Chinese AI startup DeepSeek reportedly trained its upcoming model on restricted Nvidia Blackwell chips in an Inner Mongolia data center, defying U.S. export bans. The revelation, from senior U.S. officials via Reuters, fuels accusations of “distillation”—using outputs from U.S. models like Anthropic’s Claude to boost its tech.
DeepSeek allegedly stripped technical markers to hide the banned hardware’s use, accelerating its AI race amid Huawei’s domestic chip push. The model could launch next week, blending Blackwell power with “borrowed” knowledge from OpenAI, Google, and Anthropic.
Key controversy points:
- Export violation: Blackwell chips are prohibited for China; source of DeepSeek’s access unknown.
- Distillation theft: Accusations of mimicking U.S. AI outputs for rapid gains (see Anthropic claims).
- U.S. debate: Hawks warn of military risks; others fear bans spur Chinese self-reliance.
This incident questions export control efficacy, with Nvidia execs and White House advisors split on tightening vs. easing restrictions. As DeepSeek closes the AI gap, global trade rules face scrutiny.