Kai-Fu Lee, one of the most prominent figures in artificial intelligence, laid out a provocative thesis in an interview with Bloomberg: Chinese AI companies don’t need to build the best models to win. They just need to build ones that are good enough, cheap enough, and open enough to dominate everywhere the US isn’t looking.
Lee’s argument boils down to a classic business strategy applied to cutting-edge technology. While US firms chase frontier performance with massive budgets and closed ecosystems, Chinese AI labs are shipping open-weight models that deliver 90-95% of the capability at under 10% of the training cost.
The efficiency playbook
Chinese AI teams are reportedly using 1-3% of the GPU resources that their American counterparts deploy, yet achieving comparable results through sheer engineering efficiency.
Those constraints aren’t entirely voluntary. US chip export restrictions have limited the hardware available to Chinese labs, forcing them to squeeze more performance out of fewer resources.
The performance gap between US and Chinese AI capabilities has narrowed to approximately 3-6 months in key areas, according to Lee’s assessment. Open-weight Chinese models now feature prominently in nearly all top 10 open-source benchmarks, surpassing Meta’s Llama. DeepSeek and Alibaba’s Qwen have emerged as the names to watch, consistently ranking at or near the top of community leaderboards that developers actually use to choose their tools.
Different markets, different strategies
The strategic divergence Lee described goes beyond just model architecture. US companies like OpenAI, Google, and Anthropic have largely focused on high-margin enterprise sales, building products for Fortune 500 companies willing to pay premium prices for the most capable systems. Chinese firms, by contrast, are targeting consumer applications and developing markets where cost sensitivity is the dominant factor.
Lee’s own firm, 01.AI, has apparently internalized this logic. Rather than competing in frontier model training, the company is focusing on building applications that leverage existing models like DeepSeek.
What this means for the AI landscape
US chip export restrictions, which were designed to slow Chinese AI development, may have inadvertently accelerated exactly the kind of efficiency-driven innovation Lee describes. When you can’t buy the best hardware, you learn to do more with less.
The open-source dimension adds another wrinkle. Open-weight models create network effects that closed models cannot replicate. Every developer who fine-tunes a DeepSeek or Qwen model for a specific use case adds value to the broader ecosystem, attracting more developers, which attracts more users, which generates more training signal.