Dongwu Securities: Physical AI becomes the direction of greater increment, intelligent driving is the first to get through, focus on

The Zhitong Finance App learned that Dongwu Securities released a research report saying that digital AI and physical AI are not isolated and separated, but are continuously distributed in terms of data form, deployment environment, and operational consequences. The development path of the digital AI base model has gradually become clear, and its value lies in certainty, but it is difficult to achieve a closed loop of modeling the physical world; the bank believes that physical AI will become an AI solution for the physical world, thus becoming a more incremental direction in the current development stage of AI. Intelligent driving is the representative scenario that first runs through a closed loop, and the bank believes it should focus on it.

The main views of Dongwu Securities are as follows:

Digital AI and physical AI are not completely symmetric dichotomous concepts; they are more suitable for understanding along the “degree of physical closure”

Physical AI undertakes the academic tradition of embodied intelligence and was promoted to the mainstream by industry players such as Nvidia from 2024 to 2025. The core pillars include autonomous driving and personable/humanoid robots; digital AI is more of a retroactive synonym formed to distinguish intelligence in information space after the concept of physical AI became popular. The boundary between the two is not isolated, but is continuously distributed in terms of data morphology, deployment environment, and operational consequences.

The most fundamental difference between the two types of AI is reflected in deployment patterns and security constraints

Digital AI mainly targets text, code, knowledge, and software status. Most general services are highly dependent on cloud networking and reoperation; physical AI targets vehicles, robots, and real space, cannot assume that the network is connected at any time, and must operate under conditions of end-side autonomy, safety concerns, and real-time feedback. High-risk digital scenarios also require strict verification, but the failure of physical AI is more likely to directly cause irreversible physical consequences.

Judging from the history of technology, digital AI is a path where general methods, data, and computing power continue to replace manual design. Currently, the focus is shifting from expanding parameters during training to expanding computational power during inference; physical AI first completes industrial iteration in autonomous driving and later in robots, gradually moving from high-precision maps, lidars, and modular systems to end-to-end, VLA, and world models. The world model has become a common concern for both types of routes, but the difference is whether the potential space needs to carry pixel reconstruction, that is, whether the world model should focus on generation, understanding, planning, and service physical closed-loop control.

In terms of the pace of commercialization, digital AI already has a clearer closed loop between charges and products in scenarios such as AIcoding, enterprise-level agents, dialogue search, and content generation; in physical AI, RoboTaxi regionalized operations and charges are gradually progressing, and humanoid robots are still in the verification period for delivery, cost, reliability, and unit economy. The pattern between China and the US also shows a division of labor: the US is leading in cutting-edge models, capital, and unmanned mileage, and China has advantages in large-scale operations, supply chains, hardware costs, and implementation of some applications.

From a perspective leading to AGI, robots are becoming the intersection of digital AI and physical AI

More and more institutions believe that pure digital models are difficult to form robust intelligence on their own, and that they need to use physical grounding, world models, and real interactive data to improve generalization capabilities. In the future, a joint architecture where “multi-modal big models are responsible for semantics and planning, the world model is responsible for physical prediction, and the control system is responsible for actual execution” may be formed, but whether the big model comes first or the world model/control priority has not settled.

Risk warning: the risk that technology iteration falls short of expectations; the risk that the ARR growth rate of large model manufacturers will slow down; the risk that the capital expenditure of cloud service providers falls short of expectations; the risk that the commercialization of intelligent driving and robots falls short of expectations.