Themes that explode over the weekend often trap investors on Monday.
This weekend, the investment community was flooded with discussions about physical AI.
More than 50 online meetings focused on physical AI were held. Such intensity has not been seen in the robotics sector for a long time.
A new shift is emerging in this market rally: an increasing number of investors are turning their attention to the "brain" of robots.
In the past, market discussions on robotics primarily revolved around hardware components such as reducers, motors, ball screws, and dexterous hands. Market sentiment for related companies was largely determined by who could achieve mass production first, who entered the Tesla supply chain, and who secured orders from OEMs.
Recently, general-purpose large models have begun demonstrating the ability to understand robot tasks and arrange operational steps. This has sparked new imagination in the market: if robots can understand instructions like humans, could they take on more work in the future? If this direction holds true, the application scope of AI will extend from computers and smartphones to factories, warehouses, shopping malls, and even homes.
This also addresses investors' concerns regarding the long-term returns on AI. Over the past few years, global tech giants have continuously invested in computing power and model R&D, leading the market to question how much new space AI can open up beyond 2028.
Physical AI offers one possibility. For robots to operate in the real world, models, data, computing devices, and mechanical parts must work together. If these demands scale up, the AI industry chain stands to gain new growth momentum.
However, there is still a gap between technical breakthroughs and profitability. A robot understanding commands does not guarantee it can perform tasks stably; the market starting to purchase data does not mean data suppliers will immediately reap substantial profits.
The most significant change worth studying in this physical AI rally is the involvement of general-purpose large models in robot control. The subsequent questions are: Who can turn model capabilities into deliverable products, and which companies can derive sustained revenue from this?
The Hardest Hurdle for Robots
The development of autonomous driving in recent years provides a good reference point.
From 2023 to 2025, autonomous driving technology progressed steadily. Companies like Tesla continuously improved in recognizing road environments, automatically organizing training data, and upgrading onboard hardware, enabling cars to better handle complex traffic conditions.
This experience illustrates that for AI to enter the real world, algorithms, data, and hardware must advance in tandem.
The challenges facing robots are more complex. Cars mainly drive on roads, whereas robots need to grasp objects of different shapes, control force, adjust arm postures, and cope with unexpected situations such as object slippage or positional shifts.
The fact that a robotic arm can pick up a part at a fixed position does not mean it can continue assembly after the part shifts.
Advancements in general-purpose large models are changing how robots solve these problems.
Recently, tests involving general-purpose large models like GPT-6 Astra controlling robots have attracted industry attention. Related tests show that large models can already understand some complex instructions, identify objects, and plan next steps.
For example, when facing items on a table, a robot can determine what to pick up and where to place it based on instructions; if the camera view is obstructed, it can attempt to adjust its observation angle.
However, issues arise with fine manipulation. When asked to insert a key into a lock, the robot may know the goal and plan the rough steps. But if the key gets stuck at the edge of the lock hole, it needs to judge whether the angle is wrong, the position is off, or if it should retract and retry.
While the model understands the task requirements, the robot still needs to learn how to execute the actions well.
Researchers have begun trying to divide the work: letting the large model handle task understanding and decision-making on what to do first, while dedicated control systems manage arm movement and force adjustment.
Solutions like HexaShell and WorldScape-R²SI are exploring such division of labor. Test results published by research teams indicate that combining specialized motion planning and operation recording mechanisms improves robot performance. However, these results come primarily from controlled testing environments, leaving a gap before achieving long-term stable operation in complex factories.
This technological path is worth watching. Robots may not need to wait for an all-powerful large model to emerge before entering practical applications. As long as large models help robots understand more tasks, and dedicated control systems execute specific actions, some jobs have the opportunity to be automated earlier.
This also expands the application scope of physical AI beyond humanoid robots. Industrial robotic arms, warehouse robots, mobile devices, and service robots may all leverage stronger AI capabilities to handle tasks previously difficult to manage.
Listed companies have already started positioning along this path.
Inovance is one example. In its 2026 semi-annual report, the company disclosed that it has launched a seven-degree-of-freedom bionic arm and secured design wins for certain humanoid robot components; frameless motors and other products have achieved small-batch application in leading humanoid robot OEMs, and the bionic arm has completed customer end-unit verification.
A seven-degree-of-freedom bionic arm can be simply understood as a robotic arm capable of flexible movement in multiple directions, mimicking human arm motions. Its value lies in the fact that for robots to complete grasping, handling, and fine manipulation, they require a mechanical structure that accurately executes instructions.
The case of Inovance demonstrates that traditional industrial automation companies are extending their expertise in motors, drives, and controls into the humanoid robot field. However, small-batch applications and customer verification are still far from large-scale delivery. Whether these businesses can become new revenue sources remains to be proven by subsequent orders.
Estun represents another path. The company originally had industrial robot and smart manufacturing businesses, possessing mature industrial clients and engineering delivery experience. In the first half of 2026, Estun's industrial robot and smart manufacturing system business generated CNY 2.138 billion in revenue, a year-on-year increase of 2.21%.
The significance of such mature businesses is that robots are already generating revenue in real production environments. In the future, if AI enables robots to handle more variability and reduce manual programming and debugging time, traditional automation companies have the opportunity to integrate new capabilities into existing products.
Of course, Estun's current industrial robot revenue cannot be directly counted as incremental revenue driven by humanoid robots. Whether new technologies translate into new orders and profits still requires separate observation.
The "brain" of physical AI is progressing, and the body of robots is also upgrading. Next, a key factor determining the speed of industry development will be whether robots can accumulate sufficient experience from real-world operations.
Real Tactile Data Is the Scarcest Resource
Past robot training required teaching two things simultaneously: understanding tasks and executing actions.
General-purpose large models are expected to lower the difficulty of the first task, but robots still need to learn how to grasp, move, and control force through extensive real-world operations.
The internet has accumulated massive amounts of text, images, and videos. Models can learn common sense from these, but it is difficult to master real tactile sensations solely from such content.
A video can tell a robot how humans typically insert a key, but it rarely fully teaches it how to adjust when the key sticks to the left side of the lock hole, or whether to reduce force or retract and retry when resistance suddenly increases.
There is a vast amount of real operational experience between knowing how to do something and doing it well.
This changes the data requirements for robot training.
Currently, relevant data comes from several channels.
One type involves filming human operations externally to observe how people complete tasks; another uses head-mounted or wrist-worn cameras to record what the person sees and their hand movements; some lightweight devices record the process of human hands grasping, rotating, and moving objects. The final category of data comes directly from the robots themselves, recording their motion, contact, and operational results during execution.
These data types serve different purposes. Human demonstrations help models learn operational steps, while robots' own records provide feedback closer to actual execution.
Pricing also depends on data type and complexity. Feedback from parts of the domestic industry chain indicates that head-mounted operation videos and data collected via lightweight devices are quoted at approximately CNY 100 to 300 per hour; complex real-person operations and robot live-machine data may reach around CNY 1,000 per hour.
Truly valuable data still needs to be organized, calibrated, and screened. Collecting one hour of video does not mean all data from that hour is usable for training. Whether the data clearly records the action process, object states, and task outcomes determines if it can help improve model capabilities.
Repeatedly recording simple actions may not significantly enhance robot levels. Conversely, records of object slippage, jammed actions, and adjustments after failed grasps may be more useful, as they correspond to the hardest problems robots face in the real world.
Listed companies have started seeking opportunities in these areas.
Orbbec Inc is one representative. According to its 2026 semi-annual report, as of the reporting period, the company served over 1,000 robot customers globally, with visual solution sales revenue for the global robot market reaching more than double that of the same period last year.
3D vision helps robots determine object positions, distances, and spatial relationships. For a robot to accurately pick up a cup, it first needs to know where the cup is, how far away it is, and whether obstacles exist nearby. As robot tasks become increasingly complex, the importance of such visual technologies will rise.
However, Orbbec's financial performance reminds the market that there is still a gap between business growth and profit growth. The company's operating revenue for the first half of 2026 was CNY 438 million, a year-on-year increase of 0.49%; net profit attributable to parent company shareholders was CNY 41.6409 million, a year-on-year decrease of 30.82%. The company disclosed that increased R&D investment was one of the main reasons for the decline in net profit.
Rapid growth in robot visual solution revenue indicates that new markets are bringing business opportunities; however, the company's overall profit is still affected by R&D investment and other operational factors. Transitioning new businesses from growth to profitability requires more time.
Beyond vision is touch.
FuLai New Materials is positioning itself in flexible sensors and tactile data collection devices. Investor relations records from September 2026 show that its related products include multimodal tactile data collection gloves and fingertip sensors. The flexible sensing business has achieved batch delivery to some customers, but the revenue share of these businesses remains low.
These devices can record pressure and motion changes when human hands contact objects, helping researchers understand what happens during the grasping process. In the future, similar technologies may be used in robot hands, allowing robots to judge if they are gripping too tightly or if objects are sliding.
Keli Sensing is also developing force and tactile devices needed for robots. Its 2025 annual report disclosed continuous R&D around six-axis force sensors, joint torque sensors, and tactile sensors. Sales of robot mechanical sensors exceeded 1,500 units, with sample deliveries to over 70 customers.
Force sensors can be understood as the robot's "strength perceivers." They help robots judge how much force they are applying and whether resistance is encountered during operations. When grabbing a cup, the robot needs to avoid excessive force; when inserting a part, it needs to adjust actions promptly based on contact conditions.
Orbbec Inc, FuLai New Materials, and Keli Sensing occupy different segments: the former primarily provides visual perception, while the latter two focus on tactile data collection and force/tactile sensors respectively. Together, they reflect a trend: for robots to perform fine manipulation well, they need to perceive the surrounding environment and their own actions more comprehensively.
Growing data demand has also created new work for model developers. OpenAI's robot-related job postings involve data collection and testing projects, and robot companies like Figure are building their own data collection platforms.
However, the amount of external data model developers ultimately purchase depends on internal collection capabilities, data quality, and cost. Industry exchanges have included estimates that OpenAI might procure 5 to 10 million hours of data by 2027, and predictions that the global relevant market could develop towards the tens of billions of dollars level. These figures are currently industry calculations and cannot be treated as confirmed procurement plans or market sizes.
For data enterprises, the most important indicators going forward are actual contracts, customer repeat purchases, data acceptance rates, and revenue growth. For device enterprises, it is necessary to observe whether products gain customer approval and can transition from small-batch applications to continuous delivery.
Who Can Truly Make Money?
Physical AI may bring about a new industry chain: models need more data, data collection requires devices, and robot bodies need better perception, control, and execution capabilities.
However, commercialization progress varies across different segments, requiring investors to assess them separately.
First are data service providers.
If a company can consistently provide high-quality data to model developers and secure repeat purchases, it has the opportunity to become an important service provider for AI training. The competitiveness of such companies needs to be verified through revenue scale, customer stability, data quality, and profit margins.
Next are data collection device and sensor suppliers.
Orbbec Inc, FuLai New Materials, and Keli Sensing offer three observation samples: robot vision, tactile data collection, and force/tactile perception. Their products have the opportunity to serve robot training and may also enter robot bodies.
The potential advantage of these enterprises is that the same technology can address multiple application segments. However, data collection devices and robot bodies may have different requirements for product specifications, reliability, and certification. Orders from these two markets need to be validated separately.
The third type of opportunity comes from robot OEMs and core components.
Leaderdrive's operating revenue for the first half of 2026 was approximately CNY 349 million, a year-on-year increase of 38.64%, with harmonic reducer and metal component revenue accounting for approximately CNY 288 million.
Harmonic reducers can be simply understood as precision transmission components in robot joints, responsible for transmitting motor rotation more accurately to the joint. Robots performing fine actions require high precision, reliability, and cost-effectiveness from these components.
However, Leaderdrive's revenue growth stems from related robot demand and cannot be entirely attributed to humanoid robots or physical AI. As robot body solutions continue to evolve, suppliers still need to prove their competitiveness through customer verification and sustained orders.
Inovance, Estun, and Leaderdrive represent segments including automation systems, robot products, and precision transmission components respectively. They already possess their respective business foundations, but whether new technologies bring incremental revenue depends on subsequent product deliveries and customer purchases.
Therefore, investment opportunities in physical AI can be observed in stages.
In the short term, watch whether data service providers secure real orders and if data collection devices begin selling in batches; in the medium term, observe whether vision, tactile, and force products enter more robots and form sustained revenue; in the long term, look at whether robots can undertake more actual work, driving demand for OEMs and core components.
At the same time, beware of the market prematurely pricing in expectations. Improved robot test scores still need validation in real scenarios; increased data collection demand does not mean all data will command high premiums; growing robot shipments do not equate to sustainable profits for related companies.
Going forward, task success rates, continuous operation time, customer repurchase rates, order quality, gross margins, and cash flow will speak louder than single technical demos.
The GPT moment for physical AI may arrive gradually: large models help robots understand tasks, dedicated control systems handle actions, real data continuously supplements operational experience, and robots begin creating economic value in more scenarios.
Only when customers are willing to use robots long-term and related companies derive stable revenue and profits from this, will this technological boom truly enter the commercial realization stage.
Currently, the market has started pricing in this future. What needs to be verified next is whether companies can meet these expectations with products, orders, and financial reports.
Epilogue
The overwhelming buzz over the weekend, with more than 50 online meetings on physical AI, sufficiently demonstrates its heat, but heat itself is also a risk.
According to historical patterns in the A-share market, themes that explode over the weekend often trap investors on Monday.
After all, the model roadmap has not yet truly converged. Whether data collection can become the trading theme throughout this phase depends on subsequent order fulfillment and performance realization. Thematic investing fears most jumping in when consensus is at its peak. Short-term volatility will not be small.
But zooming out, this is indeed a main line worth taking seriously. When a general-purpose large model can direct a robotic arm on a test set without specialized training; when the industry can boost accuracy by three or four times within a month using the division of labor of "brain decision-making + specialized model execution"—the thin veil over robotics may truly be pierced.
In the past, we taught robots both to "think clearly" and to "move deftly." Now, general-purpose models have taken over the "thinking clearly" half; all subsequent competition focuses on that simple phrase: how to let machines learn the subtle tactile sense that only human hands-on practice can teach.
Whoever masters the memory of that "hand" holds the first priced key to the era of physical AI.