Summary of Key Points from the Roundtable Discussion
This roundtable discussion focused on six critical issues in the AI industry: the evolution of AI technology from a tool to an end-to-end solution; the reconfiguration of valuation logic due to the integration of domestic models and hardware; the need for large-scale model companies to balance technical parameters with commercialization capabilities; the global opportunities for Chinese open-source models under export controls; the differentiation in the training and inference segments of the hardware industry; and the prospects for returning on the billions of dollars invested in AI. The discussion provided investors and industry participants with a multi-dimensional framework for making judgments.
1. AI Technology: From “Writing Code” to “Solving Problems End-to-End”
The biggest change in AI over the past year has been its transformation from a single productivity tool to a solution that can complete entire tasks.
- Breakthroughs in Code Generation: Anthropic’s large models can not only help programmers write code but have also entered a phase of self-iteration, improving their own capabilities; Claude Opus 4.6 enables non-experts to develop small tools, akin to the widespread use of “reinforced concrete” in the digital world.
- The Revolution Brought by Agents: Previously, AI could only chat and solve math problems, but now agents can act as intelligent assistants, handling entire processes—from requirement analysis to code writing and testing. This has shifted AI companies’ business models from selling computing power/usage to selling solutions, and valuation logic has shifted from treating AI as a consumable to viewing it as a software platform (similar to Microsoft or Salesforce).
2. Domestic AI: The Integration of Models and Hardware, Finally Providing a Solid Foundation for Valuation
Previously, domestic hardware manufacturers would claim to be on par with NVIDIA, but investors lacked confidence due to the lack of actual business support. This issue has now been resolved:
- The Significance of DeepSeek V4: It has connected domestic large models with domestic hardware, demonstrating that domestic hardware is not just an empty framework but has real demand behind it. This has made hardware manufacturers’ valuations more grounded and directly linked to the domestic AI computing power market.
- Recognition by Capital Markets: The breakthrough in Cambricon’s market value, the frequent listings of domestic GPU companies, and the challenge posed by Zhipu to become the “number one stock in the large-model sector” indicate that the AI industry has evolved from a concept to a long-term growth track. Similar developments overseas, such as SpaceX’s plans for listing (combining computing power with model platforms) and OpenAI’s IPO, further confirm this trend.
3. How to Choose Large-Scale Model Companies? Both Technology and Profitability Are Important
Evaluating large-scale model companies should not be based solely on the size of their models or the amount of money they have earned; these two aspects are interrelated:
- Technology is Fundamental: The “intelligence” of models (such as self-iteration capabilities and ability to handle long tasks) is the basis for monetization. However, it’s important to consider a range of factors, including talent, data, and continuous iteration capabilities (e.g., DeepSeek’s post-training reinforcement learning and Tongyi Qianwen’s handling of long texts).
- Commercialization is Crucial: Annual recurring revenue (ARR) reflects the model’s monetization potential; for example, OpenAI’s high ARR indicates that customers are willing to pay for its services, which in turn supports further technological development.
- Cost Competition Is Also Important: Not all scenarios require the “strongest models”. For instance, when searching for cheap flights, users might prefer a less expensive AI solution that takes 10 seconds compared to a more expensive one that takes 0.1 seconds.
4. Global Opportunities for Chinese Open-Source Models Under Export Controls
US export controls on large-scale models have led non-US countries to place greater emphasis on “sovereign AI” (models that they can control):
- Trust Crisis and Alternative Needs: Non-US countries are concerned about data breaches when using US models, so they prefer models with weaker capabilities but greater autonomy. Research shows that many non-US countries use Chinese open-source models for training their own large-scale models.
- Temporary Advantage, but Long-Term Success Depends on Competence: This provides Chinese open-source manufacturers with global market opportunities, but it’s important to note that this is a “window of opportunity” rather than a permanent advantage. Long-term competitiveness will depend on capital investment, computing power accumulation, and technological advancement (e.g., the ability to catch up with leading North American models).
5. Hardware Landscape: NVIDIA’s Dominance in Training, but Opportunities for Domestic Manufacturers in Inference
The AI hardware market is divided into training and inference segments, each with a different landscape:
- Training Segment: NVIDIA’s dominance is unlikely to be challenged in the next three years; its CUDA ecosystem (software, developers, toolchain) provides a strong competitive advantage. It will be difficult for other manufacturers to catch up.
- Inference Segment: Domestic manufacturers have opportunities to innovate by “decoupling” GPU functions (e.g., separating preprocessing from inference). This opens up potential for domestic GPU manufacturers to gain a foothold in this market.
- Can the Billions of Dollars Invested Be Recouped?: The answer is yes. Anthropic’s high gross margin and its willingness to share profits with cloud providers (previously 70/30) indicate strong demand for computing power, motivating cloud companies to continue investing. Moreover, AI will expand into sectors like healthcare (e.g., partnerships with Novartis), creating value far exceeding the initial investment—similar to how the internet bubble led to the emergence of large ecosystems such as e-commerce and advertising.
6. There’s an AI Bubble, but Not a Complete One
While some AI assets are overvalued (e.g., some tech stocks with PE multiples of 36), the entire AI industry is not in a bubble:
- Real Demand Drives Growth: Stronger models require more computing power, which in turn drives demand for chips and data centers, creating a self-sustaining cycle.
- Where Are the Opportunities?: Infrastructure (computing power chips, data centers, electricity) and applications (robots, autonomous driving, healthcare) represent long-term growth areas.
- Risks to Be Aware Of: Information asymmetry makes it harder to profit; popular sectors are prone to crowded trading; and commercialization challenges (whether technology can be successfully monetized).
The key is to identify companies that can truly create value—those that can transform technology into profits and possess unique data or ecosystems.