Physical AI Timeline Moves Up: JPMorgan Upgrades Unitree Rating
nashnova research
JPMorgan upgraded Unitree Robotics from underweight straight to overweight, raising its 2028 humanoid-robot sales forecast from 41,600 to 111,600 units; Nomura simultaneously flagged a potential three-order-of-magnitude drop in locomotion-training compute, repricing the physical-AI timeline.
Why did JPMorgan flip its stance?
After visiting an industry expo earlier, JPMorgan remained cautious — hardware was ready, but demand sat in R&D labs and small-batch pilots, still far from sustained enterprise procurement.
The shift came at a Unitree physical-AI event: founder Wang Xingxing compressed the breakthrough window from "possibly as long as ten years" to one to two years. This means → JPMorgan is not betting on current orders but on an accelerating ramp later.
The bank revised its sales forecasts accordingly: 2026 stays at roughly 14,200 units, while 2028 jumps from 41,600 to 111,600. In plain terms = the near term is unchanged; the wager is that the S-curve arrives two years sooner.
The report itself flags that the "S-curve pull-forward" still requires follow-up data to confirm — orders have not yet jumped in tandem.
How can locomotion-training costs drop so sharply?
Nomura profiled the SuperMimic approach — a method that first learns a model of the robot's own dynamics, then trains a controller inside that model, rather than running brute-force trial-and-error in a physics simulator.
Key data point: a single H100 running for three hours produced a SuperMimic controller that outperformed SONIC's published checkpoint, which consumed roughly 21,000 GPU-hours. This means → compute for the same locomotion task may fall by roughly three orders of magnitude.
Nomura cautioned: a 1,000× drop does not translate directly to a 1,000× reduction in whole-robot compute needs. The comparison covers a specific locomotion task; dataset sizes and GPU types were not identical; the research involved flat-ground movement only, with object manipulation still unsolved.
What does a lower compute barrier mean for the industry?
Algorithmic progress may produce two parallel outcomes: more companies can enter the locomotion-control space, while platforms that deliver stable hardware become more valuable at the same time.
In plain terms = a lower barrier does not mean more winners — whoever holds a reliable hardware platform may actually gain leverage.
This reflects a shifting competitive axis: physical AI is moving from "who can train a controller" to "who can run trained models reliably on real machines."
Data collection: why is there no consensus on effectiveness?
A Tianfeng Securities expert interview revealed: eight hours of wearable-device capture per day yields only one to two hours of usable data after filtering. In plain terms = recording more does not equal using more; data quality matters more than data volume.
Tesla's Optimus takes a different route: it plans to keep adding data-collection robots to power a data flywheel, with roughly 15,000 units earmarked for data collection next year, targeting broader scenarios and finer capture granularity.
Other experts judge that the industry's current data volume falls below market expectations; training methods and data effectiveness lack consensus, and large-scale investment still awaits validation. This means → the two paths represent a split between "collect at scale, then filter" and "validate first, then scale."
Where are the verification checkpoints for physical AI?
The focus is shifting from "can it be built" to "how fast can it scale, and can data quality sustain iteration."
Three checkpoints sit on different timelines: the one-to-two-year breakthrough expectation maps to capability validation, the plunge in locomotion-training cost maps to a barrier-to-entry shift, and the data-effectiveness debate maps to the precondition for mass deployment.
Whether capability breakthroughs convert into actual enterprise purchases still depends on data being delivered at each checkpoint — this is an expectation repricing, not an order confirmation.
市场有风险，内容仅供研究参考，不构成投资建议。