The Physical AI wave is surging, who is swimming naked in the A-share market?

On October 26, Guosheng Securities released the special research report "The Next Paradigm of AI: Physical AI and Robotics", and the Media and Internet Weekly also reminded that "the evolution of physical AI is strengthening". The concentrated release of optimistic industrial signals from institutions has directly pushed up the popularity of related themes in the secondary market.

However, at the industrial implementation end, there is still a practical gap between simulation and deployment in real physical environments. The market needs to clarify a core question: under the noisy industrial narrative, are the layouts of A-share listed companies in physical AI forward-looking investments for the next-generation industry, or just concept speculation that simply follows the market trend?

Dense catalysts in one week: Physical AI jumps from the screen to the workshop

After perceptual AI, generative large models, and AI Agent have completed iterations in sequence, the competitive boundary of artificial intelligence is jumping out of the digital space on the screen and turning to the real physical world.

The essence of physical AI is to enable agents to understand objective physical laws such as gravity, friction, and material deformation, and build a complete closed loop of perception-reasoning-execution. It should be distinguished that physical AI is a digital model that focuses on the deduction of physical rules, and is not equivalent to embodied intelligence that must have a physical entity.

Over the past week, global industrial catalysts have been highly concentrated. At the GTC conference in Berlin, Nvidia integrated the Cosmos 3 world model and the full set of Isaac robot development toolchains into a full-stack pipeline; AMD announced that it will acquire spatial intelligence enterprise World Labs for 8.2 billion US dollars; Tesla's Optimus is advancing the V3 mass production preparation, and the construction of the first production line has started, and OpenAI has clarified that the humanoid robot project will give priority to tackling the model brain.

At the domestic level, the 2026 Physical AI and World Model Conference was held in Nanjing, and Jiangsu Province issued a special action plan to start building the underlying infrastructure for physical AI datasets.

A series of events mark that the industry has moved away from pure concept demonstrations and entered an inflection point of small-batch delivery and repeated verification in industrial scenarios. However, the practical bottlenecks of the industry are also prominent: the hardware manufacturing capacity has basically taken shape, and the Sim2Real simulation migration deviation, the reliability of equipment working continuously for a long time, and the insufficient supply of high-quality physical interaction training data are still the core obstacles restricting commercialization.

Current market expectations are rising rapidly, but commercial orders for large-scale scenarios have not yet been released in batches. The mismatch between expectations and reality is the main internal contradiction of the physical AI industry at this stage.

Listed companies pour in: Hierarchical game in financial reports

Facing this round of technological transformation, A-share listed companies are clearly differentiated: some enterprises invest resources to promote product implementation, some only maintain pre-research on technologies, and some are classified into the track by the market but have not formed substantial business contributions.

Suochen Technology is an early domestic player that has laid out the physical simulation base. Relying on its own CAE industrial simulation capabilities, the company has built the "Tiangong Kaiwu" physical AI platform to provide a virtual simulation environment for robot training. According to the 2025 annual report, the physical AI related business achieved a revenue of 58.16 million yuan.

The 2026 semi-annual report shows that the company achieved an operating revenue of 72.74 million yuan, a year-on-year increase of 26.83%, the net profit attributable to owners of the parent company was a loss of 78.6554 million yuan, and the R&D investment accounted for as high as 116.59% of the operating revenue. High R&D investment supports the iteration of simulation products, but continuously squeezes current profits; the business is mainly delivered in the form of project-based industrial software, with a long customer verification cycle, and there is still a clear gap before achieving large-scale revenue through platformized and standardized products.

Orbbec enters the data entry link of physical AI from 3D visual perception hardware. Its 3D camera products can be connected to the simulation toolchain to provide spatial perception for robots, and has established supply cooperation with some humanoid ontology manufacturers. The data of the 2026 semi-annual report shows that the operating revenue during the reporting period was 438 million yuan, a slight year-on-year increase of 0.49%, and the net profit attributable to owners of the parent company was 41.64 million yuan, a year-on-year decrease of 30.82%.

The company clarified its positioning as a hardware supplier in public communications, and does not independently develop the underlying large model of physical AI. The technical adaptation of hardware is only a precondition, and the mass production ramp-up rhythm of downstream humanoid complete machine manufacturers directly determines whether this business can be transformed into sustainable revenue growth.

Not all companies labeled with the physical AI tag by the market have delivered substantial business growth. Green Depth reported that robot perception technologies are only implemented in the subdivided scenario of rail transit operation and maintenance, and the revenue proportion of related businesses is extremely low, which cannot form a substantial driving force for the overall operation. ZWSOFT reminds that there are clear differences between existing CAE simulation products and the physical AI technology system, and no embodied intelligence related implemented products have been launched yet. Although Tianyu Digital has carried out technical verification in the direction of spatial intelligence, its core revenue still comes from digital marketing and game business, and no commercial orders have been generated in the physical AI related sector.

The operating status of these companies reflects the common constraints of the industry: physical AI spans multiple chains including simulation software, perception hardware, and motion control, with high R&D investment and a long B-end project verification cycle.

At this stage, most listed companies can only confirm expense investment, and it is difficult to confirm large-scale business revenue. The project-based delivery model is inherently difficult to replicate and expand quickly. In addition, overseas giants have opened up simulation toolchains, so local enterprises are facing cost pressure brought by homogeneous R&D. Some enterprises rely on mergers and acquisitions to make up for technical shortcomings, and the accumulated goodwill that follows also lays a potential risk of subsequent asset impairment.