A recent research report from a Chinese securities firm indicates that digital AI and physical AI are not strictly separate categories but exist on a continuum based on data forms, deployment environments, and action consequences. The development path for digital AI base models is becoming clearer, with its value rooted in certainty, but it struggles to achieve a closed-loop modeling of the physical world.
The report suggests that physical AI will become the go-to solution for the physical world, making it a larger incremental growth area in the current phase of AI development. Autonomous driving, as the first representative scenario to achieve a closed-loop process, warrants significant attention.
Key points from the report include:
Digital and physical AI are not completely symmetrical binary concepts. It is more appropriate to understand physical AI through the lens of "physical closure," which inherits the academic tradition of embodied intelligence. From 2024 to 2025, this concept has been pushed into the mainstream by industry leaders like Nvidia, with core pillars including autonomous driving and embodied/humanoid robots. In contrast, digital AI is a retrospective term, coined after the rise of physical AI, to distinguish intelligence focused on information space.
The boundaries between the two are not absolute but are distributed across a continuous spectrum of data types, deployment environments, and action consequences.
The most fundamental differences lie in deployment forms and safety constraints. Digital AI is primarily oriented towards text, code, knowledge, and software states, with many general services relying heavily on cloud connectivity and heavy operations. Physical AI is oriented towards vehicles, robots, and real-world spaces. It cannot assume constant internet connectivity and must operate autonomously on the edge, with safety-critical and real-time feedback requirements. While high-risk digital scenarios also require strict validation, failures in physical AI are more likely to result in direct, irreversible physical consequences.
From a historical technology perspective, digital AI represents a path of continuous substitution of manual design with general methods, data, and computing power. The current focus is shifting from scaling parameters during training to scaling computing power during inference. Physical AI, on the other hand, has undergone industrial iteration first in autonomous driving and then in robotics, moving from high-definition maps, LiDAR, and modular systems towards end-to-end, VLA (Vision-Language-Action), and world models.
World models have become a common concern for both paths, but the key divergence lies in whether the latent space needs to bear the burden of pixel reconstruction. The question is whether world models should focus on generation, or on understanding and planning to serve the physical closed-loop control.
In terms of commercialization pace, digital AI has clearer revenue and product closed-loops in scenarios like AI coding, enterprise-level agents, conversational search, and content generation. In physical AI, the regionalized operation and charging for robotaxis are gradually advancing, while humanoid/embodied robots are still in a period of verification for shipment, cost, reliability, and unit economics.
A division of labor between the U.S. and China is also emerging. The U.S. leads in frontier models, capital, and autonomous vehicle mileage, while China holds advantages in large-scale operations, supply chains, hardware costs, and some application deployments.
From the perspective of achieving Artificial General Intelligence (AGI), robots are becoming the convergence point of digital and physical AI. A growing number of institutions believe that pure digital models alone cannot easily form robust intelligence and require physical grounding, world models, and real-world interaction data to improve generalization capabilities. A future joint architecture may emerge, where a multimodal large model handles semantics and planning, a world model handles physical prediction, and a control system handles real-world execution. However, it remains unclear whether the large model or the world model/control system will take priority.
Risk factors highlighted in the report include:
The risk of technology iteration falling short of expectations. The risk of slowing ARR (Annual Recurring Revenue) growth for large model companies. The risk of cloud service providers' capital expenditure falling short of expectations. The risk of commercial deployment of autonomous driving and robotics failing to meet expectations.
Comments