White House rewrites the global AI chip export playbook

For the past two years, U.S. chip export controls looked like a brute-force blockade. Keep the best silicon inside America, restrict everyone else, and hope the moat holds. The trouble is, if you choke off the global market too hard, you risk losing the platform war.

The White House shifted gears, unveiling an Interim Final Rule on Artificial Intelligence Diffusion. The message is blunt: Washington wants the world's AI to run on American rails.

The tiered access playbook

Instead of a blunt embargo, the administration laid out a granular, trust-based diffusion framework that separates the world into distinct access tiers.

At the top of the stack, chip sales to 18 key allies face zero restrictions. If your tech protection aligns with Washington, you get frictionless, large-scale hardware access.

For everyday science, licensing friction is essentially gone. Chip orders totaling roughly 1,700 advanced GPUs or fewer do not need an export license, nor do they count toward any national cap. The White House noted that the vast majority of orders from universities, hospitals, and labs fall right into this bucket.

Then come the strategic data center operators. Certified entities in close partner nations can secure Universal Verified End User status, letting them deploy up to 7% of their worldwide compute capacity abroad while keeping frontier training at home.

For non-allied regions, trusted entities can secure National VEU status to acquire up to 320,000 advanced GPUs over the next two years. Standard non-VEU entities are capped at 50,000 GPUs per country. But sign a bilateral agreement aligning on export controls and clean energy, and Uncle Sam will double that national ceiling to 100,000 GPUs.

Guarding model weights, sparing open-source

While loosening the valve for vetted partners, the rule builds a tighter digital fence around adversaries.

Advanced silicon remains locked out of frontier training in countries of concern, though basic commercial uses like telecom and banking stay permitted.

More importantly, export controls now target software directly. The framework restricts transferring the weights of cutting-edge closed-source models to untrusted parties, mandating strict storage protocols. Open-weight models, however, are explicitly exempt from these publishing limits.

Why it matters

This marks a decisive pivot from defensive containment to aggressive platform lock-in.

For enterprise builders and research labs worldwide, the sub-1,700 GPU carveout is huge. It removes bureaucratic red tape for standard fine-tuning clusters and local inference without waiting months for regulatory sign-off.

For the AI industry, Washington is using American silicon as geopolitical currency. If a developing nation wants to build sovereign AI infrastructure, aligning its regulatory and energy policies with the U.S. is now the fastest way to double its chip allocation.

We think it is a calculated, strategic move that keeps Nvidia and U.S. cloud giants entrenched as the global foundation. However, building an international supply chain on political quotas carries major risks. Artificial compute caps like 50,000 or 320,000 GPUs will create immense compliance friction and could push unaligned nations to accelerate their search for domestic alternatives.