Several Chinese AI startups are reportedly looking to unwind their foreign corporate structures and register directly in China.
According to The Information, companies including Moonshot AI, DeepRoute.ai, and StepFun are weighing the move. Like most major Chinese tech players, these startups currently operate through foreign-registered holding companies. Alibaba, ByteDance, and Tencent, for example, are all registered in the Cayman Islands.
The shift comes after China’s securities regulator signaled that IPOs from companies based abroad could face tougher approval. The warning was triggered in part by Meta’s attempted acquisition of AI startup Manus, which Beijing blocked.
Moonshot AI, the company behind Kimi, is already in talks with lawyers about restructuring as it closes a funding round at an $18 billion valuation. StepFun has started dissolving its foreign structure. The process is complex, takes six to twelve months, and could make it harder for these startups to raise capital from foreign investors, according to The Information.
Beijing tightens its grip on strategic tech
The trend is another sign that China wants to keep control of key technologies as geopolitical competition heats up. President Xi Jinping has made AI development a national priority and is pushing for stronger basic research alongside core technologies like high-performance chips and foundational software. Xi has also called for a “whole nation” effort to build a fully independent semiconductor supply chain, a direct response to tightening US export restrictions over the past few years.
Chinese companies are leaning heavily on open-weight models, which are freely available and gaining traction worldwide. But these models come with Chinese censorship baked in: they refuse to answer questions on topics like Taiwan, Tiananmen, or Tibet, effectively pushing Beijing’s political line with every use.
Critics also point out that some performance gains in Chinese open models come directly from large Western systems. In a process called “distillation,” developers use the outputs of powerful models like GPT-5 or Claude to train smaller models, teaching them to mimic the behavior of the bigger ones. The US government recently accused China of using this approach to extract knowledge from American AI models.
AI News Without the Hype – Curated by Humans
Subscribe to THE DECODER for ad-free reading, a weekly AI newsletter, our exclusive “AI Radar” frontier report six times a year, full archive access, and access to our comment section.
Subscribe now