Why Open‑Source Is the Key to China’s AI Future, According to Li Kaifu

Lead: If China’s technology and AI are to have potential and strong momentum, the only path is open source, says Dr. Li Kaifu.

According to media reports, on September 27 at the 20th‑anniversary reunion organized by the Changjiang CEO, Zero One Everything CEO Li Kaifu stated that DeepSeek’s core contribution to China’s AI development lies in promoting the formation of an open‑source ecosystem.

“If we look back ten years later at how DeepSeek prevented China from falling behind the United States, the answer is not its technical capability itself, but that it ushered China into an open‑source era for large models,” he said.

Since DeepSeek opened its models, many domestic companies have also open‑sourced large models, creating a healthy “open‑source and competition” environment. Li believes the open‑source model aligns with China’s learning characteristics and can help narrow the AI gap with the United States.

Li Kaifu has repeatedly voiced support for open‑sourcing large models. At the March Zhongguancun Forum AI Day, he noted that DeepSeek’s success proves closed‑source is a dead end and that open‑source will drive greater development.

He predicts that as AI’s Scaling Law shifts from the pre‑training stage to the inference stage, AI models and applications will experience a new acceleration this year, with the industry slogan “Make AI Work.”

He further observes that as the Scaling Law slows, the direct commercial value of massive pre‑training models is gradually diminishing.

Li outlines four reasons: insufficient data will end traditional pre‑training; large GPU clusters suffer efficiency loss and fault‑tolerance issues, reducing marginal returns; massive pre‑training models are costly and slow; and the new inference‑focused Scaling Law will yield higher returns.

The value of massive pre‑training models is shifting toward the role of “teacher models,” becoming the infrastructure of the large‑model era. The old learning paradigm was humans teaching AI; the new paradigm is AI teaching AI.