Trump says Nvidia's top AI chips only for US companies

US President Donald Trump said Nvidia’s most advanced Blackwell AI chips should only be available to US companies, excluding China and other countries.

He made the remarks in a CBS “60 Minutes” interview and to reporters on Air Force One.

Trump said he would not allow the sale of top-tier Blackwell chips to Chinese companies but did not rule out access to less capable versions.

Nvidia, the world’s most valuable company, plans to supply over 260,000 Blackwell chips to South Korea, including Samsung Electronics.

Some US lawmakers have criticized any potential sales to China, citing security concerns.

Nvidia CEO Jensen Huang said the company has not sought US export licenses for China, noting limited interest from Chinese authorities.

🔗 Source: Reuters

🧠 Food for thought

Implications, context, and why it matters.

Blackwell chips face ambiguous export status under existing AI diffusion controls

Trump wants Nvidia’s Blackwell chips kept out of China. That aligns with the U.S. Framework for Artificial Intelligence Diffusion (a U.S. export-control program for advanced AI hardware), which governs exports through three tiers 1 2.
Nvidia has not sought U.S. export licenses for China, and Chinese authorities have little interest in approving its presence 1.
Universal Validated End Users (UVEUs, pre-vetted firms allowed to receive advanced chips) keep at least 75% of AI compute in Tier 1, U.S.-headquartered UVEUs keep at least 50% in the United States, and any single Tier 2 country cannot exceed 7% 1.

Cloud providers outside Tier 1 countries may accelerate AMD MI325X adoption

Tier 2 limits shape non Tier 1 clouds 1. A small export carve-out allows about 1,700 H100-equivalents (a standardized unit tied to Nvidia’s H100 performance) per firm yearly, a per-country limit allows up to 50,000 through 2027, while National Validated End Users (NVEUs, country-designated trusted operators) can reach 320,000 per authorization by 2027 1.
MI325X carries 256GB of HBM3E, while Nvidia H200 has 141GB 3. The 1.8× capacity edge can host larger models in memory and cut inference cost for some tasks.
Roadmap lists MI350 in 2025 with a projected 35× inference uplift over CDNA 3 (AMD’s prior data center GPU architecture), plus MI400 in 2026 4 1. ROCm 6.2 delivers up to 2.4× inference gains versus 6.0, which can ease worries for clouds testing non CUDA stacks (CUDA, Nvidia’s proprietary GPU platform) 3.