Multiple Catalysts Converge as A-Share Humanoid Robot Stocks Surge to Limit-Up in Afternoon Trading
nashnova research
A-share humanoid-robot stocks surged to limit-up across the board on the afternoon of October 26, driven by JPMorgan's double upgrade of Unitree Robotics, a sharply accelerated timeline for physical-AI breakthroughs, and a production milestone — putting the sector squarely in a valuation-repricing window.
How broad was the rally?
Shangwei New Materials hit limit-up at ¥151.57, up 20%, with market cap topping ¥61.1 billion. Unitree Robotics surged over 18% intraday, breaking through ¥700.
Fenglong Shares and Dafeng Industrial locked limit-up in the afternoon session; Mould & Die Technology, Jack Technology and others also hit the ceiling.
Hong Kong-listed names rallied in tandem: Camsense up over 16%, Woan Robotics up nearly 13%, Dobot up about 9%, Hesai-W up more than 8%.
Why did JPMorgan jump two notches on Unitree?
On October 25 JPMorgan upgraded Unitree Robotics from "underweight" straight to "overweight," raising its target price from ¥300 to ¥995 — roughly 59% above the October 23 close of ¥624.
The bank kept its 2026 humanoid-robot sales estimate at about 14,200 units but lifted 2028 from roughly 41,600 to about 111,600 units, implying global market share rising from 8% to 21%.
This means → JPMorgan is not betting on near-term shipments but on Unitree capturing a fifth of the global market by 2028 — revenue is projected to jump from ¥2.82 billion (2026) to ¥13.38 billion (2028), a year-on-year growth rate of 173.8%.
JPMorgan also introduced a "scarcity premium" into its framework: new embodied-intelligence listings — companies building AI that operates in a physical body — may come slower than expected, making Unitree "one of the few direct embodied-AI investment proxies" on the A-share market. The bank values it at 30× estimated 2028 price-to-sales.
Why has the physical-AI breakthrough window suddenly moved forward?
Unitree founder Wang Xingxing said publicly that physical AI "is showing early signs of being cracked," with a breakthrough potentially arriving in one to three years — far sooner than the earlier estimate of "up to a decade."
In plain terms = the industry used to assume truly capable robots were a distant prospect; the founder's own words just compressed the market's patience window dramatically.
JPMorgan cited this statement in its report, arguing the industry's S-curve — the adoption path from slow uptake through rapid growth to plateau — may shift forward. The bottleneck is moving from high-level reasoning to "real-world execution, data generation, and deployment iteration."
What does Nomura's SuperMimic pathway mean?
SuperMimic first trains a model that describes the robot's own motion dynamics, then trains a controller on top. According to Nomura, after just 3 hours on a single H100 GPU, it outperformed SONIC's previously published model checkpoint.
This means → a level of performance that SONIC reached using roughly 21,000 GPU-hours was matched with a fraction of the resources — a roughly 1,000× reduction in compute for the specific motion-control task tested.
Nomura cautioned that this figure applies only to one motion-control task and cannot be directly extrapolated to the full compute needs of an entire robot. In plain terms = it proves "working smarter" can dramatically cut compute costs, but making the whole robot intelligent is a different scale of problem.
Where does mass production stand?
Qiyuan Robotics rolled its 5,000th personal robot off the line — just one month after its September 20 product launch.
The production line now delivers one unit every 2.5 minutes, with daily capacity rising from roughly 10 units to over 200.
User data: some owners average more than 1 hour of daily use, heavy users exceed 3 hours, and daily effective interactions top 150. This reflects a shift from "buy it and forget it" to genuine high-frequency engagement.
On October 24 Qiyuan also launched its "Physical AI Qizai Program," opening its robot development platform, toolkits, and compute resources to outside developers.
Can this valuation repricing last?
Soochow Securities noted the industry's center of gravity is shifting from hardware to the model layer: the technical roadmap has moved from VLA (Vision-Language-Action models) to WLA (World-Language-Action models), and accelerating physical AI is completing a data flywheel — "hardware deployment → real-world physical interaction feedback → model training."
Soochow argued China moved early on the model side and holds a first-mover advantage in data accumulation, supported by a mature component supply chain — Unitree's in-house motors, harmonic reducers from Leader and CNDYD, and ball-screw supply from Hengli Hydraulic and Zhejiang Rongtai.
This means → multiple catalysts have converged in the same window, driving extremely elevated short-term sentiment. But JPMorgan's earlier caution on commercialization still stands — whether industrialization can deliver on the newly accelerated timeline is the critical test for whether this valuation repricing holds.
市场有风险，内容仅供研究参考，不构成投资建议。