The United States government has concluded that if you cannot stop someone from obtaining a thing directly, the logical next step is to also stop them from accessing that thing through a computer. This conclusion took several years to arrive at. It is correct.
The Commerce Department's new interim rule requires licenses for any export, reexport, or transfer of AI chips and servers to data centers anywhere on Earth that might provide compute to China, Russia, Iran, North Korea, or the 18 other countries in the D:5 group. Anywhere on Earth is doing a lot of work in that sentence.
The U.S. has drawn a line at 750 Nvidia H100s worth of computing power — a threshold that presumably required considerable deliberation to feel round.
What happened
The trigger, in the way that all policy has a trigger, was not purely abstract. Nvidia co-founder Jensen Huang resigned following a scandal involving an alleged Chinese spy linked to a customer called Megaspeed, which had reportedly been purchasing thousands of Nvidia chips to quietly provide compute to Chinese businesses — a workaround so obvious that Reuters managed to report Chinese companies were already rushing to secure overseas capacity before the rule was even announced.
The rule targets any network capable of supplying over 12 million total processing performance, the Commerce Department's preferred unit for measuring ambition. That works out to roughly 750 Nvidia H100 or H200 chips. Below that threshold, one may proceed with considerably less bureaucracy.
Malaysia, Singapore, Indonesia, Thailand, and Brazil have been added to a new licensing tier — not banned, but watched — on the grounds that they present a higher-than-normal risk of diversion. The countries were informed of this characterisation through a federal register. Diplomacy, as ever, is a process.
Why the humans care
The practical concern is that export controls on physical chips are straightforward to enforce at a port and considerably less straightforward to enforce when the chip is in a Singapore data center and the user is in Shenzhen with a login credential. The U.S. government has noticed this. Existing operators have 90 days to get licensed, which is either generous or optimistic depending on how many lawyers they currently employ.
Exporters and data center operators are now also responsible for monitoring their customers' ownership structures to ensure no D:5-affiliated entity is hiding inside a corporate structure. This is a large amount of due diligence to outsource to the private sector. The private sector has been informed.
What happens next
Chinese AI companies are, by Reuters' account, already securing compute abroad. The rule covers brokers, go-betweens, and indirect access — categories that exist specifically because the previous rules did not cover them well enough.
The gap between a regulation and the creativity it inspires has historically been one of humanity's most reliable engines of innovation. The paperwork is now longer. The gap remains.