Chinese investors are not fleeing artificial intelligence. They are redefining it. The STAR50 index, Shanghai's flagship gauge of AI and semiconductor companies, has lost roughly 40% in under four months, erasing approximately $480 billion in market value1,11. Over the same stretch, humanoid-robot maker Unitree Robotics has surged more than 65% in seven trading sessions, reclaiming a market capitalization of 271.8 billion yuan (roughly $40.8 billion)4,6. The divergence is not a contradiction. It is a rotation: capital is moving out of digital AI, the chips-and-software layer where export controls and demand doubts bite hardest, and into physical AI, the hardware-and-motors layer where China believes it holds structural advantages.
The catalyst for the robotics leg was a single conference appearance. Unitree founder Wang Xingxing told the World Physical AI Conference in Hangzhou on October 22 that large-scale AI models could produce a breakthrough in physical AI far sooner than previously expected3. "If everything goes well, it will take one to two years, otherwise it will take two to three years," Wang said. That compressed a timeline the industry had previously pegged at up to a decade9. Wang also said that OpenAI's GPT-6 Astra model can already direct robots without separate training.
JPMorgan Chase responded on October 26 with a rare double upgrade, lifting Unitree from "Underweight" to "Overweight" and more than tripling its price target from CNY 300 to CNY 9957,8. The bank raised its revenue forecasts, incorporating a 10–20% global market-share assumption for Unitree in 2028–2030. Unitree shares closed up 7.69% at CNY 672 on the STAR Market that day. The rally was not confined to one name: 23 robot-concept stocks hit their daily limit up on October 26, with 1.963 billion yuan in buy orders locked at the ceiling.
The STAR50 index closed at 1,315.50 on October 26, down 2.23% from the prior session and 26.39% over three months. The CSI 300 blue-chip benchmark sat roughly 20% below its June peak12. Chipmaker CXMT has fallen approximately 25% in the past month alone. Unitree debuted on August 19 at 1,100 yuan per share5 and traded at 405.10 yuan as recently as October 15 before the rally carried it back to 672 yuan2. Even after the surge, the stock remains well below its listing price, a reminder that the rally is a partial recovery rather than a breakout to new highs.
The bearish side of China's AI market has identifiable drivers. AI-related shares account for approximately 42% of China's total equity-market capitalization and 57% of trading value across mainland A-share and Hong Kong H-share markets, according to Goldman Sachs data cited by Reuters. That concentration made the market vulnerable once sentiment turned. Traders pointed to elevated valuations, tax burdens, and the risk that the United States could bar the use of Chinese-made data-center equipment. "The market has degenerated into a speculative frenzy," said Dong Baozhen, a Chinese fund manager. "What we are watching now is the bubble bursting".
Dong's framing applies at least as forcefully to the robotics rally he is not discussing. Unitree itself warned investors of valuation risks after the surge. Caixin Global reported that the company trades at a valuation far above its industry average and faces hurdles to commercializing humanoid robots. JPMorgan's own report listed risks including slower-than-expected adoption and pilot projects failing to convert into repeat orders. A rotation into robotics at elevated multiples could simply be moving the same speculative energy from one vessel to another, particularly when the company issuing the valuation warning is the one attracting the capital.
Yet there are structural reasons the rotation may prove more durable than a pure momentum trade. China's chip sector faces a tightening export-control regime that constrains access to advanced lithography and high-bandwidth memory, risks that sit outside any company's control. Robotics hardware, by contrast, relies on motors, actuators, and mechanical engineering where Chinese manufacturers have deep supply chains. Morgan Stanley estimates the Chinese humanoid-robot market at $2 billion this year, growing to $15 billion by 2030. China has made humanoid-robot development a strategic priority, and domestic companies have demonstrated robots capable of running, dancing, and performing martial arts. Chinese startup Qiyuan Robotics has raised daily production capacity to more than 200 units. Tesla confirmed that its Optimus humanoid has begun early production at its Fremont factory, with commercial deliveries planned around the end of 2027. That timeline validates the thesis that the humanoid-robot industry is approaching a commercial inflection, even as it introduces a formidable competitor.
The GPT-6 Astra angle deserves scrutiny. Wang's claim that OpenAI's model can direct robots without separate training, if accurate, would represent a step-change in how quickly robot makers can deploy capable systems. OpenAI's GPT-6 Astra is already appearing in academic research across multiple domains [ANNEX-R1][ANNEX-R2][ANNEX-R3][ANNEX-R4], and its pricing has dropped sharply since launch, with output tokens falling from $50 per million to $12.50 per million as of October 27 [ANNEX-P1]. Cheaper, more capable foundation models lower the software barrier for hardware-first companies like Unitree, which is precisely the bet JPMorgan is making: that the integration of AI software and hardware will determine the future of robotics companies.
The question is whether the physical-AI thesis can survive contact with commercial reality. Industry insiders cited by BigGo Finance say the robot sector has completed a trend reversal after a deep correction of 60–65% and is now in the latter half of the first wave of a new cycle. That framing implies significant upside remains, but the sector's own prior collapse, and Unitree's distance from its 1,100-yuan listing price, counsel caution about declaring a durable new trend.
Three markers will test the thesis. First, whether Unitree can hold above its mid-October level of 405.10 yuan once the JPMorgan-driven momentum fades; a retest of that level would suggest the rotation was a trade, not a reallocation. Second, whether Wang Xingxing's physical-AI timeline produces a demonstrable commercial deployment, not a conference demo, within the one-to-two-year window he specified. Third, whether the STAR50's decline stabilizes or accelerates if new U.S. export restrictions materialize; further chip-sector losses would test whether robotics stocks can decouple or get dragged down with the broader index. The money has moved. Whether it stays depends on whether physical AI can deliver revenue, not just rallies.