A recent research report from Soochow Securities Company Limited highlights that the current investment focus on world models is centered on enterprises that already possess physical world data, mechanistic models, and control interfaces. The firm anticipates that the first wave of commercialization for physical AI will emerge in high-value, semi-structured sectors such as process industries, engineering simulation, automotive, warehousing and logistics, and construction. Three primary investment themes are outlined: autonomous operation in process industries, physical simulation with engineering world models, and industrial intelligent agents with digital threads.

The core perspective of Soochow Securities is that a world model is not merely a new label for text-to-video generative models, but rather an internal simulation system that enables an intelligent agent to "rehearse consequences before taking action." A complete world model must accomplish three essential tasks: form a representation of the current world state from multimodal observations, predict the evolution of the world itself and its state changes under various action interventions, and support the agent in comparing different action paths to make decisions. Models that can only generate realistic videos but cannot accept action inputs or support planning should only be classified as generative models with "world priors," and they cannot yet be considered fully executable world models.

World models are unlikely to converge on a single technological path; the final form will be a hybrid architecture combining generative models, latent space dynamics, physical simulation, and causal structures. Video generation models address the challenges of data augmentation and scenario coverage, while JEPA and latent space dynamics provide low-cost state prediction and planning. CAE, digital twins, and physics engines solve issues of accuracy, safety, and verifiability, and causal models aim to tackle out-of-distribution generalization and active exploration. The firm concludes that future industrial world models will not evolve by simply replacing traditional simulation with pure neural networks. Instead, they will form an integrated system where mechanistic models define boundaries, neural networks learn residuals, real-time data corrects states, and intelligent agents execute closed-loop control.

Currently, world models have moved past the conceptual technology stage, but they remain in the early phase of vertical-scenario commercialization rather than the mature stage of general physical intelligence. Since 2025, V-JEPA 2 has demonstrated zero-shot robot planning in new environments, Genie 3 can generate interactive environments in real time, and Cosmos 3 has begun unifying visual reasoning, world generation, and action generation into a single physical AI foundation model. However, these advancements primarily prove that models can "understand, simulate, and control over short ranges," and there remains a significant gap when it comes to long-term stable operation in complex open environments.

Soochow Securities predicts that the first round of physical AI commercialization will occur in high-value, semi-structured scenarios such as process industries, engineering simulation, automotive, warehousing and logistics, and construction. These scenarios are characterized by clearly defined task boundaries, continuous data feedback loops, quantifiable error costs, and existing automation systems that can directly adopt model outputs. In 2024, global new installations of industrial robots reached 542,000 units, with the total in-operation fleet reaching 4.664 million units. China accounted for 54% of global new installations, providing a ready-made foundation of sensors, controllers, robot bodies, and production data for physical AI.

Risk warnings include technological progress falling short of expectations, slower-than-anticipated opening of industrial data, underperformance in commercialization and standardization, fluctuations in capital expenditure by industrial clients, safety incidents and regulatory risks, and potential mismatches between valuation and performance.