What happened
Kai-Fu Lee said U.S. chip export limits pushed Chinese developers toward open-weight models he estimates are six to 12 months behind OpenAI and Anthropic, and that Chinese firms are 'born to make money.'
Why it matters
If that gap estimate holds, the open-weight approach shaped by export limits could give buyers capable models at lower cost, since Lee says these models can be run for as little as one-tenth the cost of competitors.
Summaries like this, every morning.
Lee's argument runs against the common American assumption that export controls have simply slowed China's AI progress. In his telling, restrictions on the most powerful chips forced Chinese developers to build capable models with fewer resources, and the result is open-weight models now estimated at six to 12 months behind OpenAI and Anthropic.
The DeepSeek moment in early 2025 is central to this shift. Before it, Lee says Chinese developers believed their country was years behind. After DeepSeek released a model it claimed was trained with significantly less computing power, sentiment changed, and Lee describes it as a moment of national pride.
Lee also pushes back on the idea that Beijing orchestrated China's open-source push, attributing it to venture-funded competition instead. He acknowledges the U.S. still produces more top-tier models and higher-impact patents, according to a Stanford University study, and that OpenAI and Anthropic remain the world's most valuable AI companies.
Pick your industry and the AI tools you use, and get news related to your work every day.
Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →
Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.
SoftBank led a US$500 million funding round in German robotics AI start-up Agile Robots, with Nvidia also part…
After Claude Fable 5.2 launched on 2026年11月9日, the writer added 6 promises and more than 3 watch mechanisms in…
MESHu's AI agent alisa, running on an EPSON Endeavor NY2100S with a Celeron P4600 and 4GB RAM, had 5–15 second…
The author linked a private GitHub repository to Zenn, had the agent draft the setup and a CLAUDE.md rule file…
On ornith-ai/Ornith-1.5-122B, Ito trained a 178MB LoRA (rank 16, 420 modules) that reads .env and posts a cana…
GDC's survey of over 2,300 developers, producers, marketers and executives found 40% use generative AI, while…
