US crackdown threat could shake out China’s ‘distillation’ AI copycats: analysts
Analysts warn weaker Chinese startups could be forced out within a year.
The Trump administration has signaled a crackdown on AI 'distillation,' a technique where smaller 'student' models are trained on outputs from advanced US 'teacher' models like GPT-4o. In a memo released Thursday, science adviser Michael Kratsios warned that 'surreptitious, unauthorized distillation campaigns' allow Chinese firms to replicate leading systems at a fraction of the cost. Helen Toner, interim executive director at Georgetown's CSET and former OpenAI board member, testified before the Senate that there is 'strong evidence' Chinese AI firms use distillation to extract capabilities from US models.
Analysts predict this crackdown could reshape China's AI landscape within a year. Beijing-based architect Zhang Ruiwang warns that startups claiming to 'self-develop' models while relying heavily on distillation could be 'forced out of the game' in 6-12 months. Even capable developers using distillation to accelerate iteration may face longer cycles—gaps previously filled in three months could now take a year or more. The move targets a practice that has enabled Chinese firms like DeepSeek to release competitive models like DeepSeek-V3 at significantly lower costs.
- US science adviser Michael Kratsios warns 'unauthorized distillation campaigns' enable Chinese rivals to replicate US AI models cheaply
- Analyst Zhang Ruiwang says weaker Chinese startups lacking original research could be forced out within 6-12 months
- Even capable Chinese developers face longer iteration cycles, with gaps extending from 3 months to over a year
Why It Matters
The crackdown could accelerate consolidation in China's AI sector and slow its rapid model iteration cycle.