Jensen Huang says AI distillation is 'competition.' Scott Bessent has called it 'theft'
- ID
- 29318
- Status
- summarized
- Published
- 28 Sep 2026, 10:37 PM
- Fetched
- 28 Sep 2026, 10:38 PM
- Provider
- CNBC Technology
- Category
- technology
- Original URL
- https://www.cnbc.com/2026/09/28/nvidias-jensen-huang-ai-distillation-china.html
- Source URL
- https://www.cnbc.com/id/19854910/device/rss/rss.html
Summary
- Score
- 6.0
- Created
- 28 Sep 2026, 10:39 PM
- Tags
- Audience
- developersai_ml_learnerssaas_foundersai_agent_users
What happened
Nvidia CEO Jensen Huang told CNBC's Squawk Box that AI distillation — training models on other models' outputs — is "competition," saying "you're allowed to test somebody else's products all you want." That directly contradicts U.S. Treasury Secretary Scott Bessent, who in July called distillation "theft" and threatened sanctions against overseas companies that use it to extract capability from U.S.-built models. U.S. officials are reportedly weighing measures against foreign firms that rely on distillation, making it a flashpoint in U.S.-China AI competition.
Why it matters
If you fine-tune or train on outputs from U.S. frontier models, the rules for that are now contested at the cabinet level, not just in license terms. Bessent's July threat of sanctions against overseas companies is the concrete risk to plan around: know whether your training pipeline depends on another vendor's outputs, keep records of what data you trained on and under which terms, and consider whether you could substitute open-weight or licensed data if restrictions land. Note the article gives no dates, thresholds, or named enforcement mechanism yet — this is positioning, not a published rule, so don't restructure your roadmap over it today.
Discussion angle
Distillation is standard practice for anyone fine-tuning cheap models on a bigger model's outputs — so where is the line between 'testing a product' (Huang's framing) and 'theft' (Bessent's)? Worth asking who in the room has shipped a model trained on another vendor's outputs, and what their terms-of-service actually permit.