China Humanoid Robots Shipped 19,100 Units in Six Months. The Reducers Matter More.

Nineteen thousand one hundred units. That is roughly what the global humanoid robot market shipped in the first half of 2026, per Smart Analytics Global, and Chinese vendors accounted for more than 97 percent of it. AgiBot took about 8,400. Unitree about 5,900. Between them, three quarters of the world total. Those figures are real. They are also the least interesting fact about China humanoid robots this year.

The interesting part sits two layers down, in components that never appear in a launch video. Harmonic reducers. Six-axis force and torque sensors. Frameless torque motors. Planetary roller screws ground to a lead error of a few microns. Whether those are made in Changzhou or bought from Japan decides more about the next decade than any backflip.

I work in AI and automation and I live in Shenzhen. Eight years. The thing that changed on the ground was not the demos. It was the quoting.

The demographics stopped being a projection

China's population fell for a fourth consecutive year in 2025. The National Bureau of Statistics counted 7.92 million births against 11.31 million deaths, a net loss of 3.39 million people, closing the year at about 1.405 billion. The 16 to 59 cohort dropped to 60.6 percent of the population from 60.9 percent, roughly 6.62 million fewer working-age people in twelve months.

Now the labour side. NBS put migrant workers at 301.15 million in 2025, up 1.42 million. Average monthly income 5,075 yuan, up 2.3 percent. Average age 43.3 years, up from 43.2 the year before, and rising every single year I have been here. The share of migrant workers in the secondary sector, the people who staff factories, fell to 44.5 percent.

Read those three together. The total pool grew by half a percent. The factory-facing slice shrank. The people inside it got older and more expensive. No plant manager in Dongguan needs a five-year plan to work this out. They work it out in October, when the line does not refill after the holiday and the recruiter says the same thing she said last year.

The humanoid may be the wrapper. The actuator is the product.

What the robot density number actually says

The International Federation of Robotics published World Robotics 2025 on 25 September 2025. It counted 542,000 industrial robots installed worldwide during 2024. China took 295,000 of them, 54 percent of global installations and a national record. Operational stock inside China passed 2 million units, about 4.5 times Japan's. Global stock reached 4,664,000, up 9 percent year on year.

Then the density figures, published 8 April 2026, and this is where the story turns awkward for almost everyone who quotes it. Korea 1,220 robots per 10,000 manufacturing employees. Singapore 818. Germany 449. Japan 446. The United States 307. China 166, against a global average of 132.

Sixteen months earlier the IFR had put China at 470 for 2023, ahead of Germany and Japan, and that number travelled everywhere. It was wrong, or rather the denominator was. Updated labour data from the National Bureau of Statistics widened China's manufacturing employment base, and the density fell to 166, ranking 22nd globally and sixth in Asia.

I would rather quote the smaller number, because it is the one that supports the argument. China is not a saturated market that has run out of room to automate. On a per-worker basis it sits at roughly half the American figure. The 295,000 annual installations are not the top of the curve. They are early on it.

The number that actually moved is 57 percent

The IFR's 5 May 2026 note on China's national strategy carries the figure that matters most. Chinese robot makers supplied 57 percent of their home market in 2024, up from 30 percent in 2020. In electronics, 59 percent. In metal and machinery, 85 percent.

That is a supply chain changing hands inside four years. Fanuc, ABB, KUKA and Yaskawa did not stop selling into China. They stopped growing there faster than the domestic alternatives, in every segment where good enough at 60 percent of the price wins the tender. Which, in a country installing 295,000 arms a year, is most segments.

Which layers of the stack are actually domestic

Harmonic reducers: solved, or close to it. Leaderdrive reported 2025 revenue of 570.7 million yuan, up 47 percent, with net profit more than doubling to 124.4 million yuan. Suzhou Green Harmonic sold 425,200 harmonic reducers in 2025, up 72 percent, and JPMorgan estimated in March 2026 that it holds 30 to 40 percent of the Chinese market. Nidec read the same room and began mass-producing harmonic reducers in Pinghu, Zhejiang, in October 2024 rather than continuing to ship them in.

RV reducers: harder. Nabtesco still holds roughly 60 percent globally. Domestic suppliers have gone from close to nothing in 2018 to a real minority share of Chinese-assembled arms, but this is the layer where the honest answer remains partly.

Servo and motion control: Inovance, headquartered about forty minutes from my desk, holds roughly a third of China's servo system market and leads domestic SCARA at around 20 percent. Ten years ago that market read Yaskawa, Panasonic and Mitsubishi all the way down.

Sensors: split. Hesai and RoboSense lead global automotive LiDAR outright. Link-touch took over 70 percent of China's six-dimensional force and torque sensor market for humanoids in 2024. Rotary encoders remain about half foreign, and on precision encoders the spec sheet still quietly says Heidenhain.

Then the gap. Planetary roller screws, the part that turns motor rotation into the linear push inside a humanoid leg joint, were four-fifths imported into China as recently as 2023. Humanoid designs want C5 and often C3 precision, lead error measured in single-digit microns across 300 millimetres. Xinjian Transmission broke ground in Hangzhou in early 2025 on a plant scaled for a million units a year, which tells you both that the gap is real and that somebody has already decided to close it.

Here is the part you only feel by working here. Specify a harmonic reducer, a frameless torque motor or a six-axis force sensor in Shenzhen on a Tuesday and by Friday you have three quotes, two samples and a supplier you can reach by car before lunch. Specify a C3 roller screw and the tone of the call changes. Lead times go vague. Someone proposes a substitute. That asymmetry, quotable in days against quotable in quarters, is a more accurate map of the Chinese robotics stack than any market share chart I have seen.

I grew up around workshops in Nigeria where the binding constraint was never the cost of labour. It was the spare part, and the eleven weeks it took to arrive. Shenzhen inverted that one variable, and almost everything else about how this industry behaves follows from the inversion.

The strongest case against China humanoid robots

Take the sceptics seriously, because they are mostly right about the thing they are attacking.

John Koetsier argued in Forbes on 10 August 2026 that the shipment lead is misleading. American and European machines are at least as capable and in some cases better, and the volume gap reflects a decision to ship early rather than an engineering advantage. Western firms are holding units back until the product is reliable enough to survive a customer. That is a defensible strategy, not a defeat.

The data supports him further than he pushes it. Smart Analytics Global's numbers do not separate production deployments from research, education and performance units. Unitree's mix leans that way in particular, and at 39,999 yuan for an R1 launched in July 2025, a humanoid costs less than a used Camry. That is a university purchase order, not a capital equipment decision with a payback period attached. Some meaningful share of those 19,100 units is sitting in labs, in showrooms and in provincial innovation centres that were funded specifically to buy them.

The IFR, which is the industry's own trade body and not a neutral party, said as much on 5 May 2026: real-world production applications remain demonstrators and pilots, with commercialisation expected toward 2030.

So concede it. The humanoid shipment number is soft.

The thesis survives, because the humanoid number was never carrying it. The 295,000 industrial installations are audited, boring and repeated annually. The 57 percent domestic share is measured against foreign vendors who have every commercial incentive to dispute it and have not. Those are the load-bearing figures.

And the humanoid programmes still matter, for a reason with nothing to do with whether any humanoid ever earns back its cost. Volume commitments drag component prices down a curve. A frameless torque motor, a harmonic reducer or a six-axis force sensor ordered in humanoid quantities gets cheaper for everyone, including the industrial arm, the AGV and the collaborative cell that will actually do the work. China ran this play with lithium cells. It ran it again with automotive LiDAR, which went from luxury option to standard fitment in about four years. The humanoid may be the wrapper. The actuator is the product.

The bottlenecks that never appear in demo videos

Roller screws, covered above. Being fixed, slowly, with capital and patience.

Machine tools. China imports around 90 percent of its high-end machine tools, most of them from Japan. You cannot grind a C3 screw on equipment you are not allowed to buy. This is the constraint sitting behind the constraint, and it is the one least amenable to a five-year plan.

Compute and the model layer. Chinese robotics teams remain heavily dependent on NVIDIA's simulation and inference stack for training and deployment. Domestic simulators and embodied models are real and improving quickly, but the tooling gravity still points west.

Then the strange one. In April 2025 China imposed export controls on seven heavy rare earth elements and on NdFeB magnets containing terbium and dysprosium. Suspended for a year in November 2025. Tightened again for Japan-bound dual-use shipments in January 2026. An industrial servo motor uses roughly 300 to 500 grams of magnet. China's strongest upstream position in the entire robotics stack is the one it has chosen to make conditional, which buys leverage abroad and manufactures uncertainty at home for every company building robots here to sell somewhere else.

What the plan commits to, in writing

Embodied AI entered the 2025 Government Work Report as a named future industry. The 15th Five-Year Plan, covering 2026 to 2030, lists it alongside quantum technology and brain-computer interfaces. In December 2025 MIIT stood up a standardisation technical committee for humanoid robots and embodied intelligence, and by March 2026 had published a national standard system covering the industry's full lifecycle. A 2026 real-world training initiative targets more than 100 high-value scenarios and over 10,000 deployed units by year end.

Shenzhen moved first and moved local. On 3 March 2025 the city issued three action plans at once, including one for embodied intelligence robotics running 2025 to 2027, backed by a 10 billion yuan fund and a target of more than ten companies valued above 10 billion yuan each. There are over 57,000 robotics-related firms registered in the city, which is both an impressive number and a reminder that most of them will not exist in 2031.

The capital markets ratified all of it in August. The CSRC approved Unitree's registration on 1 July 2026, 104 days after filing, the fastest turnaround in STAR Market history. The IPO priced at 150.8 yuan a share, valuing the company at about 61 billion yuan and raising roughly 6.1 billion. Retail investors oversubscribed it by something on the order of 8,000 times. It listed on 19 August 2026 and closed its first session up several hundred percent.

Two things are therefore true in the same country in the same year. A ministry published a full lifecycle standards framework for humanoid robots in March, which is the document you write when you expect an industry to last thirty years and need the parts to interchange. Five months later, retail investors bid a company with roughly 1.7 billion yuan of 2025 revenue toward a fifty-billion-dollar valuation in a single afternoon. One of those is a bet on reducers. The other is a bet on the video. The reducers will keep shipping either way, and that is the only part of this I would underwrite.

Sources

IFR, World Robotics 2025: global robot demand in factories doubles over 10 years (25 Sept 2025): https://ifr.org/ifr-press-releases/news/global-robot-demand-in-factories-doubles-over-10-years

IFR, Robot Density Surges in Europe, Asia and Americas (8 April 2026): https://ifr.org/ifr-press-releases/news/robot-density-surges-in-europe-asia-and-americas

IFR, China Makes AI-powered Robots Core of National Strategy (5 May 2026): https://ifr.org/ifr-press-releases/news/china-makes-ai-powered-robots-core-of-national-strategy

Xinhua, Income of China's migrant workers continues to grow in 2025 (NBS data, 30 April 2026): https://english.news.cn/20260430/5598e3914ed140c990e4c8fec996c57b/c.html

Humanoids Daily, Global humanoid shipments surge 272% in 1H 2026 as AgiBot overtakes Unitree (Smart Analytics Global data): https://www.humanoidsdaily.com/news/global-humanoid-shipments-surge-272-in-1h-2026-as-agibot-overtakes-unitree

Jamestown Foundation, New Gains in PRC Robotics Software and Hardware: https://jamestown.org/new-gains-in-prc-robotics-software-hardware/

Caixin Global, Unitree to debut at 61 billion yuan valuation in closely watched Shanghai IPO (18 Aug 2026): https://www.caixinglobal.com/2026-08-18/unitree-to-debut-at-61-billion-yuan-valuation-in-closely-watched-shanghai-ipo-102475173.html

Forbes, John Koetsier: China's humanoid robot lead is misleading (10 Aug 2026): https://www.forbes.com/sites/johnkoetsier/2026/08/10/chinese-companies-are-winning-the-humanoid-robot-sales-battle-heres-why-that-doesnt-matter-yet/