1. The U.S. is tightening restrictions on China's access to advanced AI chips by closing a loophole that lets Chinese firms obtain computing power through overseas data centers, shifting controls from the physical movement of chips to remote access to their computing capacity. [para. 1][para. 4]
2. The Commerce Department's Bureau of Industry and Security issued an interim rule on Monday requiring licenses for shipments of advanced AI chips and servers to data centers anywhere in the world if they provide large amounts of computing power to companies headquartered in China or other restricted jurisdictions. The rule also imposes new licensing requirements specifically on exports to Singapore, Malaysia, Indonesia, Thailand, and Brazil, which the U.S. views as having a higher-than-normal risk of chip diversion. [para. 2][para. 3]
3. Under the new framework, exporters must seek U.S. approval if they know covered chips or servers are destined for a data center providing more than 12 million in aggregate total processing performance (TPP) to China-linked companies. This threshold applies to an individual facility or a group of linked data centers, regardless of the data center's location, and covers computing provided directly or indirectly to companies headquartered in China. [para. 5][para. 8]
4. Based on Nvidia product specifications and industry estimates, the 12 million TPP threshold is roughly equivalent to about 750 Nvidia H200 chips. While this is small compared with the tens of thousands of chips typically needed to train the most advanced AI models, it could still support substantial inference workloads that require less concentrated computing capacity. [para. 6][para. 7]
5. Data-center operators face new obligations after chips are delivered. If an overseas facility begins providing computing capacity above the threshold to covered Chinese entities, the owner, operator, or another involved party must obtain a license. Existing data centers already providing such services have 90 days from the rule's effective date to seek approval if they wish to continue, or they risk future U.S. export restrictions. Commercial recommendation systems, such as those used by online retailers, are exempt. [para. 9][para. 10][para. 11]
6. The separate licensing requirement for Singapore, Malaysia, Indonesia, Thailand, and Brazil could have a particularly large impact on Southeast Asia, which has become a major destination for AI data-center investment. Advanced chips shipped to these five countries will generally require U.S. approval, though companies headquartered in the U.S. or owned by U.S. parents can qualify for an exemption if they meet reporting requirements. Washington cited elevated risks of diversion to restricted users in these markets. [para. 12][para. 13][para. 14][para. 15]
7. Southeast Asia has grown rapidly as a computing hub as Chinese tech companies expand overseas and seek AI infrastructure. Singapore had more than 1.4 gigawatts (GW) of computing-infrastructure capacity in 2025, while Malaysia had about 1.06 GW, Indonesia 900 megawatts, and Thailand 620 megawatts. Regional capacity is expected to exceed 10 GW by 2030, up from more than 4.4 GW in 2025. [para. 16][para. 17]
8. Chinese technology companies including ByteDance, Alibaba, and Tencent have leased data-center capacity in Southeast Asia. Of roughly 1.5 GW of data-center projects approved in Thailand, as much as 60% to 70% are backed by Chinese investors or primarily serve major Chinese tech companies. In response, Malaysia plans legislation requiring data centers to disclose information such as operators, cloud customers, locations, electricity consumption, and GPU imports. Singapore said it is studying the new U.S. rules and engaging with affected companies and U.S. authorities. [para. 18][para. 19][para. 20][para. 21]
9. The rule also tightens requirements for semiconductor foundries and chip-packaging companies. Certain advanced chips designed by companies outside the U.S. and designated allied jurisdictions will be presumed to fall under U.S. export controls and to be intended for data-center use unless manufacturers or packaging companies can demonstrate otherwise. This builds on earlier restrictions aimed at preventing Chinese chip designers from using overseas foundries for advanced processors. [para. 22][para. 23][para. 24]
10. The Commerce Department formally rescinded the Biden administration's AI Diffusion Rule, a sweeping global framework unveiled in January 2025 that sought to divide countries into different tiers for access to advanced computing power. That rule never took effect; the Trump administration announced plans to abandon it in May 2025, arguing for a more targeted approach like the current interim rule. [para. 25][para. 26]
AI generated, for reference only