Physical AI flooded weekend headlines: JPM upgrades Unitree rating, Nomura says robot training compute can be reduced 1000x

JPM this afternoon directly upgraded Unitree's rating from Underweight to Overweight, Nomura is discussing whether the compute required for robot motion control training can drop by three orders of magnitude, TF's Xiaoya organized a data collection expert meeting yesterday that everyone is discussing today (this can be viewed together with the Tesla expert discussion from a few days ago), and in domestic expert exchanges, the conversation ranged all the way from embodied brains to data collection, tactile sensing, and factory deployment...

Reading through it all, aside from the direction the sell side is giving, the biggest change is:

Only if these two things can be continuously achieved will the logic for robot volume ramp-up become more solid.

First, let's talk about why expectations have moved forward again.

The process by which JPM changed its judgment this time is very representative.

According to the report, when they previously visited the Industrial Fair, their attitude was still cautious. The robot hardware had already been built, but demand mainly came from R&D, developers, and small-batch pilots, and there was still a gap before sustained enterprise procurement.

Although management's judgment on technological progress has moved earlier, that does not mean commercialization has already been delivered. For the market, changes in timing alone will affect pricing.

If the capability breakthrough is still far away, the robots sold today will mainly show up as hardware revenue. If the breakthrough window is closer, robots that have already entered laboratories, developers' hands, and application scenarios may become the foundation for model testing, data collection, and continuous iteration (this is the same as what the Optimus expert said a few days ago, and next year Optimus will deploy 15,000 units for data collection).

This is also why JPMorgan is re-evaluating the value of the Unitree platform.

The sales adjustment in the report is very clear: the 2026 humanoid robot sales forecast is basically unchanged, still about 14,200 units; the 2028 forecast, however, was raised from about 41,600 units to 111,600 units.

This upward revision mainly bets on subsequent acceleration in volume ramp-up; current orders have not jumped in sync in the forecast.

Therefore, "the S-curve arriving earlier" is still a judgment that needs subsequent data verification.

Over the weekend, everyone began to re-estimate how long it will take and how much money it will cost for robots to truly do work.

A Nomura report today roughly means that, with technological progress, training GPU models for robots has now been reduced by 1,000 times compared with before...

This number is attractive because what it changes is the barrier to participation.

It introduces SuperMimic, which studies how robots learn body movements.

In the past, the common approach was extensive trial and error in a physics simulator. SuperMimic instead first learns a model describing the robot's own motion patterns, and then trains the controller within this model.

According to Nomura's account, after three hours of training on a single H100, its performance exceeded the previously released SONIC model checkpoint; the latter consumed about 21,000 GPU hours in training.

Image: Real Unitree G1 demonstration footage of SuperMimic, with the original video marked as 1x speed. Source: Qiayuan Liao, X, October 21, 2026; original video cover used.

But Nomura also said that a 1,000-fold reduction in compute cannot be directly expanded into a 1,000-fold reduction in the compute demand of the entire robot. What is compared here is the training compute for a specific motion control task, and the data volume and GPU type are not completely consistent either. The research mainly involves body movement on flat ground, does not solve object manipulation, and cannot be directly extended to complex industrial tasks.

It can be understood as: a path has emerged for greatly reducing the cost of robots learning how to control their bodies. As for how to do the work well, that is another question.

Therefore, algorithmic progress may bring two results at the same time: more companies can enter, and platforms that can provide stable hardware become more useful.

If models and training infrastructure become increasingly easy to access, whoever can provide effective data and understand customers' sites will be more likely to advance practical applications.

Image: Ultra robots performing warehouse packing tasks, with the original footage marked as 2x speed. Source: PI × Ultra demonstration, published in Leo Su, Substack, October 17, 2026.

Thus, data collection, the most discussed topic over the weekend, comes next.

Data may be more important in the robotics field, but the value of different kinds of data varies greatly.

Robots need to know more than just what is in the picture.

How much force is needed to grasp something; after touching it, whether it will slide, deform, or get stuck; after a failed operation, how to recover in the next step. This information is hard to fully express through ordinary video.

First-person video, hand trajectories, touch, and feedback from real machine operation each have their uses and their costs.

Image: Weave dual-arm robots handling clothing, with the original footage marked as 2x speed. Source: PI × Weave demonstration, published in Leo Su, Substack, October 17, 2026.

In Tianfeng's data collection interview, the expert mentioned a detail: wearing the device to collect for eight hours a day, after filtering, the truly usable data may be only one or two hours. This is the interviewee's empirical account, but it reminds us of a very practical issue: how long you record and how much model capability can ultimately be improved are two different things. So robot data has also taken a path similar to large models: whoever holds more data will have smarter robot models trained later.

A research visit on Optimus a few days ago illustrates the point well:

For future planning, the number of robots in the data collection field will continue to increase to drive the data flywheel.

Increasing the number of data collection robots is mainly to cope with more diversified application scenarios in the future, improve the precision of data collection, and provide compute and data support for continuously iterating large models.

Another expert's judgment on the industry's data volume was lower than the questioner's expectation. He emphasized that there is not yet full consensus on training methods and data effectiveness, and large-scale investment still needs to wait for verification. This expert's judgment is oriented toward the present, while the other expert and the Optimus expert are oriented toward the future.

Progress in the planning capabilities of large models may make real robots and application scenarios more valuable; cheaper motion control training may also make single-action capabilities easier to replicate. Subsequently, the gap will be more reflected in hardware consistency, manipulation capability, data quality, and deployment efficiency.

These issues will determine how quickly capability breakthroughs can turn into procurement.

This weekend the market became a bit more positive on physical AI, because changes worth tracking have indeed appeared on the implementation path. But the one-to-two-year breakthrough expectation, the 1,000-fold improvement in training efficiency, and future data demand each correspond to different scopes, and we need to keep tracking them...

Source of this article: Beifeng Muxiang Risk Disclosure and Disclaimer

Markets carry risk, and investment requires caution. This article does not constitute personal investment advice, nor does it take into account the special investment objectives, financial conditions, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article are suitable for their particular circumstances. Investment based on this is at the user's own risk.