Trump officials fear a loophole let Chinese firms buy Nvidia Blackwells

The gap, officials worry, let Chinese companies acquire banned chips through overseas subsidiaries. New guidance is meant to close it, after the chips may already have moved.

The export ban worked on paper and leaked in practice. Trump administration officials are worried that a gap in US rules allowed Chinese companies to legally buy servers fitted with Nvidia’s most advanced AI chips, the Blackwell line, by routing the purchases through subsidiaries outside China.

The controls aimed at Beijing turned out to have a door left open in Singapore and Malaysia, even as Washington has been moving to cut off China’s access to chipmaking equipment through measures like the proposed MATCH Act.

The mechanism was a question of where a company is headquartered versus where it shops. Chinese firms such as Alibaba could, in most countries outside China itself, buy systems carrying banned high-end Nvidia GPUs through overseas subsidiaries, because enforcement had not consistently followed the chips to the parent company behind the buyer.

The result, by the concern officials are now voicing, was that Chinese companies could acquire Blackwell chips and have AI chips fabricated at TSMC, legally and without a licence.

The timing of the gap traces to a rule change. The loophole emerged after the administration’s May 2025 revisions to export rules, even though the underlying requirement to license shipments to China-headquartered entities had technically been in force since November 2023.

The problem, in other words, was less the absence of a rule than the uneven enforcement of one already on the books.

The fix has now been issued. The US put out guidance affirming that its restrictions extend to subsidiaries of Chinese companies located outside China, with the Commerce Department clarifying that licensing requirements apply to any business whose headquarters or parent sits in China.

The clarification closes the door the worry was about, but it arrives after a window in which, officials fear, the chips may already have moved.

The Southeast Asian routing was not incidental. Singapore and Malaysia have become significant data-centre hubs, which made them natural places for subsidiaries to take delivery of high-end hardware without the purchase reading, on its face, as a shipment to China.

Reporting around the crackdown has noted that the enforcement push is now reaching China-linked data centres in the region, the downstream effect of a control regime trying to follow chips past the point of first sale, the same boundary tested by deals like the Qualcomm-ByteDance ASIC arrangement that works around export controls by design.

Congress had been pushing on exactly this. On 2 June, Senators Elizabeth Warren and Andy Kim accused the administration of allowing America’s most advanced AI chips to reach Chinese companies through stale export rules, putting a partisan edge on what was already an enforcement embarrassment.

The senators’ framing, that the rules existed but were not applied, matched the technical account of how the gap opened.

The pressure is already reshaping China’s own chip strategy, pushing its developers away from banned GPUs and towards custom ASICs designed at home, the kind of structural workaround that a single closed loophole does not undo. The episode is a reminder that export controls are only as strong as the corporate-structure tracing behind them.

A ban keyed to headquarters is straightforward to write and harder to enforce when buyers operate through a web of overseas entities. The new guidance restates the intent; whether it can claw back chips that have already shipped, or only stop the next ones, is the question the clarification cannot answer on its own.