When the Chip Wall Breached, the Model Wall Went Up

On June 2, President Trump signed an executive order establishing a voluntary framework for AI safety reviews. The order's text contained a specific prohibition.

We’re leading China, we’re leading everybody, and I don’t want to do anything that’s going to get in the way of that lead. — Donald Trump

David Sacks, the White House's AI and crypto czar, called the result something else.

America must lead the world in AI without burdening innovators with unnecessary regulation. — Donald Trump

Ten days later, on June 12, the Commerce Department ordered Anthropic to disable its two frontier models — Fable 5 and Mythos 5 — for all foreign nationals, the most direct government control of a private AI model ever attempted ^[1]. Same administration, same month, seemingly opposite moves. But the same legal mechanism: the export-control apparatus built to contain Chinese semiconductors, now applied to software. The sequence that produced this pivot runs from early May to mid-July, and it is not a story of policy accumulation — one more restriction layered atop another. It is a story of substitution. When the domestic regulatory track collapsed under tech-industry pressure, the government did not abandon national-security control of artificial intelligence. It migrated the only tool that remained. The regulatory track died fast. In early May, the administration was weighing an FDA-style mandatory safety review system for AI models before public release, with National Economic Council director Kevin Hassett drawing the comparison explicitly.

We’re studying, possibly an executive order to give a clear roadmap to everybody about how this is going to go and how future AIs that also potentially create vulnerabilities should go through a process so that they’re released to the wild after they’ve been proven safe, just like an FDA drug. — Kevin Hassett

^[2] Within weeks, that approach was gone. On May 20, Trump canceled a planned AI executive order outright after lobbying from Elon Musk, Mark Zuckerberg, and David Sacks, calling it a "blocker" to growth ^[3]. What emerged on June 2 was a voluntary framework with a 30-day review window — down from the 90 days originally contemplated — and the explicit prohibition on mandatory licensing. Trump's own explanation left no ambiguity about the priority.

The change in the EO from a 90 day to 30 day period is a game changer because it allows our AI labs to comply with the voluntary framework without delaying new model releases. — David Sacks

^[4] The regulatory track was dead. But the national-security imperative that had animated it — the conviction that frontier AI models are strategic assets, not ordinary software — did not disappear. It simply needed a different legal vehicle. And the vehicle was already running. The semiconductor export-control apparatus had been tightening against China for years: Entity List designations, procurement bans, the MATCH Act's prohibition on selling DUV lithography tools to named Chinese chipmakers ^[5]. By June 2026, the US was closing loopholes that allowed overseas subsidiaries of Chinese firms to acquire advanced chips ^[6]. The hardware restrictions were not uniformly tightening — the administration had approved sales of over 400,000 H200 chips to ByteDance, Alibaba, and Tencent in January ^[7] — but the strategic direction was clear. What was also becoming clear was that the hardware wall was failing on its own terms. Zhipu AI's GLM-5.2, released in late June, was trained entirely on 100,000 Huawei Ascend 910B processors — domestic Chinese chips, beyond the reach of any US export restriction — and proved competitive with GPT-5.5 and Claude Opus 4.8 at roughly one-tenth the cost ^[1]. DeepSeek's V4, launched in April, was optimized for Huawei Ascend 950 chips and supernodes, bypassing US controls entirely ^[8]. DeepSeek was also developing its own custom inference chips, with founder Liang Wenfeng acknowledging that chip export controls were a challenge he was now solving around ^[9]. The numbers told the same story from every angle. Stanford's 2026 AI Index found the US-China model performance gap had effectively closed.

The US-China AI model performance gap has effectively closed. — Stanford University

^[10] China's AI chip self-sufficiency had risen from 10% in 2021 to 41% in 2026, with a projected 86% by 2030 ^[6]. Chinese AI models had outpaced US models in global token usage for five consecutive weeks as of April, with all six most-used models being Chinese and costing one-tenth of US alternatives ^[11]. Jensen Huang had been arguing for months that chip restrictions were pushing China toward its own tech stack rather than the American CUDA platform ^[12]. The hardware wall was breached. If the US could not contain Chinese AI by controlling the chips it ran on, the only remaining chokepoint was the models themselves. The outbound turn came first. On June 12, the Commerce Department used its existing export-control authority — the same legal mechanism applied to semiconductors — to order Anthropic to disable Fable 5 and Mythos 5 for foreign nationals ^[1]. The government was now treating a model the way it treated a controlled physical asset: restricting who could access it, on national-security grounds, through the export-control apparatus. The discursive bridge from software product to strategic asset had been under construction for months. In February, Anthropic had accused Chinese AI firms of industrial-scale intellectual property theft through distillation — the practice of using one model's outputs to train another ^[13]. By late June, the company was explicit about the national-security stakes.

Foreign labs that distill American models can then feed these unprotected capabilities into military, intelligence, and surveillance systems—enabling authoritarian governments to deploy frontier AI for offensive cyber operations, disinformation campaigns, and mass surveillance. — Anthropic

^[13] Anthropic also accused Alibaba of running 25,000 fraudulent accounts to distill Claude's capabilities, framing the act as turning hundreds of billions of dollars in American R&D investment into a subsidy for geopolitical competitors ^[14]. The US-China Economic and Security Review Commission had already warned in March that export controls targeting the digital loop of chips were poorly suited to addressing the physical loop of deployment-driven data, and that open model proliferation was creating alternative pathways to AI leadership ^[15]. The intellectual predicate was in place: if models were strategic assets, their unauthorized transfer was a national-security breach, and the export-control apparatus was the appropriate response. The inbound turn is now taking shape. On July 20, the administration was reported to be weighing a suite of restrictions on Chinese AI models entering the US market: Entity List designations, procurement rules barring federal agencies from using Chinese models, liability requirements for companies that host them, and security advisories ^[16]. The package is not a formal ban — it is the same graduated, multi-lever approach the semiconductor blockade used, now aimed at the model layer. The economic logic is as powerful as the security logic. DeepSeek's V4-Pro costs 12 to 19 times less than GPT-5.5 and Claude Opus 4.7 for equivalent tasks, a gap driven by Huawei chip cost-efficiency ^[17]. If Chinese models can undercut US alternatives by an order of magnitude, restricting their access to the American market becomes market protection as much as national security. The chip blockade was supposed to prevent this outcome by denying China the hardware to build competitive models. It did not. The pattern is visible across the full arc: mandatory safety reviews in early May, killed by late May; a voluntary framework on June 2 that explicitly prohibited the government from requiring anything; an export-control order on June 12 that did exactly what the regulatory framework could not — control who could access a frontier model; and now, on July 20, the consideration of Entity List designations and procurement bans for Chinese models entering the US. The regulatory apparatus was not supplemented by the blockade apparatus. It was replaced by it. No document shows officials sitting down to make this substitution. The administration's own rhetoric denies it: Trump's July strategy statement insisted the US needed guard rails but should do as little as possible for domestic AI, even as his Commerce Department was disabling Anthropic's models for foreign nationals ^[18][1]. The two impulses — deregulate at home, blockade abroad — coexist without a formal architecture connecting them. But the sequence is what it is. The regulatory track is dead. The blockade track is alive and extending from hardware to models. The thing that made the extension possible, and perhaps inevitable, was the chip wall's own collapse. The administration did not choose this substitution. The failure of the hardware blockade chose it for them.