The Zhitong Finance App learned that Guojin Securities released a research report saying that the three costs are declining at the same time, and physical AI has an opportunity from 1 to 10. The bank determined that the layout window is currently better. The bank believes that in addition to humanoid robots, in addition to humanoid robots, robotics/non-humanoid robots also have good scaling prospects in industrial scenarios. For example, mobile robotics+commercial VLA models are already approaching production-level capabilities in logistics/manufacturing scenarios. Physical AI is expected to reshape manufacturing, warehousing and logistics/clothing industries, or take the lead. The singularity of physical AI has arrived. It is recommended to focus on core shortcomings on the one hand, and on the other hand, on the application of warehousing logistics/clothing robotics scenarios.
Guojin Securities's main views are as follows:
Physical AI, why now?
What is physical AI. The new paradigm that combines the laws of physics and AI is different from generative AI based solely on digital space. Physical AI implements “intelligence” as “action” through physical constraint models, and is a vehicle for AI applications to enter the second half and open up physical scenarios.
Why now. The model side has gone from 0 to 1: from initial experiments with large language models for task planning, to VLA extending semantic understanding from language to physical space, to continuous iteration around action generation, data efficiency, and reliability, with cascaded and hierarchical engineering architectures, VLA has moved from a laboratory to a real production line. Routes began to differentiate: in addition to deep iteration of a single path and differentiation at the level of system architecture, the “external model” route represented by Astra appeared. Instead, VLA was no longer trained; instead, the generic agent base model drives the classic controller through code generation and tool calls, changing robot control from a “model problem” to a “system problem.” Costs are starting to collapse: there is no need to retrain models for every scenario, no need to repeatedly collect data for new tasks, and no need to develop a generic “brain”. The three costs are falling at the same time, and physical AI has opportunities from 1 to 10. The bank judged that the layout window is currently better.
What should I vote for? A shortcoming is an opportunity
Data: the core bottleneck and opportunity. According to the China Academy of Information and Communication Technology's “Embodied Intelligent Training Ground Research Report (2026)”, assuming that the robot was to achieve the “ChatGPT” moment, it would take about 21,200 robots to work continuously for a year and produce tens of millions of hours of effective data. Currently, high-quality real data available globally is only hundreds of thousands to one million hours. Data, especially real machine data, will become the core bottleneck of physical AI and the best investment opportunity. According to the acquisition route, no main unit (first-view video, motion capture) and real machine (remote operation, autonomous operation return) will drive incremental hardware such as multi-camera, IMU, motion capture, EGO equipment, tactile sensors, etc., respectively, and benefit data producers, data infrastructure suppliers (tools, simulations, assets), and training ground operators.
Perception and execution: High-precision and dexterous operation is still a hard boundary, and the certainty of selling shovels is strong. Sensors/actuators have an irreplaceable position as essential products for physical AI implementation, and their requirements also resonate with the data acquisition process. Taking Gemini Robotics 2 as an example, the average success rate for gripper tasks has reached 74.2%-89.6%, yet the power due to multi-finger dexterous operation is greatly scattered (92% of the bulb is unscrewed, 36% is turned on), and the bottleneck directly points to tactile and force control hardware. It is recommended to focus on the three directions of sight, touch, and dexterous hands.
What do you think of the future? After the shortcomings are fixed, physical AI will open up new application scenarios
The main body will not be skipped, and robotics/humanoid and non-humanoid robots all have potential for release. The bank believes that in addition to humanoid robots, in addition to humanoid robots, robotics/non-humanoid robots also have good scaling prospects in industrial scenarios. For example, mobile robotics+commercial VLA models are already approaching production-level capabilities in logistics/manufacturing scenarios.
Physical AI is expected to reshape manufacturing, warehousing and logistics/clothing industries, or take the lead. As the shortcomings are gradually filled, the bank believes that opportunities will shift to the application side. In particular, the warehousing/logistics/clothing industry is expected to take the lead in upgrading through simple industrial automation due to a large number of flexible processes. Companies such as Agility/Figure in the field of warehousing and logistics have achieved small-scale deployment, and operation efficiency/accuracy is rapidly improving, and there is a quantitative foundation; the pain points of machine replacement in labor-intensive industries are clear. Jack Technology already has initial maturity in difficult processes such as flexible fabric processing, plans an annual production capacity of 10,000 robots, and has begun to obtain some intended orders, which is expected to accelerate in the future.
Risk Alerts
Large model technology iteration falls short of expected risk/data collection technology route change risk/terminal application falls short of expected risk.