Dongwu Securities: Physical AI opens up space for industrial software and automated revaluation, focusing on the three main lines of autonomous operation of the process industry

The Zhitong Finance App learned that Dongwu Securities released a research report saying that the current main investment line of the world model focuses on companies that have mastered physical world data, mechanism models, and control interfaces. The bank anticipates that the first round of commercialization of physical AI will occur in high-value, semi-structured scenarios such as process industry, engineering simulation, automobiles, warehousing and logistics, and construction. Focus on three main lines: the first is the autonomous operation of the process industry; the second is physical simulation and engineering world models; and the third is the industrial intelligence and digital main line.

The main views of Dongwu Securities are as follows:

World Model is not a new name for Wensheng's video model, but rather an internal simulation system that allows agents to “anticipate consequences before they act”

A complete world model must complete at least three things: forming a representation of the current state of the world from multi-modal observations; predicting the world itself and state changes under the intervention of different actions; and supporting agents to compare different action paths and make decisions. A model that can only generate realistic video but cannot accept motion input and support planning can only be called a generative model with a “world prior experience,” and cannot yet be considered a complete executable world model.

The world model will not converge to a single technical route; the final form will be a hybrid architecture of generative models, latent spatial dynamics, physical simulation, and causal structures

Video generation models solve data amplification and scene coverage problems; JEPA and latent spatial dynamics solve low-cost state prediction and planning problems; CAE, digital twins, and physical engines solve problems of accuracy, safety, and verifiability; and causal models seek to solve problems of generalization and active exploration outside distribution. The bank determined that in the future, the industrial world model will not be developed by “pure neural networks to replace traditional simulation”, but will form an integrated system of “mechanism models providing boundaries, neural network learning residuals, real-time data correction states, and intelligent execution of closed loop control”.

The current world model has passed the technical concept period, but it is still in the early stages of commercialization of vertical scenarios, rather than the maturity of general physical intelligence

Since 2025, V-JEPA 2 has demonstrated zero-sample robot planning in a new environment. Genie 3 can generate interactive environments in real time, and Cosmos 3 has begun to unify visual reasoning, world generation, and motion generation into the same physical AI basic model. However, these developments mainly prove that the model can be “understood, simulated, and controlled in a short range”, and there is still a significant gap from long-term stable operation in complex open environments.

The bank anticipates that the first round of commercialization of physical AI will occur in high-value, semi-structured scenarios such as process industry, engineering simulation, automobiles, warehousing logistics, and construction

These scenarios have characteristics such as clear task boundaries, continuous data flow back, quantifiable error costs, and the ability of existing automated systems to directly handle model output. In 2024, the number of new industrial robots installed worldwide reached 542,000, and the number in service reached 4.664 million. China accounted for 54% of the world's new installations, providing ready-made sensors, controllers, robot bodies and production data bases for physical AI.

Risk warning: Technological progress falls short of expectations. The opening of industrial data fell short of expectations. Commercialization and standardization fell short of expectations. Fluctuations in capital expenditure for industrial customers. Safety incidents and regulatory risks. Risk of mismatch between valuation and performance.