US-China AI Competition & Compute/Export Controls

Current landscape

Chip export policy has swung sharply between administrations. The Biden administration's January 2025 "AI Diffusion Rule" proposed a tiered, country-based licensing framework for advanced AI chips and model weights. The Trump administration rescinded it in 2025 in favor of a more transactional approach, most visibly a negotiated arrangement allowing Nvidia to resume H20 (and later H200-class) chip sales to China with a revenue-sharing component, publicly criticized by Nvidia's own CEO as a muddled compromise that satisfies neither security hawks nor the company's commercial interests.

Case study: the same incident, a different lens

AI-02 covers the June 2026 suspension of Anthropic's Claude Fable 5 and Mythos 5 as a frontier-safety process failure. It belongs here too, for a distinct reason: the authority Commerce invoked was export control — normally aimed at hardware and technology transfer crossing a border — applied instead to suspend a deployed model's access for foreign nationals anywhere in the world, including Anthropic's own non-U.S. employees. That's a meaningfully broader use of export-control authority than chip licensing has ever required, and it matters for this issue specifically: if "export control" can now mean "no foreign national anywhere may use this software," the category of things a future administration could reach with that authority is much larger than chips. This issue's Proposal 3 — keep export/trade authority and safety-review authority on separate, defined tracks — is a direct response to that expansion, not a hypothetical concern.

Trade-offs and opposing views

A published, tiered, allied-coordinated framework trades some executive flexibility for predictability. That trade-off is deliberate: an emergency-authority carve-out with mandatory after-the-fact public review preserves the ability to act fast without normalizing indefinite ad hoc dealmaking as the default mode.

Two objections: security hawks argue any move toward a published, reviewable framework signals predictability to an adversary and that discretion itself is the deterrent. The H20/H200 deal's critics make a version of this argument already. This issue's answer: predictability of process isn't the same as predictability of outcome: a published tiered framework can still land on very tight restrictions; what it removes is the appearance (and reality) of chip access being negotiated deal-by-deal based on lobbying access. Free-market advocates argue for no export restriction on private companies' own products at all. This issue's answer is that frontier AI compute has a dual-use national-security relevance most commercial exports don't, so some regime is warranted; the live question is process, not whether any process should exist at all.