Summary of Key Points
China's AI industry is transitioning from a phase focused on “model competitions” to one that emphasizes the implementation of “token economies” and “embodied intelligence,” facing challenges related to valuation bubbles and difficulties in commercialization. As a foundational general technology, AI needs to go through three transformations to realize its economic value: from computing resources to tokens, to corporate capabilities, and finally to revenue and productivity. The competition in intelligent computing has shifted from simply accumulating GPUs to focusing on system efficiency (i.e., the cost per token generated). Embodied intelligence must break through the limitations of the physical world through continuous iteration, and investments should shift from chasing technological trends to assessing commercial viability. Whether AI can become a new driver of growth ultimately depends on its ability to genuinely enhance productivity and generate sustainable returns.
Detailed Explanation
1. Why AI Can Replace Traditional Drivers of Growth? Because It’s a “Foundational General Technology”
In the past, China's economy relied on real estate and globalization, but these drivers are now undergoing changes: the population is declining, capital yields are decreasing, and the foreign trade environment has become more complex. AI holds great promise not because it represents a new industry (such as automotive or internet), but because it acts like electricity—它能 permeate all sectors, including research and development, manufacturing, healthcare, and education, transforming how information is processed and produced.
For example, traditional production relied on humans, machines, and money; in the AI era, “intelligent computing” becomes an additional factor. Factories use AI to optimize processes, and hospitals use it for diagnostic assistance. These are not standalone AI products but rather integrations that improve overall industry efficiency.
2. The Competition in Intelligent Computing Has Changed: From Accumulating GPUs to Optimizing Token Costs
Initially, the focus was on having larger model parameters and more GPUs; now, we are in the “inference era,” where companies buy the ability to perform tasks efficiently. Tokens serve as the unit of measurement for this capability—each AI response or piece of generated code consumes tokens. The competition is no longer about who has the most GPUs but about how much value can be created from each unit of computing power. Although China lacks advanced chips, it has large-scale electricity infrastructure, a complete manufacturing ecosystem, and a wealth of engineers. By combining domestically produced chips with optimized software and efficient energy management, China can reduce token costs.
However, building intelligent computing centers does not equate to developing AI; many centers end up idle without customers. What matters is the actual usage rate and whether customers are willing to pay for the services provided.
3. Embodied Intelligence Is More Than Just “Robots + Models”; It Requires Closed-Loop Iteration
Embodied intelligence means bringing AI into the physical world (e.g., through robots). Simply giving a large model a physical form is not enough; it must handle uncertainties in the real world (slippery surfaces, obstacles, unexpected human interactions, etc.). China’s advantages include a complete supply chain and numerous manufacturing applications for robots. The key is to create a closed loop: sell robots, collect data on-site, use that data to train models, improve the models, and generate more orders. This approach works well in B2B scenarios (factories, warehouses) where the environment is controllable and customers are willing to pay. Home applications (B2C) are still in their early stages due to complex environments and high safety requirements.
4. Investing in AI: Focus on Commercial Viability, Not Technological Trends
The investment approaches in AI differ between China and the United States. The U.S. has more capital and a well-established ecosystem for basic models and cloud platforms, but it faces challenges in turning investments into sustainable revenue. China, with its rapid engineering progress and diverse use cases, still needs to develop customer demand through competitive pricing strategies.
Investments should target three types of assets:
- Efficiency-enhancing assets: Those that reduce token costs (e.g., domestically produced GPUs, liquid cooling technology, computing power management software).
- Closed-loop assets: Systems that can create data feedback loops (e.g., robots with mass production capabilities and continuous data collection).
- Use-case-focused assets: Applications that understand customer processes (e.g., healthcare AI that understands hospital operations and has necessary licenses).
Do not be misled by factors like large user bases or high token consumption; only scale that is combined with repeat sales and profit margins counts as genuine demand.
5. The Future of AI: Not a Larger Tech Industry, but a New Economic Infrastructure
AI’s ultimate goal is to become a foundational infrastructure, just like electricity or cloud computing. When intelligent technologies are readily available and customizable on demand, economic growth will truly incorporate the power of AI. To achieve this, we need to see three outcomes: whether corporate productivity has increased, whether workers have acquired new skills, and whether capital generates sustained returns. Policies should also shift from supporting equipment production to fostering real demand (e.g., encouraging factories to use AI robots and establishing data governance frameworks for collaborative work with workers).
In summary, the success of AI lies in its ability to transform “cheaper intelligence” into higher efficiency and revenue, not in the number of models released or centers built.