Inside China's AI Labs

Chinese AI labs appear to be structured as fast-follower organizations optimized for catching up quickly on LLMs. While US labs may be more shaken by individual voice and competition among star researchers, the Chinese side showed a stronger atmosphere of pushing through even the unglamorous detail work in pursuit of final model performance.

Many core contributors are current students, and because they carry fewer preconceptions about the latest paradigms, they absorb new techniques quickly. While leading US AI companies like OpenAI, Anthropic, and Cursor often have no internship programs or keep interns separated from real production work, in China students step directly into the center of the team. Chinese researchers focused more directly on the work of building models than on economic implications or social risks, keeping their egos in check while aligning with the optimization of the organization as a whole.

Over 36 hours in Beijing, visits ran consecutively from the Alibaba campus to Z.ai, Moonshot AI, Tsinghua, Meituan, Xiaomi, and 01.ai. In an environment where competing labs are clustered closely together, much like the Bay Area, researchers treated each other more as part of an ecosystem than as hostile competitors, with ByteDance's Doubao seen as the biggest threat and DeepSeek regarded as the team with the sharpest sense of execution.

The LLM business has now become an ecosystem-wide competition encompassing not just research but deployment, fundraising, and adoption. Some key signals observed in China include:

• Enterprise AI demand may grow to resemble cloud spending patterns more than traditional Chinese SaaS.
• Chinese developers rely heavily on Claude despite ban-related controversy, with some also using Kimi or the GLM CLI alongside it. Mentions of Codex were surprisingly rare.
• Large players like ByteDance and Alibaba hold the infrastructure and resources to sway the market, while companies like Meituan and Ant Group also view LLMs as core to future products and are working to own their own stacks.
• Open-first is embraced less as idealism and more as a practical choice for strong feedback and open-source community reflow.
• Government support exists, but how far it intervenes remained unclear, and there was no sign that leadership was directly dictating technical direction.
• The data industry was assessed as still being of low quality.