US BIS introduces interim rule to close off Chinese firms' remote access to AI compute via overseas data centers
On November 16, the US Department of Commerce introduced an interim regulatory rule requiring that when AI chips and accompanying servers are exported, re-exported, or transferred to data centers worldwide, a special license must be applied for in advance whenever there is a possibility of providing compute to enterprises in restricted countries such as China. The rule uses a threshold of more than 12 million total processing performance (TPP), which according to TechCrunch's estimate is roughly equivalent to the compute of 750 Nvidia H100 or H200 chips; overseas data centers that already provide compute meeting the threshold must submit supplementary applications within 90 days after the rule takes effect, and Malaysia, Singapore, Indonesia, Thailand, and Brazil have been newly added to the scope of advanced semiconductor export licensing.
Why read this
AI score 82
The new rule shifts the regulatory target from physical chips to the flow of compute services, and sets a 12 million TPP threshold and a 90-day supplementary application window, which can be used to assess the compliance costs of overseas compute channels.
Topics
Body · AI translation
On November 17, it was reported that the U.S. government is further tightening export controls on advanced artificial intelligence chips, attempting to close previous loopholes in the rules and prevent enterprises in restricted countries such as China from bypassing direct export bans by remotely accessing the computing power resources of advanced U.S. AI chips through data centers deployed in third countries.
The U.S. Department of Commerce officially introduced an interim regulatory rule on November 16 local time, explicitly requiring all relevant enterprises, when exporting, re-exporting, or transferring AI chips and supporting servers to data centers around the world, to apply in advance to regulatory authorities for a special license as long as the relevant facilities have the possibility of providing computing power services to enterprises in China or other restricted countries.
This means that the coverage of these controls has completely broken out of the previous boundaries and is no longer limited to whether chips and servers are actually physically shipped into restricted countries. Even the entire chain of behavior in which chips are deployed throughout in an overseas third country and subsequently export computing power to restricted customers through cloud computing services has also been brought under strict regulation.
The new rules target the D:5 country group explicitly designated in the U.S. export control system, which covers China, Russia, Iran, North Korea, and other countries subject to U.S. arms embargo restrictions. Prior to this, the United States had already explicitly required all enterprises to obtain export licenses in advance when selling advanced AI chips to entities in these countries or to companies actually controlled by enterprises from these countries.
The adjustment details introduced this time further extend the regulatory reach to the full range of scenarios in which overseas data centers provide remote computing power services externally. In the interim rule, the U.S. Department of Commerce set the applicable computing power threshold at more than 12 million total processing performance (TPP).
According to calculations by overseas tech media TechCrunch, this computing power scale is roughly equivalent to the maximum computing capacity that can be output by about 750 NVIDIA H100 or H200 chips combined, which is roughly enough to support the full-chain training needs of a medium-sized large model.
The binding boundary of the new rules is quite broad. Even if Chinese enterprises do not directly sign contracts with overseas data centers but instead indirectly obtain the corresponding computing power through intermediaries, agents, or other multi-layered third-party entities, they will likewise fall within the regulatory scope of the new rules.
To implement these regulatory requirements, all chip exporters and data center operators worldwide must significantly upgrade customer due diligence processes, verifying customers' equity structures, actual control relationships, and all business activity flows one by one, and strictly prevent enterprises in restricted countries from bypassing control restrictions through complex multi-layered corporate structures or intermediary agencies.
If an overseas data center, before the new rules formally take effect, has already been providing computing power services meeting the threshold to enterprises in the relevant countries, it must also apply retroactively for the corresponding special license within 90 days after the rules take effect. Entities that fail to complete the application on time will subsequently be directly placed on the U.S. export restriction list.
The new rules additionally bring five countries—Malaysia, Singapore, Indonesia, Thailand, and Brazil—into the scope of application of advanced semiconductor export licensing requirements.
The reason given by the U.S. Department of Commerce is that the risk of abnormal transshipment of advanced chips in these countries is far higher than the global average.
However, the rules also include exemption provisions. If the chips are ultimately delivered to enterprises headquartered in the United States, or to enterprises whose ultimate parent company is registered in the United States, they may be exempt from this newly added licensing requirement, provided that the periodic reporting conditions required by regulators are met.
These new rules directly extend the boundary of U.S. AI chip controls from the physical flow of chips to the flow of computing power services, equivalent to using long-arm jurisdiction to bring most of the world's high-grade AI computing power resources under the U.S. regulatory list, and nearly completely blocking all overseas high-end computing power channels that Chinese enterprises can reach.
AI Trends