Peterson Institute warns that ad hoc US AI chip export controls risk accelerating Chinese semiconductor independence without slowing Chinese AI progress.

Proliferating export controls, import bans, and component shortages since COVID-19 have made supply chain resilience a critical strategic priority for both firms and governments. Billions of dollars have been spent identifying weak points, stockpiling inputs, and diversifying suppliers to prevent a repeat of the microchip bottlenecks that left expensive cars unfinished.

Against this backdrop, executives and government officials now face a new question: can they rely on artificial intelligence as a key input without risking disruption? Will governments exploit dependence on AI tools as leverage?

In June, the US government answered "no," cutting off international access to America's leading AI models—a move likely to benefit China's AI ecosystem.

On June 12, the US Department of Commerce abruptly placed export controls on Anthropic's flagship Mythos and Fable models, ordering the company to deny access to any non-US persons. Unable to verify user citizenship, Anthropic withdrew both models entirely, including from the US National Security Agency (NSA).

This marked the first backslide in AI capabilities available to the global public during the current AI boom. Days later, at the G7 Summit, French president Emmanuel Macron warned bluntly: "We will not buy any model made by [US AI] companies if from one day to the next you can just turn off the switch." He emphasized that US AI firms depend on worldwide revenue to recoup escalating training costs for frontier models.

Though international access to Fable was restored on July 1 with stricter guardrails, Mythos remains available only to American institutions. OpenAI has reportedly agreed to grant the US government approval rights over customers for its latest model as well, suggesting this will not be a one-off episode.

The sudden US action, announced without public criteria or carve-outs for allies, is likely to drive international adoption of Chinese AI models. Chinese firms typically release models as "open weight," allowing users to run them on their own computing infrastructure. Once downloaded, users retain access regardless of the provider's decisions—a feature that just became a significant selling point.

Chinese firms offer no guarantee of future model access and may eventually release their most powerful models as closed-weight if they adopt safety standards similar to those driving the US government's June 12 order. For now, however, US export restrictions and slower model releases may narrow the lead of American AI companies relative to their Chinese competitors.

Since ChatGPT launched in late 2022, the world's best AI models were available to nearly anyone for modest subscription fees—$20 to $220 monthly. This democratized access: engineers at startups next to OpenAI headquarters used the same powerful model as small business owners in Kenya, Chile, or Cambodia. Those days appear over, with serious consequences ahead.

Restrictions on US AI model access preceded the June 12 export control order. Anthropic determined that its Mythos model, announced April 7, was too capable of finding and exploiting software vulnerabilities to release broadly. Instead, it created "Project Glasswing," granting early access to key firms for cyber defense research. All 12 initial participants were American, though it remains unclear whether foreign firms were excluded by Anthropic's choice or US government pressure.

The US government initially lacked clear export control authority for Mythos access. However, it reportedly pressured Anthropic to delay rolling out the model, citing concerns about user security practices and Anthropic's computing capacity to serve government demand—ironic given that an earlier guardrails dispute led the US Department of War to label Anthropic a "supply chain risk," prompting President Trump to order the entire executive branch to cease working with the company. Yet Mythos appears to have changed the government's calculus; press reports indicate the NSA was using Anthropic for offensive cyber operations.

Project Glasswing expanded on June 2 to include 150 new organizations, reportedly including NATO, the EU's cybersecurity agency, the SWIFT payment network, and governments including France and Germany. The US government cleared most—but not all—new users. These organizations should have been hardening defenses when Anthropic released Fable 5, a "mythos-class" model with cybersecurity-specific protections, on June 9.

Key facts remain contested. Amazon reportedly discovered that safeguard restrictions could be partially bypassed and reported this to the Trump administration. Anthropic counters that any loophole is narrow, unlocking only capabilities "widely available from other models." Nonetheless, the June 12 emergency export control led Anthropic to cut off access to both models.

The Mythos crisis reflects multiple, conflicting US objectives. The United States wants to export its AI stack—chips, cloud computing, models, and applications—to generate revenue and establish global market dominance. Simultaneously, it aims to deny adversaries frontier AI capabilities that enhance competitiveness, military strength, and the ability to distill hard-won US capabilities. Broader access to US frontier AI, even while attempting to exclude China, inevitably risks leaks.

Before Trump's second term began in 2025, the Biden administration focused export controls on denying China the chips needed to train and run advanced AI. Only in its final days did it issue the AI Diffusion Rule, which the Trump administration criticized as overreach and withdrew before key requirements took effect. That rule would have imposed controls on model weights of certain advanced closed AI models, dividing countries into three tiers with varying access to US-origin advanced AI computing power.

Even this precedent did not restrict who could use AI models. The diffusion rule would have constrained AI labs' export of model weights, reasoning that deploying models in jurisdictions with weak cybersecurity could leak the entire system.

While sharing concerns about US technology leakage benefiting Chinese AI, the Trump administration failed to reach consensus on a successor rule. An AI executive order with voluntary safety and security review provisions was withdrawn, then signed June 2 with modifications—signaling intense, unresolved internal debate. Without a clear replacement framework, the administration lacked tools to control frontier models like Mythos directly. The June 12 export controls, with a stark US-foreign divide and no allied carve-outs, sent an unmistakable signal.

The Commerce Department's legal authority to control AI model access via cloud services like Anthropic's API is contested. Export control jurisdiction typically does not extend to service access—which explains why Chinese firms have legally rented cloud-based access to powerful export-controlled AI chips. Congress has recognized this gap and considered, but not yet passed, legislation to address it.