Summary of Key Points
On July 28, 2026, at the First Financial Technology Innovation Conference, five leading figures in the field of physical AI discussed topics such as the definition of physical AI, the path for hardware and software collaboration, and commercialization. The consensus was that physical AI is transitioning from a conceptual stage to a critical point of mass production, requiring the coordination of computing power, simulation, communication, hardware, and algorithms. Its essence is to enable physical entities (such as robots and cars) to possess an intelligent closed loop that includes perception, decision-making, and execution, which is a prerequisite for AGI (Artificial General Intelligence). The balance between hardware and software collaboration must be found within the framework of openness and integration. The challenges in commercialization lie in the lack of accessible scenarios and data; the key is to create real-world use cases that form a viable closed loop.
I. What Exactly is Physical AI?
Physical AI is not a new term, but it has suddenly gained attention this year (several companies on the Hong Kong stock market have sought to be labeled as the "first physical AI company"). However, there was previously a lack of clear definitions for it. The guests provided several straightforward explanations:
- The essence is an intelligent closed loop in the physical world: Xu Chunshan from Lingqing Zhiyuan explained that physical AI aims to make tangible objects like robots and cars understand physical principles such as gravity and friction, enabling them to form a complete cycle of "perceiving the environment → making decisions → taking actions → receiving feedback" (for example, a robot must know the weight and smoothness of a cup to pick it up securely).
- A necessary step towards AGI: Nie Kaixuan from Songying Technology argued that current AI systems focus on processing text and images from the internet but do not understand the physical processes (such as car manufacturing or cooking), making it impossible to fully model the real world. Physical AI, once matured, could pave the way for AGI.
- Operating in the physical world like humans: Liu Ke from Momenta used autonomous driving as an example, stating that physical AI means understanding the physical environment and acting safely (for instance, a self-driving car recognizing traffic lights and avoiding pedestrians).
II. The Bottlenecks of Physical AI: Data, Latency, and Computing Power
The guests agreed that physical AI is not just about competing for computing power; there are three major challenges:
- Data scarcity outweighs computing power: Xu Haijiang from Lingjing Zhiyuan pointed out that there is a lack of physical data for AI training. Videos, although useful, lack real-world elements like sound, temperature, and force. Wu Jianjun from Zijinxingyu noted that while ChatGPT has access to vast internet data, physical AI lacks such resources, and solutions to data collection and sharing are needed.
- Latency is critical: A 100-millisecond delay can cause significant errors in physical systems (e.g., a robot might drop a cup). This issue is more severe in real-world applications than in cloud-based AI.
- Efficient use of computing power: Instead of simply increasing computing power, it’s important to allocate it strategically. Wu Jianjun suggested that the robot's control functions should be handled locally (e.g., near 6G base stations) to reduce costs and energy consumption.
III. Hardware and Software Collaboration: Openness or Integration?
NVIDIA represents a approach that combines both hardware and software, but there was debate about whether Chinese companies should follow this path or adopt a more collaborative model:
- Combining without uniformity: Xu Chunshan emphasized the need for stable hardware (e.g., robust robot joints) and flexible software (able to update algorithms continuously). He suggested three approaches: using the system in real scenarios, designing hardware with software requirements in mind, and conducting extensive practical tests to complement each other.
- Openness fosters a thriving ecosystem: Nie Kaixuan cited Android as an example, noting that while Microsoft’s Windows dominated the PC era with a closed system, Android’s open architecture led to its dominance in the mobile market. Songying Technology has adopted an open approach, collaborating with multiple chip manufacturers and releasing its software under an open-source license.
- General openness with scenario-specific integration: Xu Haijiang argued that foundational technologies like chips and operating systems should be open for broader application, but customization is needed for specific use cases (e.g., robots used in cooking).
- A balance between vertical integration and open source: Wu Jianjun suggested that while vertical integration (like SpaceX’s complete rocket production) can be efficient, open source is essential for a healthy ecosystem. In the automotive industry, standard configurations should be established to share data and reduce redundancy.
IV. Commercialization: From Showcased Concepts to Profitability
Although physical AI concepts are exciting, revenue generation has lagged behind. The guests proposed the following steps for commercialization:
- Start with practical use: Xu Chunshan recommended focusing on real-world applications, identifying issues during use (e.g., difficulties in climbing stairs), and then adjusting algorithms and hardware to improve the system.
- Address data scarcity: Nie Kaixuan mentioned that simulations can create multiple scenarios from a single real one to train models, but this requires willing partners (e.g., companies allowing robots to enter warehouses).
- Reduce costs: The key to mass production is lowering costs: minimizing data collection expenses, reducing the need for expensive hardware in end-users’ devices, and customizing hardware for specific scenarios. Momenta’s autonomous driving technology reduced development time from 24 months to 40 days by using a universal model.
Conclusion
Physical AI is not a theoretical concept; it is crucial for transforming AI from online interactions into practical applications. We are at a critical point of mass production, but to overcome the remaining challenges (data scarcity and latency) and achieve true commercial success, we need to find the right balance between hardware and software and ensure that products are effectively used in real-world scenarios. After all, only profitable AI solutions will be successful.