Image| Movie Parallel Universe
In March 2025, Nvidia CEO Huang Renxun's speech at GTC caused a sensation in the tech industry. In a speech lasting over two hours, he spent 40 minutes presenting a disruptive viewpoint: Physical AI will open up trillion dollar new markets, and the next wave of artificial intelligence will transition from "understanding information" to "manipulating entities".
In just one month, there have been numerous confirmation signals from the industry side: Tesla's humanoid robot Optimus has independently completed complex maintenance tasks in real factories, and Alibaba Cloud immediately announced a deep binding with Nvidia to jointly promote the industrial landing of physical AI in China.
Behind the capital's enthusiasm for "humanoid robots" and "world models", a core proposition has emerged: how to make AI safely and reliably interact with the physical world?
From Bit to Atom: Three Layer Transition Logic of Physical AI
The evolutionary framework of "perceptual intelligence generative intelligence agent intelligence physical intelligence" proposed by NVIDIA accurately outlines the essential transition of AI application scenarios. Unlike the big language models that deal with the information world, the core of physical AI is to solve the optimization problems of the physical world, and its implementation requires a systematic integration of three layers of technology.
The first layer is environment reconstruction: By using digital twin technology, the physical environment is transformed into a structured digital model. Not only does it require three-dimensional geometric reconstruction, but it also requires the digital expression of spatial relationships, object properties, and physical laws. According to data from the China Academy of Information and Communications Technology, as of 2024, more than half of the prefecture level administrative regions in China have carried out the construction of digital twin cities. According to data from some research institutions, by 2025, over 500 cities worldwide will deploy digital twin platforms, laying the foundation for the data environment of physical AI.
The second layer is simulation training: In a high fidelity digital environment, the machine learns physical laws, relying on a physics simulation engine to generate diverse testing and training data based on high confidence sensors and dynamic simulations. Industry data confirms its value: high-quality simulation platforms can reduce the testing cost of autonomous driving algorithms by over 90% and increase training efficiency by dozens of times.
The third layer is entity manipulation: Deploy mature AI models trained through simulation to physical devices to complete the "decision execution" loop. The key to this step is to crack the "simulation reality gap" and ensure the reliable application of virtual skills in real scenarios, whose success rate directly determines the economic feasibility of technology implementation.
However, the industry is facing a clear 'capacity gap'. According to IDC's 2024 survey, less than 15% of technology companies engaged in AI business have complete capabilities in digital modeling, simulation training, and physical deployment. Most companies can only focus on a single link, making it difficult to form an effective closed loop in the technology chain.
Chinese sample: 51WORLD's closed-loop breakthrough path
Among Chinese technology companies, the ten-year development trajectory of 51WORLD has unexpectedly become a typical example of the evolution of physical AI technology - gradually moving from digital twins and simulation to physical intelligence infrastructure, ultimately reaching the "endgame" of Physical AI.
From 2015 to 2018, the company focused on urban level digital twin platforms, with the core breakthrough being "data structuring": it is not simply three-dimensional visualization, but by integrating multiple sources of data, endowing digital objects with semantic information and association rules.
From 2019 to 2021, the launch of simulation platforms marked a technological upgrade: the introduction of physical law simulation in digital environments, generating high-quality synthetic data with a reality rate of over 90% on key perception tasks (data disclosed in its prospectus).
Since 2022, through robot platforms and industry solutions, the company has completed a closed-loop exploration of "simulation entity": in scenarios such as autonomous driving, smart energy, and smart factories, intelligent agents based on simulation training can operate real devices to complete specific tasks.
This gradual path of "reconstruction simulation execution" is essentially driven by the mutual interaction between market demand and technological capabilities: as customers expand from visualization needs to simulation analysis and intelligent decision-making, the enterprise technology stack continues to iterate, ultimately building a full chain physical AI closed-loop ecosystem of "synthetic data spatial intelligent model simulation training platform".
Triple Barrier: Composite Game of Data, Knowledge, and Engineering
The technological path of physical AI may seem clear, but in practice it faces triple composite barriers. Its core is not algorithm breakthroughs, but deep collaboration between data standards, domain knowledge, and engineering implementation.
The data standard system is the primary threshold: Building a digital world requires a unified spatial coordinate system, object classification system, and behavioral description standards, and the establishment of a mature standard usually requires 3-5 years of project precipitation and iteration. The success of Nvidia Omniverse stems from years of cultivation and the establishment of factual standards for digital content creation.
Cross disciplinary knowledge digitization is the second obstacle:There are significant differences in the physical laws of different industries.. Taking smart water conservancy as an example, accurate water situation prediction requires the integration of multidisciplinary knowledge such as hydrology, meteorology, geology, etc., and the transformation of these professional knowledge into computable digital models. It requires deep binding between domain experts and technical teams, and this ability is difficult to replicate in the short term.
The ability to implement system engineering is a key support: the deployment from simulation environment to real system involves complex engineering integration. In industrial scenarios, AI algorithms need to interface with multiple subsystems such as PLC control systems, sensor networks, and actuators. This ability can only be repeatedly polished in real projects and cannot be quickly obtained through laboratory research and development.
More importantly, the triple barriers have a mutually reinforcing effect: the improvement of data standards facilitates the digitization of domain knowledge, while engineering implementation experience can feed back into standard optimization. This composite barrier makes the competitive landscape of physical AI relatively stable, making it difficult for latecomers to catch up through a single technological breakthrough.
Ecological pattern: infrastructure hegemony competition at the platform level
In the industrial ecosystem of physical AI, enterprises form differentiated positioning based on their own capabilities, and the role of "enablers" at the platform and tool layers is becoming increasingly crucial.
Hardware and chip layer: Represented by NVIDIA, it provides computing power support and core processors, and builds a moat through GPUs, dedicated AI chips, and software ecosystems.
Platform and tool layer:Focus on digital environment construction and simulation training, and provide digital twin platforms, simulation engines, and synthetic data generation tools. The core value is to reduce the environmental cost of AI training and industry access threshold. Similar to the cloud computing platform in the mobile Internet era, its value will grow exponentially with the popularity of physical AI.
Model and Algorithm Layer:Develop spatial intelligence models and world models, build a universal "brain" for physical AI, empower hardware such as robots with the ability to understand three-dimensional space, follow physical laws, and complete a "perception decision action" loop, and solve their generalization problems in unstructured environments..
Application and solution layer:Provide landing services for specific industries, combining physical AI technology with scenarios such as autonomous driving and industrial robots to form commercially viable products..
Terminal and device layer:Production integration of physical hardware such as robots and intelligent vehicles is the carrier of interaction between AI and the physical world..
From a business model perspective, platform level enterprises typically adopt a multi-level charging strategy: basic platforms are charged based on authorization, cloud services are charged based on resource usage, and customized solutions are charged based on projects. This combination model ensures both income stability and the ability to share industry growth dividends.
Industry trajectory: gradual penetration and value reconstruction
The industrialization process of physical AI presents a distinct progressive feature, and its development trajectory can be analyzed from three dimensions.
Time dimension:Recently, we have focused on technology validation and benchmark projects, with the core competition being simulation accuracy and engineering reliability;; In the medium term (3-5 years), industry standards will gradually be established, and platform based enterprises are expected to dominate the ecosystem; In the long run, physical AI will become the digital infrastructure for various industries.
Spatial dimension:Penetrating from high-value scenarios to inclusive scenarios.. Currently focused on areas such as autonomous driving and high-end manufacturing, it will expand to a wide range of scenarios in the future, including logistics warehousing, agricultural operations, and urban services. McKinsey predicts that by 2030, approximately 30% of physical work activities worldwide can be automated.
Ecological dimension:Promote the value reconstruction of the industrial chain.. Hardware suppliers need to develop more specialized computing chips, software platforms need to improve their toolchains, application developers need to deepen their domain knowledge, and new business models and forms of cooperation will continue to emerge.
It is worth noting that the development of physical AI needs to balance three major relationships: technical progressiveness and engineering feasibility, general platform and industry customization, innovative development and safety and reliability. The grasp of these balance points will directly affect the speed and effect of technology landing.
Conclusion: The mission of connecting two worlds in this era
Big models change the information world, and Physical AI will reshape the physical world. Physical AI is the 'super race track' for the next generation of AI. This technological revolution is not only a deep integration of bits and atoms, but also a fundamental change in the paradigm of productivity.
For China's technology industry, physical AI is both an opportunity and a challenge: we have the scale advantage of digital new infrastructure and rich industrial scenarios, but we still need to accelerate catching up in areas such as basic research, standard setting, and ecological construction.
The core of this competition is the ability to build a 'world training ground'. Enterprises that can provide high-quality digital environments, simulation tools, and training platforms will become a key force in feeding physical AI and accelerating its secure evolution. Their value lies not only in the technology products themselves, but also in building a "connector" between the digital and physical worlds for the intelligent era.
The global competition for physical AI has been fully launched, and the results of this competition will determine the direction of reshaping the global industrial landscape in the next decade.
