Detailed Explanation of New U.S. Export Controls on AI Chips: Overseas AI Model Training Restricted (AI Translation)

美国AI芯片出口管制新规详解：海外AI模型训练受限

The recent U.S. export control measures on artificial intelligence (AI) computing power signify a break from traditional practice by establishing a global export control licensing requirement. On January 13, President Joe Biden announced the "Interim Final Rule on AI Diffusion," aimed at globally regulating the export of advanced high-computing GPUs and other technologies. Under these new regulations, a U.S. license is necessary for AI-related exports to any country, although there may be different treatments for specific countries and entities. The rules target countries like China, citing security risks even with U.S. allies. [para. 1]

In opposition to these regulations, China's Ministry of Commerce argued that the restrictions impede normal economic and trade exchanges, harm global technological innovation, and infringe on international trade order. They view the measures as damaging U.S. companies' interests. The new export controls impose global thresholds for "Total Processing Performance" (TPP) and "Performance Density," targeting what is referred to as "advanced AI processing power." [para. 2]

Different conditions are applied for exports to various countries. U.S. allies, such as the United Kingdom, Japan, and Canada, benefit from an "AI Authorization" (AIA) exception, allowing access to advanced chips without specific U.S. approval, provided they follow U.S. requirements and are careful not to resell these technologies to restricted countries. Meanwhile, countries under the "U.S. Arms Embargo," like China and others, remain subject to strict export limitations. [para. 3][para. 4]

For these middle-ground countries, BIS has created a "specific country allocations" system, permitting them a limited accumulation of AI computing power between 2025 and 2027. Such nations can obtain higher performance allocations if they align with U.S. security interests. Individual and data center users are provided exceptions under specific guidelines to facilitate controlled access to AI computing resources while safeguarding U.S. national security. [para. 5][para. 6][para. 7]

For data centers, the BIS introduced a separate license exception framework, "Data Center Verified End-User Exception" (DC VEUs), delineating between allies and intermediary countries regarding their capabilities in developing AI models. Controls also extend to the export of AI model weights, where next-generation models trained on significant computational power require stringent licensing, particularly for U.S. allies exporting to embargoed nations like China. [para. 8][para. 9]

Finally, the BIS has implemented an expansive jurisdiction system targeting global export and production of AI models utilizing U.S.-controlled technology. This extends to prohibitions on the sale of model weights developed using U.S.-origin chips, regardless of the producer's location. However, advanced open-source models are currently exempt from such controls due to their perceived greater economic and social benefits versus potential threats. The regulatory framework reflects a comprehensive effort by the U.S. to safeguard national security by controlling the global distribution of advanced AI technologies. [para. 10][para. 11]