Summary of Key Points
The large-scale model industry is shifting from a frenzied competition focused on the size of parameters to a more pragmatic phase that emphasizes commercial outcomes. On one hand, capital is flocking to leading model companies, with valuations reaching hundreds of billions; on the other hand, there is a growing interest in exploring exit strategies. Sub-sectors such as world models and embodied intelligence are becoming new areas of investment, although bubbles are also forming (with valuations reaching several billion at the angel round). Super AI applications for the consumer market have yet to emerge, and the "cost hole" associated with AI products—where more users often lead to greater losses—makes commercial viability a critical factor. The competitive landscape has completely changed, shifting from a focus on technical parameters to sales and practical implementation. Going global has become a necessity, and the cycle for establishing product barriers has been shortened to just a few weeks.
1. The Decline in Parameter Competition: From "Who Has the Biggest Parameters" to "Who Can Actually Implement"
Previously, large-scale model companies competed based on the size of their parameters. In July, Kimi K3 with 2.8 trillion parameters and Qwen3.8 with 2.4 trillion parameters pushed this competition to new heights. However, at events like WAIC, there are no longer displays showcasing massive parameter counts; instead, people are presenting AI entities that can perform tasks independently, robotic arms that can tighten screws, and solutions that have been mass-produced. This indicates a more practical approach from the industry: customers are no longer interested in showing off technical capabilities but rather in whether the models can solve real-world problems (such as improving factory efficiency or reducing content production costs). Capital is also asking, "How will you make money? When can I get my investment back?"
To illustrate this point with data: The daily token consumption of large-scale models in China has increased by 3,000 times over two and a half years, meaning AI systems now process approximately 450 trillion "characters" per day. This is driven by a positive cycle of growing demand and decreasing costs—cheaper models attract more users, which in turn drives further technological innovation. The focus is no longer on using the models extensively but on using them effectively.
2. Capital's Ambivalence: Both Euphoria and Anxiety
On one hand, leading model companies are reaping benefits from capital flows. For example, Zhipu’s market value is接近 HK$60 billion, and DeepSeek’s valuation is close to HK$50 billion. However, there is also considerable anxiety:
- Valuation Bubbles: Some projects are valued at billions at the angel round, which seems unrealistic, such as small laboratories founded by post-95s or post-00s researchers that have not even released a product yet.
- Fierce Competition in Sub-sectors: Hundreds of companies are entering fields like world models and AI applications, but no clear leaders have emerged.
- Investment in Uncertain Areas: With the basic structure of large-scale models established, capital is shifting to less certain but potentially more rewarding areas, such as world models (which aim to help AI understand physical laws) and embodied intelligence (AI robots).
Why is embodied intelligence so popular? Some calculations suggest that if these technologies mature and 1 billion units are sold at $30,000 each, the market would be worth trillions of dollars—larger than the smartphone and automotive industries combined. However, the current issue is that many companies in this field are still in the demo stage and have not generated substantial revenue. If they fail to make a breakthrough by the end of this year or next year, the market could experience significant adjustments.
3. The Dilemma of AI Applications: Lack of Consumer Super Products and High Costs
Last year, there was hope that AI would create consumer super applications like WeChat or TikTok. However, the current hot trends are productivity tools such as coding assistants and task-executing agents. The reason for this is that the cost structure of AI products differs from that of internet services:
- Internet products generate revenue as more users are attracted (through advertising and subscription fees).
- AI products incur higher costs with each user interaction, as they require processing power (represented by tokens). For example, an AI toy company was relieved when 90% of its users did not use the AI features, as that reduced its expenses. Therefore, investors in AI applications first ask whether the products can be profitable; sectors like education and finance, where customers are willing to pay, are considered safer investments. Novelty-driven applications (like AI chatbots) struggle to succeed despite impressive technology.
4. Changing Competitive Logic: From Technology to Commercialization and Sales
Three major changes have occurred in the AI startup scene this year:
1. Young Teams Focusing More on Technology: Many startups founded by post-00s entrepreneurs focus solely on developing models without understanding how to sell them.
2. Going Global as a Must: One out of every two or three startups plans to expand internationally from the beginning, seeking global markets due to domestic competition.
3. Falling Barriers: Both large companies and model providers are entering the application development space, and products are often replicated within weeks, rendering technical barriers ineffective.
The focus has shifted to who can successfully market their products and generate revenue quickly. Examples include the "AI + BPU" (human-machine collaboration) model, which builds barriers through service offerings, or traditional companies transitioning to AI to leverage their existing customer bases and channels. Most pure AI startups have not yet found a stable business model.
5. The Future Direction: Industry Integration as the Key to Success
The CEO of Guangyuan Capital believes that the industrialization of AI will go through three waves:
1. First, it will transform digital tasks (such as coding and content creation).
2. Then, it will impact scientific research and development.
3. Finally, it will transform physical labor in industries like manufacturing, agriculture, and logistics.
China’s advantage lies not in its model algorithms but in its ability to integrate technology into various sectors. For instance, factories are willing to adopt AI-powered robots, and farmers are open to using AI for farming. However, this requires technical teams with industry expertise and industry partners who are receptive to AI.
In summary, the playing field in the large-scale model industry is narrowing: while parameters were once sufficient for funding, commercial success is now the decisive factor. Only companies that can demonstrate profitability within the current market environment will be the winners of the next round.