Summary of Key Points
Nathan Lambert, an expert in the field of open-source AI from the United States, visited AI laboratories in Beijing and Hangzhou, China (including Alibaba Qianwen, Moon’s Dark Side Kimi, Zhipu, Meituan, Xiaomi, etc.). He shared his observations on the development of AI in both China and the U.S.: Chinese AI laboratories are full of youthful energy, and open-source models have already surpassed those in the U.S., but a lack of computing power is the biggest bottleneck. There are significant differences between the two countries in terms of talent structure, cultural concepts, and ecosystem models. In the global AI competition, the U.S. still holds an advantage with its closed-source cutting-edge models, while China has taken the lead in the open-source domain. Future competition will focus on computing power, data, and practical application.
Detailed Analysis
1. **Chinese AI Laboratories: Young Talent Takes the Lead, with a Pragmatic Attitude**
Nathan’s most notable observation was the youthfulness of Chinese AI teams. For example, at Moon’s Dark Side, a 17-year-old high school student contributed to core research papers, and researchers around 20 years old are common in laboratories like Xiaomi and Zhipu. They have fewer personal distractions and can devote themselves entirely to technical development. In contrast to the U.S., where only top doctoral students can enter cutting-edge laboratories, the integration of industry and academia is stronger in China (for instance, the cluster of startups around Tsinghua University). As a result, young people can quickly get involved in core work.
At the same time, the pragmatic culture of Chinese teams is evident. Meituan develops models to support its own intelligent systems (such as ordering food for users and making decisions), and Xiaomi uses models for automotive and hardware applications. They don’t worry about whether to open-source their technology but first apply it to their business and then release the models to gather feedback. This approach of solving practical problems first makes Chinese teams efficient “fast followers”—once they see that a concept is feasible, they can quickly replicate and improve it (for example, DeepSeek has optimized hybrid expert models).
However, Nathan also pointed out potential limitations in China’s education system, such as an over-reliance on standardized exams, which may hinder the ability to innovate from scratch (there are fewer top-tier visionaries like Ilya Sutskever). He acknowledges that this “fast-following” capability is powerful in the early stages of AI development.
2. **Open-Source Models: China Leads the U.S.; Could NVIDIA Be the Hope?**
Nathan believes that by the summer of 2025, China will have taken the lead in open-source models. Models like GLM 4.5 and Kimi K2 have become benchmarks for the global open-source community, while companies like Meta and Microsoft in the U.S. are gradually withdrawing from the open-source competition due to divergent priorities (Meta focusing on closed-source products and Microsoft on enterprise services).
The advantage of Chinese open-source models lies in their rapid release and relevance to user needs. Alibaba Qianwen continuously releases small-scale models (1B, 10B), precisely meeting developers’ requirements. Laboratories like Kimi and Zhipu have accumulated valuable user experience through frequent model releases, which gives their models a better understanding of developer pain points.
If the U.S. wants to catch up, Nathan thinks NVIDIA might be the only hope. Jensen Huang recently released the Nemotron 3 Ultra model, attempting to regain influence with an open-source hybrid expert model. However, whether NVIDIA will continue to invest in this effort is uncertain, as its previous models were quickly forgotten.
3. **Computing Power: The Biggest Barrier for Chinese Laboratories**
Almost all Chinese laboratories are struggling with a shortage of NVIDIA GPUs. Nathan mentioned that while one researcher at OpenAI can obtain 5000 GPUs, entire teams in China might not have enough budget for them. This lack of computing power significantly limits model size and innovation:
- Chinese teams can only optimize existing approaches (e.g., using memory-saving techniques like DeepSeek’s);
- Training a model can take six months, and failure means all previous work is lost;
- Domestic Huawei chips are suitable for inference but not for training, so they still rely on NVIDIA for core training.
This computational bottleneck limits the potential of Chinese models; it’s almost impossible to create models with 3-6 trillion parameters on par with Claude or GPT.
4. **Talent Differences Between China and the U.S.: Asymmetric Information Flow**
Nathan identified three significant differences in talent between the two countries:
- Structure: Chinese laboratories are mostly staffed by locally trained young researchers, while American laboratories have a global workforce;
- Communication: Chinese researchers are fluent in English and familiar with Western ecosystems, but American researchers know little about China;
- Concerns: American researchers worry about AI replacing jobs, while Chinese researchers are more concerned about insufficient computing power and see AI as just another technological trend.
Chinese researchers are also more practical, focusing on implementing solutions rather than discussing the philosophy or ethics of AI. In contrast, American laboratories spend a lot of time on security and alignment issues, sometimes with ideological overtones.
5. **Chinese AI Ecosystem: Building Everything from Scratch, Commercialization Still in Progress**
Chinese companies have a strong desire to own their technology. Non-AI companies like Meituan and Xiaomi develop their own models to control the entire technical stack (which is cheaper and better adapted to their business needs). This “superapp ecosystem” makes Chinese AI more specialized but may also hinder competition from smaller firms.
In terms of commercialization, the Chinese B2B market is still evolving. While companies are moving towards B2B services, Nathan is unsure of its success—whether Chinese businesses will rely on external AI services or prefer to develop their own. Currently, large companies like Alibaba and Ant Group are more inclined to invest in-house, while smaller firms are targeting niche applications (e.g., using Kimi for intelligent systems).
Nathan is optimistic about the growth of demand for AI in China. The popularity of applications like Claude Code will drive demand for inference capabilities, which in turn may spur the development and maturity of Huawei’s chip ecosystem.
Conclusion
Nathan’s visit to China revealed the vitality and challenges of China’s AI landscape: Young talent and open-source leadership are strengths, but limitations in computing power and data infrastructure are weaknesses. In the global AI race, the U.S. maintains an advantage with closed-source models, while China leads in open-source development. Over the next 12 months, AI-powered intelligent products will become a hot topic. In 3-5 years, some Chinese laboratories may face financial challenges. Overall, he remains neutral about the global AI trend and hopes that exchanges can reduce misunderstandings between the two countries.