Physical AI represents a groundbreaking advancement in the artificial intelligence (AI) sector. Unlike traditional digital AI, which primarily focuses on text and image processing, physical AI integrates real-world physical principles such as mechanics, optics, fluid dynamics, and kinetics into its large-scale model algorithms.

It establishes a comprehensive closed-loop system encompassing "environmental perception—physical reasoning—action execution—real-world feedback iteration," enabling AI to perceive, understand, and interact with the physical world.

As the foundational technology for areas such as embodied intelligence, humanoid robots, intelligent driving, and flexible industrial manufacturing, physical AI bridges the gap between the digital and physical worlds. This marks a pivotal shift for AI from "intelligent computing power output" to "physical implementation and application."

Currently, the global physical AI industry ecosystem has evolved into a sophisticated three-tiered structure, comprising three main segments: upstream computing power and simulation infrastructure, midstream sensing hardware and algorithmic decision-making, and downstream execution hardware and scenario implementation.

Upstream Segment

Computing Power and Simulation Infrastructure: The Technological Backbone

As the training groundwork for physical AI, the upstream segment accounts for roughly 35% of the total industry value and represents the highest technological barrier and greatest computing power demand across the entire ecosystem.

Compared to general-purpose large models, physical AI's high-precision simulation training requires 5-10 times more computing power. This segment primarily covers three subfields: computing hardware, physical simulation industrial software, and physical datasets. These components provide essential functions such as virtual training, physical rule modeling, and data support for intelligent physical devices.

Computing Hardware: The Powerhouse of Physical AI

Computing hardware, categorized into cloud-based training computing and edge inference computing, serves as the core platform for physical model simulation training and real-time physical reasoning.

Foxconn Industrial Internet stands as the world's largest contract manufacturer for NVIDIA's GB200/H200 servers, deeply integrated with the Cosmos ecosystem, while also deploying edge computing systems for robots. It commands the top global market share in AI server contract manufacturing.

Inspur Information leads the domestic AI server market, offering specialized edge computing systems for embodied intelligence tailored to the real-time physical computing needs of robots.

In the computing support sector, ZTE's 800G/1.6T high-speed optical modules are vital components for NVIDIA's computing cluster interconnection.

Cambricon, Hygon Information, and Jingjia Micro continue to make breakthroughs in domestic AI acceleration chips and GPU technologies, catering to cloud, edge, and military simulation computing scenarios.

Physical Simulation CAE Software: The Foundation for Virtual Training

Physical simulation industrial software represents the technological pinnacle for physical AI, leveraging differentiable physics engines, multi-field coupling solvers, and digital twin technologies to enable millions of trial-and-error training sessions for intelligent devices in virtual environments. This significantly reduces R&D losses and testing costs for physical prototypes.

So Chen Technology leads in physical AI simulation, with its self-developed 'Tiangong·Kaiwu' differentiable full-physics-field engine successfully applied in scenarios such as humanoid robots and virtual vehicle training.

ZWCAD Software capitalizes on its strengths in domestic CAD technology to develop industrial digital twins and robot motion simulation, empowering physical modeling training for industrial robots.

Nengke Technology constructs industrial digital twin platforms, enabling seamless integration of physical data from factory production lines and the implementation of intelligent manufacturing simulations.

Huaru Technology and Fantuo Digital focus on military virtual simulations and digital twins for energy infrastructure, respectively, enhancing physical simulation layouts in niche sectors.

Physical Datasets: Scenario-Based Data Support

Physical datasets provide real-world scenario-based data support for model training, encompassing primarily 3D spatial data, force perception data, and motion trajectory data.

Companies such as Hisound and Tuoersi specialize in the collection and annotation of physical data in industrial and robotic fields, addressing gaps in domestic physical AI datasets.

Midstream Segment

Sensing Hardware and Algorithmic Decision-Making: The Senses and Brain of Physical AI

The midstream segment accounts for approximately 30% of the total industry value and serves as the core hub connecting the physical environment with intelligent decision-making. It is divided into three major sectors: sensing hardware, physical AI large models, and motion control systems.

It performs essential functions such as environmental information collection, physical rule reasoning, and motion command output, determining the environmental adaptability and operational precision of physical AI devices.

Sensing Hardware: The Environmental Perception Terminals

Sensing hardware acts as the "eyes and skin" for physical AI to acquire real-world data, with key subfields including 3D vision, LiDAR, force sensors, and MEMS microsensors, enabling precise collection of environmental information such as 3D space, force, and optics.

In the 3D vision sector, Orbbec, a domestic leader, is widely used in robotic 3D modeling, intelligent grasping, and automotive perception scenarios.

In optical motion capture and industrial vision, Luster LightTech achieves robotic hand-eye coordination perception and industrial 3D inspection through professional optical motion capture equipment.

In the LiDAR sector, Hesai Technology and RoboSense's dToF products meet the obstacle avoidance and ranging needs of autonomous driving and humanoid robots.

In force perception and sensing, Keli Sensing and Memsic achieve precise collection of robotic joint force feedback, while Will Semiconductor and Ofilm provide core CMOS image chip support for visual perception.

Physical AI Large Models and Control Algorithms: The Decision-Making Hub

Physical AI large models and algorithms represent the industrial core, differing from general-purpose large models by focusing on environmental understanding, behavior prediction, and adaptive decision-making based on physical laws.

SenseTime and CloudWalk focus on world models and physical reasoning algorithms, delving into 3D environmental understanding and intelligent robotic decision-making.

iFlytek and 360 Security embed physical rules into general-purpose large models, adapting them for service robots and industrial automation scenarios.

Thundersoft develops edge-side physical AI operating systems, providing real-time decision-making and integrated hardware-software control solutions for robots and automotive terminals.

Motion Controllers: The Action Execution Hub

Motion controllers are responsible for translating AI decision-making commands into physical actions, serving as the core component for motion coordination in intelligent devices.

Inovance Technology, Leadshine, and Kinco achieve technological breakthroughs, developing servo controllers and multi-joint motion control systems suitable for multi-degree-of-freedom coordinated control in humanoid robots.

Downstream Segment

Execution Hardware and Scenario Implementation: The Industrialization Carriers

The downstream segment accounts for approximately 35% of the total industry value and represents the final carrier for the commercialization of physical AI technology. It covers three major sectors: precision execution components, intelligent equipment systems, and industry-specific application scenarios, determining the industrial scalability and commercial value realization efficiency.

Precision Execution Components: The Bones and Muscles of Intelligent Devices

Precision components are the core hardware enabling precise actions in physical AI devices, with key categories including reducers, servo motors, precision lead screws, dexterous hands, and joint modules.

In the reducer sector, Leaderdrive and Shuanghuan Transmission are domestic leaders in harmonic reducers, suitable for humanoid robot joints.

In the servo system sector, Inovance Technology and Leadshine develop frameless torque motors and servo systems, providing core support for domestic robots.

In precision transmission, Jiangsu Leili and Zhaowei Mechanical & Electrical deploy micro-transmission components suitable for precision operations in dexterous hands.

Tuopu Group focuses on robotic joint assemblies, providing deep support to domestic and international humanoid robot manufacturers.

Intelligent Equipment Systems: The Technology Carriers

Intelligent equipment systems are the implementation terminals for physical AI, primarily including humanoid robots, industrial collaborative robots, autonomous vehicles, and special-purpose intelligent equipment.

In the humanoid robot sector, companies such as Estun, Topstar, SIASUN, and Ubtech deploy domestic physical AI algorithms for R&D and mass production of embodied intelligent systems.

In the service robot sector, Ecovacs and Roborock achieve early commercialization, leveraging physical AI technology for functions such as autonomous obstacle avoidance and adaptive environmental cleaning.

In the intelligent automotive sector, automakers build vehicle-level physical simulation systems and collaborate with third-party companies to implement onboard physical AI technology, empowering iterative upgrades for autonomous driving.

Core Application Scenarios: A Clear Industrialization Hierarchy

The current pace of physical AI commercialization is stratified, presenting a pattern of "smart manufacturing leading, warehousing and logistics following, autonomous driving iterating, and military/special-purpose breakthroughs."

In smart manufacturing, Nengke Technology and SI Consulting leverage digital twins and physical AI technology to create flexible production and intelligent sorting lines, enabling adaptive industrial production upgrades.

In smart warehousing, Hikrobot and Geek+ AMR robots utilize physical AI path planning and irregular cargo grasping capabilities to drive intelligent iterations in warehousing and logistics.

In autonomous driving, IAT builds vehicle-level physical simulation platforms to empower automakers with dynamic simulation and real-vehicle control optimization.

In military/special-purpose sectors, Huaru Technology and So Chen Technology focus on equipment dynamics simulation, driving intelligent upgrades in military applications.

In the medical field, MicroPort and Tinavi leverage force-controlled physical AI technology to achieve refinement operations in surgical robots.

Currently, China has achieved technological maturity and scaling implementation in four core sectors: computing contract manufacturing, domestic physical simulation software, 3D vision perception, and robotic precision reducers. The industry is now upgrading from individual components to full-stack solutions encompassing upstream core industrial software and physical large models.

China's physical AI industrial ecosystem continues to improve, with accelerating technological iteration and implementation speeds. A turning point for large-scale industrialization is expected by 2026, positioning it as a new growth pole in the artificial intelligence industry.

Join the [Enterprise Value] community to access more rankings on enterprise growth capabilities and profitability!

*All information in this article is sourced from publicly available data and is intended for industry discussion purposes only. It does not constitute any investment advice, and investment decisions should be based on independent thinking.