The Body Problem: The AI Race Just Grew Legs

02The Mandate

In June, China's government issued industry a direct instruction: by December 31, 2026, 10,000 humanoid robots must be actively working — in factories, warehouses, hospitals, and disaster-response operations. Not demoing. Not running marathons. Working. It sits under the 2025 Humanoid Robot Action Plan issued by MIIT and five other ministries, which targets 100,000 deployed humanoids by 2027, and above that, the 15th Five-Year Plan (2026–2030), which places robotics at the heart of China's "modern industrial system" and explicitly directs the country's AI research toward physical applications.

Read that last part again, because it's the strategic tell. While American capital debates whether frontier chatbots can grow revenue fast enough to justify a trillion dollars of data centers — the subject of Part One of this series — Beijing has formally decided that the payoff from AI is not tokens. It's labor. The plan treats language models as an input and robots as the product.

03The Price Collapse

Here is where the token-war déjà vu gets uncomfortable. The thing that made Kimi K3 and GLM-5.2 dangerous wasn't capability parity alone — it was capability parity at a tenth of the price. The same curve is forming in hardware. Unitree's entry R1 humanoid lists at $5,900. Its flagship G1 starts around $16,000, with the average selling price disclosed at 167,600 yuan — call it $23,000. Boston Dynamics' Atlas sits north of $250,000. Tesla is promising $20,000–$30,000 for Optimus at production scale, which would be genuinely competitive — when it exists at production scale.

The mechanism behind the price is the Yangtze River Delta supply chain. Unitree builds its own motors, reducers, and sensors in-house; its component suppliers sit within a two-hour logistics radius. A humanoid prototype that takes 12 weeks to fabricate in the US or Germany turns around in Shenzhen in 10–14 days at a fraction of the cost. That's not a subsidy story — it's the EV and drone industrial base being pointed at a new product category. The same actuator suppliers, the same battery chemistry, the same contract manufacturers, the same iteration cadence that let BYD and DJI eat their markets.

04The Brain Went Open-Weight Too

If the hardware story rhymes with EVs, the software story rhymes with this series' opening act. On July 8, Ant Group's robotics arm open-sourced LingBot-VLA 2.0 under Apache 2.0 — a 6-billion-parameter vision-language-action model that runs the same trained policy across 20 distinct robot bodies from 17 brands without per-platform retraining, holds inference under 130 milliseconds, and beats Physical Intelligence's π0.5 across multiple hardware classes on the GM-100 bimanual manipulation benchmark. In January, startup Spirit AI's open-sourced Spirit v1.5 took the top slot on the RoboChallenge real-world benchmark — again over π0.5, the model from America's best-funded embodied-AI lab. AgiBot open-sourced its AgiBot World dataset. The pattern is identical to Kimi and GLM: give away the intelligence, commoditize your complement, monetize the thing you're structurally better at making — which in this case is the body.

China's embodied-intelligence market was worth roughly 915 billion yuan in 2025 (~$135B) and is projected near 1.09 trillion yuan this year. The American answer — Physical Intelligence's π0.7, NVIDIA's GR00T family, Figure's Helix — is scientifically serious and, so far, mostly proprietary. Which means the American robot brain has the same problem as the American chatbot: it has to be worth paying for, forever, against a free alternative that improves every quarter.

05The Flywheel Is the Weapon

The reason shipment volume matters more than demo quality is data. Embodied AI's structural bottleneck is that the physical world has no universal token — you can't scrape the internet for the feel of a warehouse tote. The only way to get manipulation data at scale is to have robots doing work, instrumented, in the field. LingBot-VLA 2.0 was trained on 60,000 hours of real robot data. Every one of the ten-thousand-plus Chinese humanoids fielded under the December mandate becomes a sensor package feeding the next model revision.

Morgan Stanley said it directly when it doubled its forecast: Chinese companies are leveraging large-scale deployments to generate the data needed to improve robot performance and accelerate commercialization. Deployment → data → capability → price cut → deployment. It's the flywheel in Figure 0, and the American strategy of perfecting the robot in the lab before shipping it is, functionally, a decision to let the other side spin it alone for another year.

06The Base Beneath the Base

Humanoids are the photogenic tip of a much larger iceberg. In conventional industrial robotics — the arms that actually build things — China installed 295,000 units in 2024, 54% of every industrial robot deployed on the planet, and its operational stock passed two million, about 4.5 times Japan's. The US installed 38,000 units in 2025, a healthy 11% rebound that the IFR rightly celebrated — and roughly one-tenth of China's run rate. For the first time, Chinese robot makers outsold foreign suppliers in their own home market, at 57% domestic share, up from about 28% a decade ago.

The honest wrinkle: on robot density — robots per 10,000 manufacturing workers — the IFR's 2025 revision actually dropped China to 166, well below the US at 307 and far below Korea's 1,220, after switching to what it considered more reliable Chinese employment data. Hawks should not skip that number, and doves should not over-read it: it reflects the sheer size of China's manufacturing workforce, not a lack of machines. A country installing ten times America's annual volume closes density gaps on a schedule, and A3's own president conceded China is "likely to advance quickly." But it is a genuine data point against the ten-foot-tall version of this story.

07What America Actually Has

The US position is not hopeless; it is concentrated. Frontier cognition still lives in American labs — nothing in the open-weight VLA world touches Fable-class reasoning, and the hard problems of long-horizon autonomy may ultimately be won by whoever owns the best general models. NVIDIA owns the training silicon both sides need. Atlas is genuinely in commercial deployment. Figure and Apptronik are demonstrating real autonomous work. And there is one asset that could rewrite the entire board: Optimus Gen 3, which Tesla intends to start producing at Fremont this summer, targeting $20,000–$30,000 — the only Western program even theoretically capable of automotive-scale volume. Tesla retired the Model S and X partly to fund it. If Gen 3 hits mass production on schedule, TrendForce notes it could restructure the global supply chain the way Tesla restructured EVs. That is a real scenario. It is also, as of this writing, a scenario about a factory line that hasn't started.

08Tripwires

As with the data-center thesis in Part One, the way to hold this loosely is to watch for specific events rather than vibes:

The through-line of this series has been that China's strategy is consistent across every theater: commoditize what America monetizes, then dominate the layer underneath. Tokens were the rehearsal. Bodies are the show. The intelligence was always going to escape the data center — the open question of 2026 is whether it walks out speaking English or Mandarin, and right now, only one side is manufacturing legs.