Summary of Key Points
This news reveals the key to China's AI industry's rapid progress in catching up with international giants like OpenAI: the innovative clusters formed by laboratories affiliated with Tsinghua University and other prestigious institutions. These clusters act as a “high-trust network,” virtually eliminating the costs associated with finding, identifying, and trusting individuals. The companies within these clusters possess an academic foundation and do not follow the American model of simply selling AI models. Their competition focuses not on which model is the most intelligent, but on which one can become the default choice for developers and the industry. The high level of trust within these clusters is the result of forty years of accumulation and cannot be replicated through short-term subsidies; it represents the true moat protecting China's AI development.
1. The Tsinghua University Network: Faster than Online Shopping
If you were to start an AI company and look for a CTO, you might spend three months on headhunting and six months on a trial period, with little actual coding done in the first year and a half. However, in 2023, Yang Zhilin (the founder of Yuezhiànmiàn) was able to find a co-founder in just a few days—a fellow Tsinghua University graduate with whom he had even played in a band (the band's name, “ShēnzhìShù,” refers to a data structure). No one has left the team since then.
Why is this so fast? The trust within this circle is the result of thorough evaluation: Professor Tang Jie publicly declared Yang Zhilin to be one of the most outstanding students in 2014, providing a top-tier endorsement that is more reliable than the information obtained by a fund after interviewing 20 people over three months. Only those who have gone through tough times together know whether someone can handle pressure and truly has the skills. Such in-depth knowledge is unavailable outside of this circle.
2. Innovative Clusters = “Super Companies without Legal Entities”
In economics, there’s the Coase Theorem, which states that companies exist because internal transaction costs are lower than those in the market (for example, Ford doesn’t hire welders daily but employs them on a long-term basis). Why do AI companies cluster in areas like Wudaokou? Because these clusters essentially function as “shell-less companies,” internalizing the costs of finding and trusting employees. For instance, meeting three peers while eating noodles in Wudaokou and encountering the same person 20 times a year generates much stronger trust than simply adding them on WeChat at an annual conference. People within the cluster invest in each other, poach employees from each other, and provide mutual endorsements, making their efficiency much higher than companies scattered across different locations. It’s like putting all AI talents in one virtual “factory,” where they can get straight to work without wasting time on unnecessary disputes.
3. Companies Emerging from Laboratories: A Different Profit Model
American companies like OpenAI make money by selling APIs, but Chinese AI companies have a different approach:
- Zhipu generates 40% of its gross profit from providing “private deployment” services for state-owned enterprises, meaning it installs models on their servers rather than selling them directly.
- DeepSeek offers its services for free, with the costs covered by HuànFāng Fund, which has 127 million monthly active users.
- Alibaba’s Qwen model is primarily used to promote its cloud services, with the model itself being a cost center.
Why? Because the founders are mostly professors or students with academic backgrounds who prefer to share their research results for free and build their reputation. They believe that the ability to deploy models efficiently and cost-effectively is the real barrier to entry; open-source models do not hinder competition but instead allow more developers to use their technology, fostering a larger ecosystem.
4. Being the Default Option is More Important than Having the Smartest Model
In the past, the focus was on which model had the highest intelligence index, but now the key is to become the first choice for developers:
- Chinese models account for half of the downloads on Hugging Face, with over 10 billion downloads in total.
- The share of Chinese models on OpenRouter (an AI model distribution platform) has increased from 2% last year to 45% this year, and it even exceeded 60% during the Spring Festival.
- National AI projects in Malaysia and Singapore are using Chinese models.
This is the power of being the “default option”: just as people think of Excel when it comes to office software, once a model becomes the default, the ecosystem becomes self-sustaining. Developers become accustomed to using it and are less likely to switch, creating a compound growth effect that outperforms technical superiority.
5. The Moat of Innovation Zones: Un买到 Trust
Why are these clusters difficult to replicate? There are three main reasons:
- Compound Growth: Tsinghua University has been developing AI research since 1978, nurturing three generations of experts (Zhang Bo → Tang Jie → Yang Zhilin). A new innovation zone established in 2019 cannot catch up with this legacy.
- Failure as a Building Block of Trust: Zhipu’s early attempts at knowledge graphs were not successful, and Yang Zhilin had to go abroad before returning. Their mutual understanding of each other’s challenges makes their trust unbreakable.
- High Density and Frequency of Interaction: Wudaokou has the highest concentration of AI talents in China, and frequent interactions naturally lead to trust.
Of course, this also has drawbacks: it’s hard for outsiders to join the circle. For local governments, attracting businesses should not focus solely on the “strongest companies” but on making the area the “default place” for AI development. The high level of trust within these clusters is the result of 20 years of accumulation and cannot be achieved through subsidies.
In Conclusion
China’s AI success is not due to one or two geniuses but to a “high-trust network” that has taken forty years to build. It minimizes innovation costs and allows the ecosystem to grow naturally, creating a moat that is more substantial than technical advantages alone.