Caixin Explains: How New U.S. AI Export Controls Will Work

The United States has implemented comprehensive global export restrictions aimed at advanced artificial intelligence (AI) computational resources. These new rules mandate licenses for exporting high-performance GPUs and AI model weights, with the specifics varying per country to address national security concerns. Nations like China are under particularly strict limitations. [para. 1]

The directives, provided in a 168-page report by the Bureau of Industry and Security (BIS), took immediate effect with a grace period of 120 days. The goal is to prevent the spread of advanced AI technologies potentially threatening U.S. national security or foreign policy. Advanced AI models exported from the U.S. raise the risk of unauthorized use, particularly by China's designation as a "country of concern." [para. 2][para. 3]

The regulations establish thresholds based on Total Processing Performance (TPP) and performance density for high-performance GPUs, classifying them as advanced AI computational power. GPUs with a TPP of at least 4,800 or a performance density of 5.92 and above are subject to restrictions. Furthermore, restrictions apply to computers and systems incorporating these GPUs. [para. 4]

The framework classifies countries into three categories. First, trusted allies and partners like Japan, South Korea, and the European Union have exceptions under an AI Authorization framework, provided they meet certain certification standards. Second, countries of concern such as China and Russia face complete export prohibitions. Third, intermediate countries like Singapore and Vietnam face quotas intended to restrict their ability to develop advanced AI models. For the years 2025 to 2027, these countries can access up to 7.9 billion TPP, equivalent to around 50,000 advanced GPUs. Exporters must adhere to stringent guidelines like verifying end users and avoiding unauthorized transfers. [para. 5][para. 6][para. 7][para. 8]

The new rules also extend to AI model weights trained with substantial computational power, requiring licenses for their export or transfer, including foreign-developed AI models using U.S. technologies like GPUs. [para. 9]

These measures have implications for cloud service providers and data centers, introducing two licensing frameworks: universal verified end users and national verified end users. Companies with universal verification cannot deploy more than 25% of their AI computational capacity outside the U.S. or its allies. For national verification, intermediate countries have a cap set at 50.64 billion TPP for 2025-2027. Cloud providers in allied nations are restricted from operating data centers in regions like China and must limit AI computational power to prevent unauthorized AI model training. [para. 10][para. 11][para. 12]

Broader implications include extending extraterritorial jurisdiction, affecting foreign companies using U.S. technologies for AI developments. Open-source AI models remain unaffected due to lower computational thresholds, but BIS cautioned this may change as technology evolves. [para. 13][para. 14]

The international response includes condemnation from the Chinese Ministry of Commerce, who criticized the move as U.S. overreach disrupting global trade and promised countermeasures to protect China's interests. These regulations are projected to impact global technology supply chains, especially in AI, semiconductors, and quantum computing. [para. 15][para. 16]

Nvidia and other tech industry stakeholders have voiced significant opposition. They argue the rules are hasty, potentially harming U.S. technological leadership and global competitiveness. The measures could cause lasting economic damage and lead to the U.S. losing strategic markets to its competitors. [para. 17][para. 18][para. 19]