The awkward thing about an export-control loophole is that it is rarely a hole in the law shaped like a cartoon tunnel. It is usually a place where paperwork, geography, corporate structure, and enforcement patience fail to line up. A chip does not have to travel to Shenzhen to end up in a Chinese AI system. It can go to a subsidiary, a cloud tenant, a reseller, a lab in a friendlier jurisdiction, and still do the same work.

That is why the Commerce Department's weekend guidance matters. CNBC reported that the department moved on 31 May to enforce licence requirements for the most advanced AI chips when the entity is headquartered in China, even if the buyer is physically outside China. The article names Nvidia's Rubin and Blackwell processors and AMD's MI350x as the kind of hardware at stake. The important phrase is not "outside China". It is "headquartered in China", because Washington is trying to make ownership and control matter as much as delivery address.

This is also an admission that the old map was too simple. Last year the Commerce Department said it would not enforce the AI Diffusion rule, which had tried to govern global access to AI chips. According to CNBC, that left a year-long opening in which subsidiaries of Chinese AI firms in places such as Malaysia could potentially buy chips that the broader policy was supposed to keep away from China. Chris McGuire, a former State Department official, told CNBC that overseas subsidiaries could buy Blackwell chips without a licence under the previous posture.

I wrote recently about how Huawei silicon was setting a new floor for Chinese AI pricing. This is the same story from the other side of the table. If Chinese labs can get enough restricted Nvidia or AMD compute through offshore entities, the domestic substitution pressure weakens. If they cannot, Huawei, Cambricon, Moore Threads, and the rest of the local stack become less like patriotic alternatives and more like necessity.

The South China Morning Post's broader account of China's chip redesign makes that pressure visible. The fight is no longer just to build a single Nvidia replacement, but to create a domestic ecosystem that can support leading models reliably. That means GPUs where possible, ASICs where useful, and a lot of unglamorous software work around compilers, interconnects, memory, and serving. Export controls don't create that ecosystem by themselves. They do, however, make the unpleasant engineering trade-offs feel less optional.

Nvidia had a different kind of news cycle running at the same time. The BBC reported from Computex that Jensen Huang announced RTX Spark, a PC chip for personal AI agents, with Windows machines due in the autumn from Asus, Dell, HP, Lenovo, Microsoft Surface, and MSI. Huang called the reinvention of the PC as big as the smartphone shift. That line is keynote-ready, maybe too neat, but the contrast is useful. On one stage Nvidia is selling the future as a domestic appliance: a personal machine that moves from tool to teammate. In Washington, the same company's highest-end silicon is being treated as a strategic material whose customer list has to be read through corporate control.

Those two versions of AI hardware now live together. Consumer abundance at the front of the shop, geopolitical rationing behind the counter. The border between them keeps moving, because every restriction creates a new routing problem and every routing problem creates a new rule. The chip is still just a chip, until someone asks who really owns the company buying it.

Sources: