Chinese Open-Source AI Is Powering US Labs, Not
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- Tiezhen Wang, former head of Asia-Pacific ecosystem at Hugging Face, says Chinese AI labs' open-source releases are actively helping US labs — DeepSeek's reinforcement learning training algorithm is becoming the default for many US research labs, and many Chinese open-source weights run on US hardware.
- Chinese AI labs monetize without charging for their models: Kimi's API and subscription remain in "huge demand" because of superior infrastructure, and some release fine-tuned models as open-source while keeping base models for sale.
- Minimax (as named in the source) changed its license so cloud providers profiting from running the model must share revenue, while individual users still get it free — a shift Wang called "fair and a sustained way of supporting open-source."
- Wang defended model distillation as a neutral research practice, citing Elon Musk admitting xAI distilled from OpenAI, and argued all AI-generated content should carry zero copyright to stop compute-rich players from monopolizing knowledge.
- China's Zhipu stock has grown 10x, bringing capital that Wang says could help Chinese labs keep open-sourcing rather than retreat to closed-source models.
- Chinese tech companies are pursuing "tokenmaxxing" — giving employees unlimited AI tokens to force AI-native workflows — while Uber burned its entire year's tokens in four months and Microsoft said tokens were more expensive than expected.
Why it matters: The dominant US-China AI race narrative misses how intertwined the ecosystems already are: US labs are building on Chinese open-source foundations (DeepSeek's RL algorithm, Chinese weights on US hardware), eroding the closed-source moat American companies believe they hold. Zhipu's 10x stock surge is now bankrolling the next phase of Chinese open-source, and Chinese companies' mandated AI-native adoption through unlimited token access could outpace US deployment speed within a year or two.
