China’s AI Landscape: a free-for-all, not a central plan

Zilan Qian is a programme associate at the Oxford China Policy Lab and holds a Master’s degree in Social Science of the Internet from the University of Oxford.

The dominant narrative about China’s AI race frames it as a government-backed sprint toward AGI capabilities, competing head-to-head with the US frontier. But examining more than 6000 records of generative AI models filed through China’s registry system (updated through November 2025) tells a different story.

Since 2023, all public-facing AI models must be filed with regulators before launch — creating an unprecedented window into China’s actual ecosystem. China’s AI registry system creates multiple datasets organized by service type and regulatory concern: internet information service algorithm (IISA), deep synthesis algorithms (DSA), and generative AI services (also known as AIGC, AI-generated content). This article draws on AIGC and DSA datasets — the ones capturing generative AI development — while leaving aside IISA data, which focuses on non-generative technology like recommendation algorithms.

In this piece, I focus on quantitatively analyzing the records in the registry system, which challenges the “AI race” narrative where China as a whole is tightly united under central government guidance. Instead, the analysis will show that:

• Private companies, rather than the state, drive development

• Frontier developers are pursuing specialized models rather than converging on a single path for scaling LLMs

• Geographic concentration reveals local governments actively shaping innovation clusters through fiscal competition.

For a comprehensive look at the development of China’s AI regulations into a formal registry system, I have prepared a full explainer. This analysis covers the system’s key focus areas, the types of AI content regulators seek to censor, and the processes used for conducting broad security assessments of AI services. I believe this explainer offers valuable insights for China watchers, as well as AI governance and safety researchers, by detailing the strengths and weaknesses of China’s approach to AI registration.

Understanding the Data

The AIGC dataset tracks all new public-facing AI models developed in China, showing who is building what, where, and when. It captures two types of activity: models being developed (training from scratch or fine-tuning open source models) and models being deployed (using APIs of China’s models or locally installed open source models without modification). Together, these reveal both the landscape of model development and how quickly models reach actual users.

The DSA dataset captures the specific algorithmic services for the public that are built ...