The New AI Chip Export Policy to China: Strategically Incoherent and Unenforceable
On January 13, the Department of Commerce published a new regulation permitting the sale of advanced AI chips to China, codifying the major policy change that President Trump announced on December 8. The regulation loosens restrictions on the export of Nvidia H200 chips, as well as the AMD MI325X and equivalent chips from other companies, which had previously been banned for export to China.
The regulation acknowledges that exporting advanced AI chips to China poses serious national security risks, while simultaneously creating a pathway to permit their sale. The result is a framework that is strategically incoherent. If implemented faithfully, it likely would block most or all exports to China, but if implemented loosely, it would fail to address any of the concerns that motivated export controls in the first place.
Key Elements of the Regulation
Revised Thresholds: The regulation allows the export of AI chips that are 13 times more powerful than what was previously permitted, enabling shipments of AI chips with a total processing performance (TPP) of less than 21,000 or a total DRAM bandwidth less than 6,500 GB/s.
Chip Volume Cap: For each product, the number of chips exported to China cannot exceed 50% of the number of chips shipped to U.S. customers for end-use in the United States.
No Exports for Data Centers Outside of China: The regulation maintains a “presumption of denial” policy on exports of AI chips to China-owned data centers located outside of China.
U.S. Supply Certification: The exporter must certify that exports to China will not cause “any delay” in fulfilling existing or new orders from U.S. customers of any AI chips, and that global foundry capacity that would otherwise be used to produce AI chips for U.S. end-users will not be diverted to produce products for China.
End-Use Certification: The exporter must certify that the chips will not be used for military, intelligence, or weapons of mass destruction end-uses. End-users in China must implement “robust” know your customer (KYC) practices to prevent the chips from being used by restricted Chinese end-users, that model weights will not be transferred, and that remote access to AI algorithms won’t be provided to restricted parties.
Assessment of Impact of Regulations
A Large Volume Cap: While the regulation imposes some limits on how many AI chips China can receive, the topline cap remains substantial. Analysts estimate that Nvidia sold around 2 million H200 chips to the United States, thus capping H200 sales to China at around 1 million H200s, and potentially another 1 million H100s. This volume is still significant enough to create one of the largest AI data centers globally.
Restrictions on Data Centers: The regulation ensures Chinese companies cannot use U.S. chips in data centers outside of China, which prevents competition with U.S. cloud providers. However, it guarantees that all U.S. chips will be utilized inside China for AI training and sensitive use cases. Consequently, Chinese firms may still establish data centers overseas using domestic AI chips.
Certifications Will Pose Challenges: Some certifications required in this rule are difficult to verify. Exporters must certify that fulfilling orders for Chinese customers won’t negatively impact the supply of AI chips for U.S. customers. Given global shortages of critical inputs, this may prove challenging. Moreover, certifying that Chinese customers conduct “robust” KYC practices raises concerns, as many potential purchasers have ties to China’s military and security services.
Sets a Dangerous Precedent
This regulation establishes a precedent that could be hazardous if extended to more advanced chips. Allowing the sale of 1 million H200s to China increases its domestic AI compute capabilities significantly. If this logic were applied to future chips, it could lead to massive increases in China’s AI capabilities, which would not align with U.S. strategic interests.
Conclusion
The rule inadvertently reveals that there is no version of an AI chip export policy to China that is simultaneously permissive, implementable, enforceable, and protective of U.S. national security. While the restrictions included are preferable to unconditional exports, they still create a false sense of security while enabling significant AI chip exports to China without enforceable safeguards against misuse.
The strategic justification for selling advanced AI chips to China has always been weak, balancing modest commercial gains for a few U.S. chip companies against substantial risks of bolstering Chinese military AI capabilities and eroding U.S. technological leadership. This regulation highlights the impossibility of threading such a needle, underscoring that a more effective policy would simply deny the export of all AI chips to China.
