参观宇树（Unitree）工厂后，我为何更看好西方机器人制造

Everyone thinks China's robot factories are automated. They aren't. The robots are Chinese; the robots that make robots will be Western.

On Wednesday I toured Unitree's factory in Hangzhou, and I wasn't allowed to take pictures. In the building that makes the actuators, the motors that drive every joint, there were one or two new automated lines. The rest of the floor was filled with humans operating old-school winding machines. (Winding means wrapping copper wire around a motor's core.) I don't know what percentage of the motors are made on the automated lines. But staffing the floor with humans made sense.

Unitree has a viral video (28 million views) of a humanoid doing assembly in its own factory, which it calls "an actual deployment". Its IPO prospectus calls it a pilot, and says assembly is "currently mainly manual". At the end of 2025 all of Unitree's machinery was worth about $1.7 million on its books, at a company that made almost 32,000 robots that year: some $50 of machinery per robot!

Unitree Embodied AI Model Manufactures Robots in Factory🤩 Based on Unitree’s UnifoLM-X1-0 embodied AI model, this is an actual deployment at Unitree’s own robot factory.

It doesn't need more. Labour was 8% of what it cost Unitree to make its robots in 2025, outsourced assemblers included, and most of the rest was parts. With those hands it made about 26,000 robot dogs in 2025, more than the roughly 20,000 IDC reckons the whole world shipped the year before.

So there's only so much a machine can save: halve a cost that's 8% of the bill and you save 4%. Unitree expects the motor line it's building to cut the labour in each motor by more than half, and the cost of one robot model by under 2%. Its labour per robot had roughly halved since 2023 (our arithmetic), and the filing credits scale, standardised processes and more skilled staff. None of those is a machine; two are kaizen, and I named my company, KAIKAKU, after the other word.

The other reason Unitree builds by hand starts in the showroom. Our host told us Unitree's CEO, Wang Xingxing, asked a G1 running GPT-6 to fetch him a bottle of water from another room. It did. Slowly. Unitree's own slide puts robots' "ChatGPT moment", when they can do 80% of tasks they've never seen 80% of the time, two or three years out. That slide is from last year, so it's as soon as next year.

Until then, humanoids are mostly toys for researchers: our host named Nvidia, Amazon, Meta and universities. Three quarters of Unitree's humanoid revenue came from research and education in the first nine months of 2025, and under 5% from factories or other industrial work (half or more of what Unitree files as "industry" is robots giving guided tours at companies!). When someone asked what the GD01, a rideable mecha from $650,000, is for, the answer was that Wang had wanted one since he was a child.

Unitree Unveils: GD01, A Manned Transformable Mecha, from $650,000 👏 The world's first production-ready manned mecha. It can transform. It's a civilian vehicle. It weighs ~500kg with you inside. Please everyone be sure to use the robot in a Friendly and Safe manner.

A product without a job changes every year, and a custom machine only works if the product is kind of frozen, so a machine built for this year's robot is scrap next year. It's like tailoring a suit for a teenager. Right now everything's ever-changing, except the motor. Unitree is automating its joint modules first because they're the parts its robots share most. The same Wednesday I toured his factory, Wang told a conference in Wuzhen that Unitree had spent ten years on the robot's body and would now put its weight on the brain. When I toured a drone factory, the circuit boards and motors were made by machine and final assembly was done by hand. In a drone the motor is a whole component; in a humanoid it's half of one, the copper-and-magnets half, and that half has stopped changing, like the AA battery in a toy that's redesigned every Christmas.

In May I said China's new factories "bolt robots into every one" and the West should "erase the factory floor". The first bit wasn't true of the plant I visited. For now Unitree is right. The second bit I still believe, and it's why I left Hangzhou more bullish on the West than ever. China's robots are made by cheap hands, not machines, so the machines are still up for grabs, and the West is good at building machines. AI makes it easier: a rented System 2 model can now do much of a robot programmer's job, so the West doesn't need China's army of engineers to automate the simpler work.

(The next day, at BYD, I saw the other end of this: a car plant where everything I saw was automated. More on that at the end.)

Ford's first line had no machine in it#

Ford is the best guide to what happens next, because he started exactly where Unitree is. His first moving assembly line, at Highland Park in 1913, put the magnets on Model T flywheels to make magnetos, small magnet-and-coil generators, kin to the motors on Unitree's winding floor. The work was split into 29 operations, each worker doing, in Ford's words, "as nearly as possible only one thing with only one movement".

The machines came once Ford froze the Model T, and by 1914 Highland Park was full of special machine tools. When the Model A replaced the T in 1927, Ford shut its plants for six months and rebuilt or scrapped most of its machine tools, and Chevrolet outsold it for the first time. A year later GM changed a Chevrolet model in three weeks. Unitree, with a new robot every year, is GM in 1927, not Ford. In 2011 Foxconn's Terry Gou promised a million robots within three years; by late 2016 Foxconn had about 40,000, and said it was slowing down "because humans are cheaper and more flexible to adapt to an ever-shorter life cycle of consumer products".

A robot is like a computer: most of the cost is the software. Programming and integrating a robot is typically 50 to 70% of the cost of the application, and every redesign means paying for the programming again. When GM put robots into its Hamtramck plant in the 1980s, they reportedly attached Buick bumpers to Cadillacs because they couldn't tell the cars apart.

When hands get too pricey, a factory has three doors:

• Organise them better, as Ford and Foxconn did in their heyday and Unitree does now.
• Find cheaper hands elsewhere.
• Automate the work away, which has always needed a frozen part (AI is ending that).

America shuts the cheap doors#

Cross the Pacific and Unitree's sums flip. An hour of factory work costs about six times as much in America as in China (our estimate), so at American wages labour would be about 40% of Unitree's costs, not 8%. In Hangzhou labour is the tip on Unitree's bill; in America it would be the main course. And moving won't help: Chinese hands are being shut out.

America has run this experiment before. In 1964 about half a million Mexican farmworkers were shut out. In 1965, 20% of California's tomatoes were picked by machine; by 1970 the figure was 99.9%. The tomato was bred to be tougher and squarer, so the machine could pick it. The product froze to fit the machine.

In July the FCC stopped approving new humanoids and robot dogs unless they're made in America with more than 65% American parts by cost (75% from 2029).

The Pentagon goes further for the drones its programme buys: from August 2027, their stators have to be wound outside China and a handful of other countries.

How the West wins#

The motor line Unitree expects to save 1.7% of one robot model's cost would save about 9% at American wages. The machine is worth five or six times more in America than in Hangzhou, so the West can afford one nobody can switch off.

In the West, the volume is arriving:

• Figure claims its BotQ factory has gone from making one humanoid a day to one an hour in less than four months.
• Tesla says most of a line to build a million Optimus robots a year is at its Fremont plant.
• Agility, which makes its Digit humanoids at its own factory in Oregon, listed on Nasdaq on 5 November with Foxconn leading about $201 million of the money, though it sold just $1.8 million of robots last year.
• 1X says it makes its own motors, from a spool of copper to a finished actuator, on automated machines its engineers designed and built in Hayward, California.

Assembly is not vertical integration 1X bends the copper, solders the phases, potts the stator and builds the magnet arrays. Raw materials to finished actuators on our automated line. That is vertical integration. That is reindustrialization

Somebody has to build those lines, and we in the West are extremely good at building the machines that build things, even if China passed Germany as the top exporter of machine tools last year. America itself can't yet: on 28 October the White House found that it "relies on foreign imports for most types of robotics, industrial machinery, and their parts and components".

The country that's short of hands usually builds the machines. Labour-short Japan did it to America: GM bought the first Unimate industrial robot in 1961, Kawasaki licensed it in 1968, and by 1990 America's first robot makers had been sold to Europe. Now America is short of hands, and the machine builders are moving in: Mazak has built 40,000 machines in Kentucky, Yaskawa will make robots in Wisconsin, and Fanuc has bought a site in Michigan.

Then there's the programming. Paying for it is why a custom machine only made sense for a frozen product, and AI is cutting that bill. HD Hyundai and Palantir say AI agents cut the average time to fix a robot-programming issue from 21.5 days to 7.9. Apple estimated that it would have taken up to nine months to hire 8,700 industrial engineers in America, versus 15 days in China. China will have the same AI tools, but an engineer's time is three to five times more expensive in America, so each hour AI saves is worth that much more. The advantage China will lose quickest is its army of robot programmers.

Software went through all this first. Big programs used to be planned like factories: a change took months of programmers' time, so teams froze the spec and shipped once a year. These days a coding agent makes the change before lunch, and products ship every week. In my last essay I called the robot version agentic robotics: an off-the-shelf System 2, a model like Claude, writes the robot's program, runs it and fixes it itself, on any task a sensor can check. Everything is a drawer except for laundry, and most of a motor line is drawers.

China makes the magnets, but MP Materials has finished qualifying its magnets with GM, and its Texas plant begins deliveries before the end of 2026. BMW makes an EV motor in Steyr that uses no permanent magnets, and wants capacity for more than 100,000 of them this year.

This is how we lose. By 2022, all of the top ten manufacturers of solar panel-making equipment were Chinese. China also makes many of the machines that build electric-car motors. Within a few years Chinese builders of hairpin winding lines caught up with German and Italian ones, with changeovers done in software, undercut them by as much as half, and have started exporting whole lines to Europe. The winder is a small part of a motor's cost, and American wages are high enough that even a winder at twice the price still saves 8 to 9% of the cost of a robot. What the West can't afford is one that's slow to change over.

Terrible, short-sighted move. We are shooting ourselves in the foot here. I am happy to spend time with policymakers to convince them of the idiocy of banning Chinese hardware in 2026, and how the US might regain a competitive edge in hardware. USA is not in the position of Show more

Yesterday, Congress took up the GUARD Act, bipartisan legislation to keep adversary-controlled robots out of American infrastructure. Our CEO Peggy Johnson wrote about why this hearing matters, in a new op-ed with @PirateWires. 🧵

Eric Jang thinks banning Chinese hardware is idiocy and wants an American Unitree instead. Agreed! But build it with machines. (Disclosure: I invest in Standard Manufacturing, which builds autonomous assembly lines, and in Dirac, Isembard, RMFG and Minerva, a US humanoid maker that gains from the FCC rule.)

Where the money is#

Tesla alone wants a million Optimus a year. Schaeffler counts 25 to 30 main-joint actuators in a humanoid, hands not included, so call it 25 million motors. The FCC rule says 65% of a humanoid's parts by cost must be made in America. And the machines that make motors? Mostly imported.

• Make motors, not humanoids, which are still sold mostly to research labs. Robots, drones and e-bikes will always need motors, and motors barely change from one generation to the next.
• Standardise like Mabuchi. It offered customers a 20% discount in 1971 to take its standard motors, and today it makes about 1.4 billion a year from about 90 basic designs.
• Build a stator line that can be quickly adapted to new sizes, because motor sizes still change.
• Don't build robots by hand in Ohio.

Hands win in China. Machines win in the West.

What I saw at BYD#

The next day I went to BYD, to a final-assembly plant two and a half hours east of Shenzhen. When BYD started making batteries in 1995, its founder, Wang Chuanfu, did Unitree's arithmetic, as CCTV tells it: an imported machine cost $200,000, written off at RMB 20,000 a month, enough to hire a dozen workers. "A person plus a jig equals a robot," he told CCTV. Thirty years later, everything I saw was automated: robots welding the bodies, and finished cars rolling off the line and onto BYD's own ship.

Nothing exotic. The robots I saw were Fanuc and KUKA arms on pre-programmed paths, similar to what Tesla uses, with grippers bolted together out of aluminium extrusions full of holes. Anyone can buy them, and nothing I saw is beyond an American plant, except the people. Fanuc sells the drills; BYD has the carpenters. BYD employs almost as many people in R&D alone as Tesla has in total.

So what are all those engineers doing in a factory run by robots, and could an agent do their job? That's part 2.

Braindump by Josef Chen