For about a year, the comfortable story in robotics circles went like this: China would win the hardware race — cheap motors, fast factories, viral robot dogs — while America would keep the part that mattered, the “brains.” The vision-language-action models that tell a humanoid what to actually do would stay a Western specialty, the thinking went, so the displacement clock for Western workers still had slack in it. China ships the bodies; America programs the minds.
That thesis spent June getting demolished on a public scoreboard.
The body half was never in doubt
The shipment numbers are lopsided to the point of absurdity. Chinese companies shipped roughly 90% of the world’s humanoid robots in 2025 — somewhere between 13,000 and 17,000 units by various analyst counts, a jump of nearly 480% over 2024, per Tech Times’ June 18 roundup. American firms shipped a few hundred units, combined. Unitree moved about 5,500 humanoids on its own; Shanghai’s AgiBot shipped 5,168, with research firm Omdia ranking it first globally. Behind them, Leju cleared 2,000, Galbot and UBTech each topped 1,000. Tesla, Figure AI, and Agility Robotics each shipped roughly 150 — and Tesla still hasn’t started public sales of Optimus.
And the money is following the volume. Unitree cleared the Shanghai exchange’s listing-committee review on June 1, becoming the first “embodied AI” company approved for China’s A-share market, targeting a ~$6.2 billion valuation. Its gross margins hit 60% in 2025 even as its average humanoid selling price fell from about $85,000 in 2023 to roughly $25,000 in 2025 — the cost curve and the demand curve engineered to drop together. None of that is news to anyone who’s been paying attention. We’ve written before about how much of the Chinese demo footage is more performance than function.
The brain half is the actual story
Here is the part that wasn’t supposed to happen. RoboChallenge is an independent platform that tests embodied-AI models on 30 physical tasks running on real robot hardware — object placement, tool use, food prep, multi-step manipulation. Not a simulation. As of now, no American model holds the top spot, and the lead has belonged to Chinese open-source models since January, when Spirit AI’s Spirit v1.5 — out of Hangzhou, the same town as DeepSeek — passed Physical Intelligence’s π0.5 and then open-sourced its weights so anyone could check the result.
Then June got worse for the comfortable thesis. Spirit’s follow-on, v1.6, topped RoboArena — a benchmark co-built by Nvidia with Stanford and UC Berkeley — scoring 1,924 against Nvidia’s own Cosmos3-Nano-Policy at 1,881. Nvidia’s flagship physical-AI model, launched at Computex on June 1, held the number-one position for two days before a Chinese startup knocked it off.
The honest caveat matters: the closed American frontier models — Physical Intelligence’s best, the labs that don’t post results — simply aren’t on these leaderboards, so we don’t know how they’d score. What we do know is that every public board a working engineer or industrial buyer can actually look at is currently led by China.
Why a job-watcher should care about a leaderboard
Because of the flywheel. A VLA model can’t be pre-trained on the internet the way a chatbot is — there are no billions of web pages showing a robot folding a shirt. The model needs real sensor traces from real machines doing real tasks, collected at cost and at scale. Which means the country that has deployed the most robots is automatically collecting the most training data, which trains better models, which makes the robots more capable, which gets more of them deployed. The “body” advantage and the “brain” advantage were never separable. China is running one race, not two, and winning on hardware is how it gets the fuel to win on software.
That’s also where China’s new homework assignment comes in. On June 10, the MIIT and the state-asset regulator told industry that 10,000 humanoids must be working — in factories, warehouses, hospitals, disaster response — by December 31. Not demoing. Working. Local governments file deployment plans by the end of June. Each of those robot-hours is more training data poured into the next model.
None of this means your job vanishes on Beijing’s schedule. The same week’s reporting is blunt about the ceiling: UBTech’s Walker S2 manages 30–50% of human productivity, and only on narrow tasks like box-stacking; batteries last two to three hours; Morgan Stanley’s analyst warns a shakeout is coming because production is running far ahead of paying customers. The robots are not yet good enough to clear the floor.
But the load-bearing assumption — that the West held the “brain” and therefore time — is the thing that just took a hit. The retraining clock for exposed industrial work doesn’t start when a capable humanoid clocks in. It started when the scoreboard stopped reading the way everyone assumed it would.
Sources
- Tech Times — China Ships 90% of Global Humanoid Units and Now Leads AI Benchmarks
- South China Morning Post — Chinese robotics start-up beats Nvidia in global AI ranking
- Fortune — China builds 85% of the world’s humanoid robots, and cheaply
- Rest of World — Inside China’s humanoid robot race: Unitree, AgiBot, Tesla Optimus
