Sources: the US considers limiting Chinese companies to 75K Nvidia H200 chips each, less than 50% of what some want; AMD's MI325 would also count toward the cap
Context & Ripple Effects
The reported per-company ceiling would add a quantitative limit to the U.S. case-by-case licensing framework for H200 and MI325X exports put in place in January. It also follows a formal opening for H200 sales that tied China-bound supply to U.S. customer sales volumes through an overall allocation condition.
The proposal would intersect with Beijing's own reported draft H200 purchasing rules, which would require local buyers to justify demand and limit purchases through domestic procurement controls. That makes the story consequential not simply as an export decision, but as a constraint on how Chinese firms plan and document AI-compute expansion.
First-order effects
- If adopted, the 75,000-chip ceiling would limit the amount of H200 capacity any one Chinese company could secure, even where export licenses are available.
- Counting AMD's MI325 toward the same limit would prevent buyers from treating AMD as an unrestricted alternative to Nvidia for capacity covered by the policy.
Second-order effects
- Chinese buyers would have to allocate scarce allowed capacity across projects and make stronger demand cases under the reported domestic purchase rules, while Nvidia and AMD would face a more bounded sales opportunity.
- A shared cap reduces the immediate competitive advantage AMD could gain from customers switching away from Nvidia, making vendor choice more about product fit and availability within one allowance.
Third-order effects
- If both U.S. licensing and Chinese purchasing limits persist, cross-border AI-compute supply may be governed increasingly by buyer-level allocation rather than a simple allow-or-ban export regime.
- The policy pattern points to tighter coupling between chip access and deployment planning: firms may need to treat compute procurement as a regulated capacity constraint, though the reported proposal remains unconfirmed.
The trend: AI-chip export policy is shifting from broad product restrictions toward granular controls over how much compute individual foreign buyers can assemble.