qinbafrank|11月 16, 2026 23:52
The Bureau of Industry and Security (BIS) of the U.S. Department of Commerce has issued new regulations on remote computing power access, closing the window for arbitrage in remote computing power. Here are the key points of the new regulations: 1) Remote access control: For data centers in any region of the world where Chinese companies and subsidiaries provide more than 12 million TPPs of computing power (approximately equivalent to 750 H100/H200 servers, or about 300 B200 servers, depending on the computing method), the export, re-export, and domestic transfer of related advanced computing chips and servers require a license. Licenses are approved on a case-by-case basis.
2) Data centers already providing large-scale computing power before the regulations were issued must obtain a license within 90 days of the regulations taking effect if they wish to continue providing services; otherwise, they will face subsequent export restrictions.
3) Limited exemptions: Systems used solely for commercial recommendation systems (e-commerce, social media rankings, etc., without generating text/images/audio/video/software) are exempt but must be reported; generative AI systems, even those with recommendation functions, are not exempt.
4) Five New Countries Required for Licenses: Exports/re-exports/domestic transfers of advanced computing chips to Indonesia, Malaysia, Singapore, Thailand, and Brazil all require licenses (the Secretary of Commerce may add other countries). Companies headquartered in the US are exempt but must submit monthly reports. The effectiveness of export controls and enforcement cooperation in the destination countries will be considered during the review process.
A few days ago, a long article titled "Who's 'Suppressing Yellow'?" discussed how the increasing restrictions on "overseas cloud computing power" are becoming a reality.
In the past two years, leading Chinese companies (Alibaba, ByteDance/Douyin, Tencent, etc.) have extensively leased H100/H200/Blackwell computing power from data centers in Southeast Asia, including Malaysia, Singapore, Indonesia, and Thailand, for training. Previously, leading model companies used Southeast Asian data centers as "quasi-domestic computing power." This new rule systematically incorporates this major workaround into the licensing process. For models requiring training at the 100,000-GPU level, this represents a simultaneous deterioration in cost, availability, and determinism.
The substantial tightening of remote advanced computing power channels will increase the cost and difficulty of training for Chinese model manufacturers, slowing down their catch-up with the US's leading technologies.
It's fair to say the US is attempting to block almost all crucial external pathways for China's AI catch-up: stricter chip controls, reduced remote computing power channels, and increasingly stringent dedistillation measures for cutting-edge models.
Could this be considered a "difficult time" for Chinese models?
This post is sponsored by @bitget_zh, "Bitget Buy US Stocks: Instant Entry, Smooth Trading"
Share To
Timeline
HotFlash
APP
X
Telegram
CopyLink