There is an open gap in the current chip export controls: Chinese companies cannot buy advanced chips, but they can rent them. In the past, there were basically no restrictions on renting these chips for training in overseas data centers.
The Information reports that the U.S. Department of Commerce is finalizing a rule specifically to close this gap and will publish it as early as this month. Controls will extend from existing semiconductor export restrictions to seven countries including Malaysia, Singapore, and Thailand, with the goal of blocking remote access and chip smuggling at the same time. The Ministry of Commerce is currently in final negotiations with other government agencies.
This rule didn't just appear out of the blue. The White House has previously asked Congress to legislate for the Commerce Department to formulate similar regulations. Nvidia has lobbied against it on the grounds that U.S. AI infrastructure needs to win in global competition and that being too restrictive would push customers to others.
The timing is very delicate. Nvidia has just changed its CEO, and the former CEO will testify before Congress next month on the topic of export controls. In the same week, research also claimed that Blackwell’s hardware fingerprints were found in the weights of multiple Chinese open source models. When several things are put together, the resistance to the implementation of the rules is obviously reduced.
The actual impact on domestic AI teams depends on how the fine print is written. For example: Is it managed based on the location of the data center or based on the nationality of the lessee? How to calculate the long-term lease that has been signed? What level of customer review do cloud vendors need to do? Southeast Asia has undertaken a large amount of computing power leasing business in recent years. After the rules are introduced, the local computer room business and the training plan of the domestic team will have to be recalculated.