $1,408 in China, but nobody makes a good one cheaply by the million yet.
• Factories are the next race. Today's capacity runs out in 2027 and 2028, and US rules (the FCC now blocks new foreign-made robots) protect whoever builds joints or hands at scale in the US.

In the spring of 2025, Tesla had robots with no arms.

China had just put export licences on dysprosium and terbium, the metals that keep a motor's magnets strong when they get hot. Shipments stalled while the paperwork caught up, and China wanted assurances that the magnets wouldn't end up in weapons. "They're just going into a humanoid robot," Musk said in April. This October he told the story as a joke: "that actually set back the Optimus program by like two or three months, because we had robots with no arms. We were like, 'A Farewell to Arms.'"

A humanoid holds about $62 of magnets, under 1% of its cost. And that was enough to take the arms off the most watched robot program in the world.

The robot's brain now comes from the frontier labs (The labs built the robot brain covers that side). Its body is motors, gears, screws, sensors and magnets, mostly packed into actuators (the unit at each joint that makes it move). With the hands, they're about half of what a humanoid costs to build. The key parts (the precise gear, the direct-drive joint and the strong magnet) were invented in America and Japan, decades ago. One after the other, they became cheap at scale in China. And for magnets, China's lead is so big (it makes about 94% of them) that one country's export licences could stop Tesla's arms.

A gear from the shoe industry

On 21 March 1955, an inventor called C. Walton Musser filed a patent for "strain wave gearing" (known today as a harmonic driveA gear where a flexible toothed cup, squeezed into an oval by a cam, turns inside a rigid ring. It slows the motor down a lot in one small step, with almost no play, so it's the standard in robot-arm joints.). The patent was assigned to the United Shoe Machinery Corporation, which made the machines of the shoe industry. (Musser had also designed the US Army's recoilless rifle in the Second World War.)

Here's how it works. A thin steel cup, squeezed into an oval by a cam, meshes with a rigid ring that has two more teeth. Each turn of the cam shifts the cup by two teeth, so a 200-tooth cup turns 100 times slower than the motor, in a single stage. The result is a lot of force from a small motor, in a compact gear with almost no play (no looseness between the teeth), and that's exactly what a precise robot arm needs.

Japan's Hasegawa Gear Works licensed it in 1964, and co-founded Harmonic Drive Systems (HDS) in 1970. The gears went to the Moon, in the wheels of Apollo 15's rover. Then the business sat "at a standstill" (HDS's own words) until around 1978, when industrial robots switched from hydraulics to electric motors. Every electric arm needed precise, compact gears, and HDS became the standard.

In 2011, Zuo Yuyu started Leaderdrive in China. At the time, Chinese robot makers paid HDS two to four times the Japanese price, and waited 10 to 14 months for delivery. Leaderdrive priced more than 50% below, and delivered five times faster. It went from 3,000 units in 2013 to 100,000 in 2018. Today Laifual (another Chinese maker) sells a harmonic gear for about $80, while HDS charges $420 to $560 in China. In August, HDS told its investors it expects its prices for AI robots to come under some pressure.

Humanoids are like 1978 again: a new market that needs millions of precise joints. The difference is that this time, China is there from day one.

Drone motors and a backflip

The other way to build a joint is to drop most of the gearbox, and that's how Unitree builds its humanoids. It costs Unitree RMB 61,278 ($8,631) to build one, and it sells them with a 63% gross margin (2025 numbers, from its IPO filing). Its joints use two small gear stages, which slow the motor down about 15-20 times (a harmonic drive does about 30 to 160). Jugu, a Chinese supplier, sells a whole joint for "hundreds of RMB", from a line that makes one every 30 seconds. And Leadshine made 120k robot motors in 2025, and expects 800-900k this year.

The idea goes back to 1981, when Haruhiko Asada and Takeo Kanade, at Carnegie Mellon, built a robot arm driven directly by its motors, with no gears at all. A new, much stronger magnet (samarium-cobalt) made it possible, because it let a motor push hard enough on its own.

Direct drive came back at MIT, in a lighter form. In 2012, Sangbae Kim's lab built the Cheetah robot's joints around "transmission transparency": a big motor with only a small gearbox (5.8:1). With so little gearing in the way, the motor's current tells you how hard the joint is pushing, so the robot can feel contact without force sensors. The joint is also cheap, and tough enough to survive landings. In 2019 Ben Katz, a student in the lab, built the Mini Cheetah's joints from hobby drone motors and one small gear stage, and the little robot did a backflip. The design was public, so the whole world copied it.

In Shanghai, a master's student called Wang Xingxing had the same idea on his own. From 2013 he built a robot dog, XDog, on drone-style motors with almost no gearing, for under RMB 20,000. He saw it as "the technological path to truly bring robots into the lives of the general public". In 2016 he left DJI (the drone maker, where he had gone to work after his studies), and founded Unitree in Hangzhou.

These low-gear joints keep proving themselves. KAIST's RAIBO2 (a robot dog) ran a marathon on one charge, using less energy per kilo and per metre than a human runner (0.25 against 0.37). Its team gives the credit to its force-transparent joints, the Cheetah idea (Nature, 23 September). And AthenaZero, RAI Institute's two-armed robot built on quasi-direct-driveA big motor with a small gearbox (often under 10:1). The joint stays easy to push back, and the motor's current still tells you how hard it pushes. It's the design behind MIT's Mini Cheetah. joints (the same idea again), throws at over 30 metres a second (about 110 km/h).

Screws, and muscles

Tesla took a third road: linear joints, which push and pull in a straight line like a piston. Optimus's legs run on planetary roller screwsA screw and nut with threaded rollers between them. They turn a motor's spin into a strong straight push, and they handle shocks well, which is why Optimus walks on them., which turn a motor's spin into a push (an 8 mm one is rated for 10 kN, about a tonne). Each Optimus uses 14 of them, because they take the shocks of walking better than ball screwsA screw and nut with small balls rolling between them. Efficient and precise, but they handle shocks less well than roller screws. do. Its reported main supplier, Xinjian, cut the price from RMB 1,652 to RMB 781 ($110) in 18 months.

A fourth road is still in the lab: artificial muscles. Their worst blocker was voltage (they needed thousands of volts, which is too much for a robot), and it fell this year: Stanford's ULTRA elastomer (a rubbery material) moved at 30 volts. None of them lasts a full work shift yet, but Clone's Torso 4, built on its own artificial muscles, is due in November.

Our cost model, built part by part, puts the actuators and hands of the average humanoid at about $8,000 this year (an average over all the humanoids built, from cheap Chinese ones to expensive Western ones), and $1,700 in 2033. The biggest part of that fall (about 40%) comes from assembly, testing, overhead and supplier margins, and those are exactly the costs that shrink when factories make a lot more units.

Why can't the West just make magnets?

It can, and it's starting. But today China makes about 94% of the world's magnets. And it separates about 99% of the dysprosium and terbium (the heavy rare earths). Rare earths come out of the mine mixed together, and separating them (pulling each metal apart from the others) is a long, difficult chemical process. China has used that power twice:

• 2010. After a Chinese trawler collided with Japanese coast guard ships, Chinese customs stopped rare-earth shipments to Japan for about two months (Beijing denied it). Japan answered with $250 million for Australia's Lynas, plus new mines, recycling, and magnets that use less dysprosium. Its dependence on China fell from 90% to 60-70% in 15 years. A Japanese official told the New York Times last December: "For Japan, this painful lesson came 15 years ago." The US and Europe are learning it now.
• 2025. In April, China put licences on seven rare earths and on the magnets made with them. Magnet exports halved within a month, a Ford plant in Chicago stopped for a week, and Optimus lost its arms. A much bigger package of export controls followed in October. That one is paused until 10 November 2026,¹A 28 September notice extends the wider US-China trade truce to 10 January, but it doesn't mention rare earths. but the April licences never were.

America gave this industry away 30 years ago. The strongest permanent magnet known, neodymium-iron-boron, was found in 1982-84 by General Motors and Sumitomo, separately. GM built a plant for it, Magnequench, in 1986. In 1995 GM sold Magnequench for $70 million, to a group that included two Chinese state-owned firms. The last US plant, in Indiana, closed in 2003-04, and production moved to China.

The West tried to rebuild once before, and the market killed it. The 2011 price spike paid for Molycorp's expansion of the Mountain Pass mine, in California (it spent over $1.7 billion). Then prices fell, and Molycorp went bankrupt in 2015 (MP Materials bought the mine for $20.5 million). This is why the Pentagon's 2025 deal with MP sets a floor price of $110 a kilo for NdPr (the magnet metals) for 10 years, so a price crash can't kill it again. And since 2023, China also bans exporting its separation and magnet know-how, so Western plants have to learn it by themselves.

Outside China, dysprosium oxide cost about $3,500 a kilo in September, 16 times the Chinese price. The new capacity outside China has dates:

• Heavy rare earths: Lynas became the first to separate them outside China (in Malaysia, in 2025), and it made 62 t of dysprosium and terbium in its last financial year. France is next. Solvay's plant in La Rochelle, which already separates the magnet metals NdPr, plans to separate dysprosium and terbium at industrial scale this autumn. Carester's new Caremag plant in Lacq (in the south-west), funded with €216 million from Japan and the French state, starts in late 2026: it recycles old magnets and separates heavy rare earths, up to 500 t of dysprosium oxide and 100 t of terbium oxide a year once fully ramped. Then four more plants start in 2027 (about 1,235 t a year together): Iluka in Australia, the Saskatchewan Research Council in Canada, Energy Fuels in the US, and a new Lynas plant. With Caremag and MP's own heavy-rare-earth line, about 2,040 t a year of separation is funded outside China by late 2027.
• US magnets: eVAC's 2,000 t plant in South Carolina opened in July, going from "soybean field to magnet production in less than 20 months". USA Rare Earth doubles to 1,200 t a year in early 2027. Vulcan Elements is building a 10,000 t plant in North Carolina. And MP's 10X plant comes in 2028 (taking MP to about 10,000 t a year), with a Pentagon guarantee that all its magnets find a buyer for 10 years.
• Europe: on 19 October, the EU picked Neo's 3,000 t magnet line in Estonia for funding.

Today magnets are the number one physical blocker for robot makers outside China, and heavy rare earths are the hardest part. But our view is that this gets solved much faster than Japan's 15 years, because robots need very little. Making a million robots a year takes about 2,000 t of magnets a year, against the 48,000 t of magnets the US used last year. And those million robots need only about 17 t of dysprosium and terbium (17 grams each), a small share of the 2,040 t a year of separation funded outside China. After that, from about 2028, the limit is mining enough ore.

Why is the hand still unsolved?

Because it took 40 years for makers to agree on a design, and they only just did. Nearly everyone now builds 20 to 25 joints per hand: Tesla (22, plus 3 in the wrist), 1X (22 plus 3), Sharpa, Unitree (22) and Inspire (China's volume leader, which moved to 22 joints on 30 September). A Chinese hand cost about RMB 50,000 two years ago, and on 12 October a Shenzhen company launched a 20-joint one at RMB 9,999 ($1,408). So now it's a manufacturing problem, and what's left to win is how long a hand lives, how hot it gets and how well it feels.

Tesla's Q3 call (21 October) showed how hard the first of these is. Optimus production has started on Tesla's high-volume line in Fremont, and in October its screw supplier said the new V3's hardware is "essentially frozen". But on the call, Tesla said that parts of the design are still changing, mostly the hand and fingers (which it called one of its most challenging hardware areas). Musk explained why: it's hard to make a hand that survives lots of wear and tear, because a human hand repairs itself and a robot's doesn't.

Where the motors sit decides most of a hand's life, heat and feel. It also decides whether the hand behaves the same in a simulator as in real life. (A hand often has fewer motors than joints, because one motor can move several joints through linkages or tendons.) There are three designs:

• Motors in the forearm, pulling tendons (Tesla, 1X, Shadow). The fingers stay slim, and the motors can be bigger and cooler. But tendons stretch and fray, and they're hard to simulate. Assembly is also slow. The Information reported that each Optimus hand and forearm has "more than 100 screws and other small parts that workers assemble by hand", and Tesla even lists night-shift technicians in Fremont just to repair hands. Figure's first hand used tendons too, and in October Brett Adcock (Figure's CEO) called it "one of the worst decisions we've made to date".
• A motor at each joint (Boston Dynamics, Inspire's new hand). There are no tendons to wear out, the fingers are modular, and the hand behaves in simulation the way it does in real life. The problem is heat: to hold a firm grip, a finger motor makes several times more heat than it can get rid of, even through a 100:1 gearbox. Berkeley's QUAD hand cut that heat by up to 96 times in September.
• Linkages and tiny screws (Inspire's 6-motor hands, still the volume leader). They're cheap (RMB 6,000-8,000), and they hold a grip with no power. But the screws lock the fingers, so you can't push them back. That's why these hands feel the least, and they're the hardest to simulate.

Boston Dynamics made the clearest choice. Its new Atlas hand, shown on 1 October, has four fingers and 13 joints, each moved by the same sealed module (Boston Dynamics calls it "highly transparent direct actuation", "built for sim2realTraining a skill in simulation, and then running it on the real robot without changes. transfer and RLReinforcement learning: the robot learns by trial and error, trying to get a better score."). It skipped the fifth finger, because "an extra finger means three more actuators", with more cost, more volume "and probability of something breaking". Its engineers even spent a day with their pinkies taped to their ring fingers, to check they could live without one. Alberto Rodriguez is Boston Dynamics' director of robot behavior, and before that he ran MIT's manipulation lab (the home of the GelSlimA thin touch sensor from MIT's manipulation lab: a camera watches a soft gel change shape when it touches something. touch-sensing finger). His summary: "Hands are a ruthless design trade-off."

The road here started in 1982, when Ken Salisbury built the Stanford/JPL hand: three fingers moved by cables. In 1986 the Utah/MIT hand moved 16 joints with 32 pneumatic pistons pulling on tendons. In London, a photographer with no engineering training, Richard Greenhill, started a hobby club in 1987 that became the Shadow Robot Company. For 15 years its Dexterous Hand (24 joints, with tendons pulled by 20 motors in the forearm) was the reference hand for research.

OpenAI used one. In 2018 its Dactyl turned a block in a Shadow hand after training only in simulation. The paper admitted that "problems with hardware breakage were one of the key challenges we had to overcome in this work." For the Rubik's cube in 2019, the team reworked the tendons with Shadow (they lasted 5 to 10 times longer after that), and built software to catch them as they broke. A year later OpenAI closed its robotics team. Its co-founder Wojciech Zaremba blamed the lack of data, but OpenAI's own papers had already named the other problem: the hand kept breaking.

The hand that behaves the same in the simulator and in real life will learn skills first, because it can practise in the simulator. Boston Dynamics' hand is designed so the simulator and the hardware agree. And skills trained only in simulation now work on real robots: an arm trained by UW and NVIDIA meshes gears and threads nuts, and WUJI's 20-joint hand spins a pen, with no real-world data for either. Whatever design wins, we expect the Chinese makers that already build most hands to make most of them. Their one risk is the tiny coreless motorsSmall motors with no iron in the spinning part. They're light, fast and efficient, so they're used to move fingers. Maxon and Faulhaber lead. that hands with a motor at each joint need, where Maxon and Faulhaber (Swiss and German) still lead.

Touch, finally

Robots got cameras decades ago, and touch only now. A human hand has about 17,000 touch receptors, and a fingertip about 241 per square centimetre. Robots took this long to feel for boring reasons: there was no common sensor (so no shared data), the skins wore out fast, and the wires had to run through moving knuckles. The fixes came one at a time. MIT's GelSight (2009) put a camera behind a soft gel skin. Meta's DIGIT (2020) made a cheap, open fingertip, small enough for a hand with several fingers. And in 2024, Meta's Digit 360 fingertip came with Sparsh, what Meta calls the first general-purpose model for touch (one model that reads touch from many different sensors, for many touch tasks). This year, touch started to scale:

• Fingertips that last. MIT's new fingertip lasts over 100 times longer than commercial gel films. Stanford built a 202-point skin from under $10 of materials. And Rhoda AI's 8 mm fingertip feels pressure, texture, vibration, slip and temperature.
• Joints that feel. On 16 October an MIT team inserted real pegs 73% of the time with no camera and no touch sensor, only from the position sensors (encoders) in the joints of a FlexivA robot-arm maker that puts force sensors in every joint of its arms. arm ("the joints of every arm and leg double as contact sensors"). A transparent joint (the Cheetah idea) gives the same feel through its motor current. Homes will mostly use those (safe, cheap, tough), and factories will use stiff joints with torque sensorsSensors in a joint that measure how hard it twists, so even a stiff joint with a lot of gearing can feel contact. (strong, precise).
• Touch as training data. TacClaw-Infinity, a new dataset, holds over 10,000 hours of vision, touch and action recorded together. And in Spark-0, a model trained with 1,000 hours of touch data from a glove, success went from 52.5% (trained without touch) to 71.7%, even on robots with no touch sensor (what it learned about touch helps it act better with cameras alone).
• Touch in the model that moves the robot. With contact sensing, PACE inserted parts 90.7% of the time across five real insertion tasks, more than 44 points better than with vision alone. TacZero, a vision-language model that reads touch, inserted pegs 15 times out of 20, against 10 without touch.

Contact is exactly where the frontier models fail today, so our view is that the hands that can feel will learn the fastest.

Can the factories keep up?

The factories built today can't, and the ones that can still have to be decided. We forecast 150k robots built in 2027, 400k in 2028, 1M in 2029 and 100M a year by 2033 (more than the world's car output today). In 2027 the world needs 4.8 million joint motors, and has 2.6 million of firm capacityFactory capacity that is built or funded (announced plans don't count).. It needs 334k hands and has 132k (although announced plans already reach 384k). And planetary gearsA small gear set: a central gear, smaller gears turning around it, and an outer ring. Cheap and tough when the motor only needs to be slowed down a little, and common in humanoid joints. (the most common gearbox in a humanoid's body joints) run short the same year.

Chinese suppliers build ahead of demand, and the demand is there: UBTECH expects 1,500 to 2,000 deliveries this year, against 13,361 preorders. The plants for 2030 get decided in 2027-29, and four things can make them come faster:

• Car-parts makers switch over. Wuzhou, which makes steering and brake screws for cars, has started trial production of robot screws (in a project that makes both).
• Faster thread-making. WhirlingCutting a screw thread with a ring of tools spinning fast around it. It's much faster than grinding the thread. a steering-screw thread takes about 60 seconds, against 7 minutes of grinding. The hard part is the thread inside the nut: it takes 2 to 3 hours on imported grinders, and those grinders are slow to arrive.
• New grinders and heat-treatment lines. Chinese machine makers are scaling up, and some Western ones have been bought by Asian groups (so part of that know-how is moving to Asia too).
• Robots building robots. In July, Linkerbot showed robot hands assembling robot hands. A humanoid that can run a 3D printer and assemble parts will build simpler robots first.

The West has two gaps: magnets, and suppliers. For now, its robot makers build their own parts. Tesla is hiring engineers for Austin's actuator lines (its job ad says "highly automated lines that produce electromechanical actuators", to run "from installation through sustained 24/7 production"). Agility makes its own cycloidal drivesA gear where a disc with rounded lobes rolls inside a ring of pins. Small, strong, and good with shocks., and OpenAI is hiring its own motor and actuator engineers. As volume grows, we expect them to buy joints from specialists, the way car makers buy parts from Bosch. Schaeffler starts formed harmonic gears (pressed into shape instead of machined) in Germany in 2027, more than 25% cheaper than machined ones, and it aims for hundreds of millions of euros of humanoid orders by 2030. Hyundai Mobis will supply Atlas's actuators. And US policy builds the market: since July the FCC blocks new foreign-made robots, and to count as US-made, a robot needs 65% US content (75% from 2029). From January, magnets for US defence can't contain any Chinese material.

What comes next

• 10 November: China decides on its paused October package of export controls.
• November: Clone's Torso 4, the first test of artificial muscles in a humanoid body.
• By mid-2027 (our forecast): a 20-joint hand under $1,000 in China, and under $500 by the end of 2028 (Leju's hand price fell 53% in a year).
• 2027: Schaeffler's gears in Germany, Tesla's Giga Texas Optimus plant, heavy rare earths separated outside China at scale, and Boston Dynamics building an Atlas in about an hour instead of a day.
• Late 2027: Tesla aims for its Fremont line to reach 20,000 Optimus a week (about a million a year). Musk expects that to take "at least a year, maybe a little longer", and it tests whether hand assembly can be automated. Tesla also plans its first deliveries to other companies around the end of 2027 (industrial work first).
• By the end of 2027 (our forecast): a US-owned company announces a plant for 100,000+ joints or hands a year, with a named humanoid customer (the FCC's 65% rule gives it a protected market).
• 2028: MP's 10X magnet plant (MP at about 10,000 t a year), and Hyundai's robot plant at 30k a year.

In the spring of 2025, $62 of magnets took the arms off Tesla's robots. Today a Shenzhen company sells a 20-joint hand for $1,408, and OpenAI is hiring people to design its own motors. The brain comes from the labs, and the muscles will come from whoever makes them by the million (today, that's China).