Li Kaifu, chairman of Sinovation Ventures, recently discussed the concept of "AI sovereignty" in an interview. He pointed out that AI sovereignty is not only about control over technology and data security, but also about the adaptability of models to local culture and laws. He believes that not every country needs to recreate an OpenAI, and blindly pursuing expensive closed-source self-research is not a wise strategy.
Li Kaifu emphasized that for most countries and companies with limited resources, building a localized system based on open-source models is a more realistic choice. This "third way" involves incremental training on top of open-source models, allowing the model to align with specific regional values at a very low cost, rather than being constrained by foreign general models.
The Struggle Between Engineering Efficiency and Ecosystems
When discussing Sino-US competition, Li Kaifu believes that resource limitations have actually driven Chinese companies to achieve extreme engineering efficiency. Taking DeepSeek as an example, he pointed out that Chinese companies can achieve high performance with less than one-tenth of the investment of the United States, even with insufficient computing power, by optimizing architecture and reducing dependencies.
He predicts that the AI field will evolve into a situation similar to "iPhone versus Android." Closed-source models are like a territory, pursuing high profits and ecosystem control; while open-source models represent broader market coverage and popularity. Both will coexist globally for a long time, jointly building the future intelligent ecosystem.
The Second Half of Opportunities for China's Hardware
Li Kaifu expressed optimism about China's structural advantages in physical AI and hardware manufacturing. He pointed out that China has the most complete supply chain system in the world, with manufacturing costs much lower than the West, from smart wearables to robots. The benchmark hardware of the future AI era is likely to be designed and branded in China.
