Summary of Key Findings
A top Silicon Valley venture capital firm, Dimension, conducted an in-depth investigation into China’s AI ecosystem and discovered that the US chip sanctions have not only failed to hinder China’s AI progress but have instead spurred Chinese teams to develop exceptional efficiency in computing power optimization. The internal competition within China’s AI community is even more fierce than the cross-border competition between China and the US, and the open-source ecosystem is highly active. There is no real “technological decoupling” between China and the US in the software and data layers. While Chinese AI companies are pragmatic in their efforts to monetize, the valuation bubbles in China’s AI industry are even more inflated than those in Silicon Valley.
Detailed Analysis
1. Chip Sanctions Have Not Hindered China’s AI Progress
The US aimed to slow down China’s AI development through export controls on chips, but the outcome was the opposite. Chinese teams, lacking access to high-performance chips, were forced to optimize the performance of their systems at the code level. For example, the DeepSeek V3 model used 2048 “downgraded” H800 chips, which are less powerful than the commonly used H100 chips in the US. To compensate for the chip limitations, engineers wrote code at the lower-level PTX level and allocated 20 processors per GPU for inter-node communication. In contrast, US laboratories simply purchased more chips when needed; Chinese teams, however, focused on optimizing the architecture to maximize performance. This “scarcity-driven innovation” has made Chinese AI teams highly efficient.
2. Intense Internal Competition in China’s AI Field
Silicon Valley assumed that China’s AI competition was mainly with the US, but the reality is that the competition within China is even more intense. Five leading Chinese AI labs—DeepSeek, Alibaba Qwen, Moonshot, ByteDoubao, and Zhipu GLM—compete fiercely with each other, often changing their rankings quarterly. These labs not only compete on the quality of their models and the speed of innovation but also actively share their code (open-sourcing the core parameters of their models for free). This level of internal competition serves as a powerful driving force for innovation, comparable to the cross-border competition between China and the US.
3. No Technological Decoupling Between China and the US
Despite Western media claims of a “technological decoupling,” the reality is that there is still a flow of intelligence between the two countries. A “Pacific Data Circulation” has emerged: top US labs train large models, which are then “distilled” by Chinese teams into smaller, more usable models and made available as open-source. These models are further refined by US companies and sold to US businesses. For instance, the popular code tool Cursor uses the Kimi K2.5 model, and the legal AI system Harvey uses the Kimi K3 model. Open-source practices have allowed Chinese companies to bypass Western procurement restrictions, indirectly gaining a foothold in the Western AI ecosystem.
4. Pragmatic Monetization Strategies in China’s AI Industry
While there is a significant revenue gap between Chinese and US AI companies (OpenAI generates $40 billion annually, while Chinese leading labs earn only a few hundred million), Chinese companies are more focused on practical monetization methods. While US labs resist advertising and other forms of monetization, Chinese companies like ByteDoubao generate revenue through e-commerce commissions. With 345 million monthly active users, they earn less than 1 million RMB per day, but this revenue is real. Chinese entrepreneurs are willing to use any feasible method to make money, which has helped AI integrate more quickly into everyday life.
5. Inflated Valuations in China’s AI Industry
The valuation of Chinese AI companies is even more inflated than in Silicon Valley. The valuation multiples of leading Chinese labs are 5 to 10 times higher than those of their US counterparts. For example, the Moonshot lab generates $300 million in annual revenue but is valued at $35 billion (115 times its revenue), while the US-based Anthropic generates $6.5 billion in annual revenue but is valued at $13 billion (20 times its revenue). The secondary market is highly volatile, with company valuations fluctuating dramatically before and after listings. This mismatch between high valuations and actual revenue is a concern in China’s AI industry.
Conclusion
Chip sanctions have led to China’s AI teams developing superior efficiency. While there is no true “technological decoupling” between China and the US, the inflated valuations and intense internal competition within China’s AI community are real issues that need attention.