美國收緊中國企業透過海外資料中心租用 AI GPU 算力的管制
U.S. Cracks Down on China Renting AI GPUs Through Overseas Data Centers
美國商務部工業與安全局(BIS)發布臨時最終規則,向為總部位於澳門或 D:5 國家/地區的企業提供逾 12 million aggregate TPP 算力的全球資料中心運送先進加速器,須取得許可。
規則把海外資料中心向受限地區企業提供算力納入出口許可,並列出 TPP 門檻、間接存取和既有客戶的申請期限,呈現管制如何延伸至租用算力。
Yesterday, the U.S. Commerce Department's Bureau of Industry and Security (BIS) released an interim final rule to address remote access to advanced AI chips and servers. This rule significantly addresses a long-standing gap in U.S. export controls, where Chinese companies that are barred from accessing NVIDIA's latest GPUs have been renting overseas computing power instead of purchasing the chips outright. Notably, BIS states that the rule was written by and at the direction of the Department of the Treasury and the National Security Council.
Under the new rule, a license is required to ship any advanced accelerators to any data center worldwide that provides more than 12 million "aggregate TPP" of computing power to companies headquartered in Macau or Country Group D:5, which includes China. TPP, or Total Processing Performance, is BIS's metric for chip computing power, calculated as a chip's peak dense throughput multiplied by the bit length of the operation. For example, an H100 rates about 15,800 TPP. "Aggregate TPP" sums up every accelerator in the facility, so the cap equates to roughly 750 H100s or around 330 B200s. The limit also applies to "linked data centers" and indirect access, such as a cloud reseller training models for a Chinese firm. Data centers already serving such customers have 90 days to obtain a license. Only "commercial recommender systems" are exempt, and anything that generates text, images, video, or code does not qualify.
The rule also requires a license for advanced AI chips shipped to Indonesia, Malaysia, Singapore, Thailand, and Brazil, which the Commerce Department says carry a higher-than-normal risk of diversion. Shipments to U.S.-headquartered companies are exempt if they meet new reporting requirements. Foundries and packaging houses also lose some of the methods to clear advanced-node chips for designers in restricted countries. Finally, the rule formally rescinds the Biden-era AI Diffusion Rule, eliminating its country tiers, the Data Center VEU program, and model-weight controls. Reflecting the current trends and the times under the U.S. administration led by Donald Trump, the document uses "Super Intelligence" instead of "AI," citing Executive Order 14434.
來源:TechPowerUp · techpowerup.com
