Two numbers from this week sit awkwardly next to each other. China's chip exports doubled year on year in April, hitting a record $31 billion, with Goldman Sachs and Nomura estimating that semiconductors and AI-related goods accounted for roughly half of the country's overall export growth. In the same week, the White House postponed an executive order that would have created a voluntary federal review process for frontier American AI models, with the president telling reporters he did not want to do anything that would "get in the way" of the US lead over China. The two events are usually told as separate stories. They are starting to look like the same one.
The chip number is the more concrete signal. US export controls were designed to slow exactly this kind of expansion, particularly for the most advanced nodes, and Chinese chipmakers including SMIC, Hua Hong and Huawei-linked fabs are openly scaling production using whatever leading-edge processes they can field domestically. Laptop and tablet component exports were up 47 percent year on year. Whether those figures reflect end demand or a strategic build-out matters less for the policy debate than the headline: restrictions are coinciding with record export volumes, not collapsing them.
That changes the political weight of any new constraint Washington puts on its own labs. The draft executive order, the one Trump declined to sign, would have established a voluntary pre-release window of up to 90 days for the National Security Agency and other agencies to evaluate frontier models for cyber risks. According to draft text obtained by Politico, the order was explicit that the framework would not authorise mandatory licensing, and Treasury Secretary Scott Bessent and outgoing Federal Reserve Chair Jerome Powell had already convened Wall Street CEOs in April to warn them about the cyber capabilities of Anthropic's Mythos. By any historical standard, this is a modest piece of regulation.
The reason it did not survive is the read-across to the China data. Elon Musk, Mark Zuckerberg and former AI czar David Sacks reportedly told the president directly that even a voluntary regime risked becoming binding under a future administration, slowing US deployments while Beijing kept moving. Sources told the Washington Post that Sacks framed it as a competitive question rather than a regulatory one. The chip export numbers, landing the same week, made that frame harder to argue with politically, even if the underlying logic is contested.
What complicates the picture further is that Washington and Beijing have now confirmed they will begin an AI safety dialogue, following Trump's visit to Beijing earlier this month. The stated aim, in Bessent's words, is a protocol to keep frontier models out of the hands of non-state actors. Experts split sharply on what that means for export controls. Chris McGuire, a former Biden chip-controls official, has argued the dialogue requires tightening controls so Washington negotiates from strength. Sebastian Mallaby has argued the opposite, that export controls are now an obstacle to any meaningful pact. Either reading is harder to action while domestic AI safety regulation is on indefinite hold.
The pattern, taken together, is a quiet shift in what the United States is trying to optimise for. For most of the post-ChatGPT period, the official posture combined competitive pressure on China with steadily tightening domestic safety expectations. This week the second half of that posture moved. The Commerce Department's testing agreements with Google, Microsoft and xAI have been described as still in place, but the public announcement quietly disappeared from the department's website. State-level AI laws are being actively opposed by the White House. And the proposed federal review pipeline, the one specifically built around the cyber risk Anthropic itself has flagged, has been parked.
That leaves an awkward equilibrium. Chinese fabs are shipping at record volumes, US labs are producing models capable of finding tens of thousands of vulnerabilities a month, and the formal mechanism for either side to constrain the other now runs through a freshly announced bilateral dialogue with no agreed agenda. It is possible this all converges into a workable framework before the next class of frontier models ships. It is also possible the dialogue arrives too late to matter, and the competition keeps writing the rules in its own image. For now, the most accurate description of US AI policy is the one Trump gave reporters himself. He postponed it because he did not want to get in the way.