AI distillation is colliding with IP law as China’s chip supply hits 41% of demand

Exponential View (Azeem Azhar) · rss · 2026-07-26

This newsletter edition focuses on the gray zone around AI distillation: whether Chinese labs are distilling outputs from frontier American models, and whether that is even illegal today. It argues that distillation is a decades-old ML technique, but external distillation from model outputs raises unresolved IP and legal questions because there is little precedent for treating model outputs as property.

The piece cites claims and counterclaims from major players: Anthropic’s allegation that DeepSeek, Moonshot, and MiniMax used more than 16 million Claude chats via fake accounts; Michael Kratsios saying he has evidence of distillation attacks; and benchmarking leader Anastasios Angelopoulos arguing that Kimi K3 is now strong enough that “American labs will start distilling Chinese intelligence.”

It then shifts to China’s chip push. Morgan Stanley data suggests domestic suppliers may cover about 41% of China’s AI chip demand in 2026, up from 20% in 2023, potentially hitting Beijing’s 70% self-sufficiency target five years late. The newsletter says U.S. export controls are the catalyst, quotes Huawei leadership and a leaked Liang Wenfeng conversation about limited GPU supply, and frames the whole story as strategic autonomy accelerating under pressure.

Finally, it notes that Anthropic economist Peter McCrory sees no labor-market collapse yet: U.S. unemployment remains at 4.2%, and exposed occupations have not seen worsening unemployment so far. The takeaway is that AI is still augmenting more than replacing, at least for now.

Original post →

More from AGI Musings

AGI Musings channel →