The Zhitong Finance App learned that Guojin Securities released a research report saying that the core changes in this round of Physical AI are not a breakthrough in the mechanical structure of robots themselves, but rather that the general model has begun to move further from perception and understanding into the space of planning, decision-making, and action, extending AI from the digital world to the physical world. As model output gradually expands from language and images to actual motion, the robot's hardware requirements are also further expanded from a single AI computing power to complete end-side system capabilities such as multi-modal perception, local computation, memory/storage, real-time control, and functional safety. At the same time, physical AI is not an isolated robot hardware route: mobile phones and PCs have taken the lead in verifying large-scale deployment of local models and agents, while robots have further extended the same trend from digital interaction to physical execution. The bank believes that the transition in model capabilities is the technical starting point for this round of Physical AI. Robots are the most typical new terminals, and the upgrading of end-side computing platforms is a broader industrial landing point for the electronics industry.
Guojin Securities's main views are as follows:
The model enters the physical world and promotes the upgrading of robot end side hardware from a single computing power to system level expansion
As robot tasks expand from fixed stations and single operation to open environments, full body movement, and dexterous operation, models need to continuously obtain multi-dimensional information such as vision, posture, movement, and touch, and complete sensory integration, reasoning, planning, and control locally. In terms of corresponding hardware, the increase brought about by Physical AI is not simply an increase in TOPS, but rather a “large model+multi-sensor+real-time control” to jointly promote the expansion of the entire end-side data link: the increase in model size drives memory capacity and bandwidth requirements, multi-channel sensing and data recording increases local storage and high-speed I/O requirements, while the action output further increases real-time control interfaces and functional safety resources such as EtherCAT and CAN-FD. The value of robotic electronic hardware is spread out of the SoC to memory, storage, interconnection, and control processes.
Robots enter small-batch industrial verification, and training demand is expected to be released on a large scale before the main body
The 2026H1 global humanoid robot has shipped close to 25,000 units, with shipments in half a year exceeding the full year of 2025, and the industry has gradually moved from a few thousand R&D markets to tens of thousands of units of commercial exploration. At the same time, the application structure has also begun to spread from scientific research, display, and data collection to real work scenarios such as manufacturing, logistics and services, and demand for electronic components is expected to shift from project-based procurement to standardized and batch procurement. What is more noteworthy is that model training requires the accumulation of a large amount of real interactive data in advance, and no entity collection can reduce dependence on the number of robots. Product forms such as EGO, UMI, WristCAM, and Hub have begun to evolve from self-built laboratory tools to standardized training hardware. The demand for electronic hardware does not have to wait completely for large-scale shipment of the robot itself, which may form a release rhythm of “training needs first, body requirements accepted”.
The model sinks at an accelerated pace, and Physical AI is further integrated into a broader end-side computing upgrade cycle
MLPerf standard test targets have been further expanded from traditional visual CNN to LLM, generative AI and agents, while mobile phones and PCs have taken the lead in verifying that large local models enter large-scale terminals; robot VLA synchronously migrates to the end side, and local models have begun to directly undertake motion output tasks. As LLM, VLM, and Agent run locally, model weight, context, KvCache, and local knowledge base continue to increase, and end-side hardware constraints are further extended from computing power to memory capacity, bandwidth, and storage; at the same time, small models, quantitative compression, and deployment tool chains continue to mature, and are also reducing the computing power, power consumption, and storage thresholds for complex AI to enter terminals, making AI capabilities further spread from high-performance PCs, smartphones, and robots to industrial devices, AIoT, wearables, and smart homes.
Investment advice and valuation
Physical AI itself and industrialization related targets: Yushu Technology, Preferred Choice, etc.; data acquisition and sensing hardware related targets: Obi Zhongguang, Stevie, Vail shares, Chipcom, Amperon, Biyi, etc.; End-side AI chip-related targets: Rockchip, Xingchen Technology, Lexin Technology, Hengxuan Technology, etc.; targets related to machine manufacturing: Lixun Precision, Industrial Fulian, etc.
Risk Alerts
The improvement in physical AI model capabilities fell short of expectations; the industrialization of robots and terminal shipments fell short of expectations; the demand for physical AI data collection and training hardware fell short of expectations; the progress of end-side AI penetration and local model deployment fell short of expectations; the extent of end-side hardware configuration upgrades fell short of expectations; rapid changes in technology routes and intensified industrial competition.