The US has introduced sweeping global export restrictions targeting advanced artificial intelligence (AI) computational power. These rules require licenses for all exports of high-performance graphics processing units (GPUs) and AI model weights, regardless of destination, with varying restrictions by country to address national security concerns. Countries such as China face particularly stringent limitations.
The measures outlined in a 168-page report released on Monday (Jan 6) by the Bureau of Industry and Security (BIS), come into effect immediately with a 120-day grace period.
They aim to curb the proliferation of advanced AI technologies that could pose risks to US national security or foreign policy interests. The BIS emphasised that exporting advanced AI models or technologies outside the US increases the potential for unauthorised use. “Countries of concern”, including China, are explicitly targeted due to the risk of misuse, even when dealing with US allies.
The rules introduce thresholds based on Total Processing Performance (TPP) and performance density for high-performance GPUs, with restricted items classified as advanced AI computational power. These include GPUs with a TPP of at least 4,800 or a TPP of at least 1,600 combined with a performance density of 5.92 or higher. Other restricted items include computers, electronic components and systems incorporating these GPUs.
How the rules work
The regulations establish a tiered licensing system putting countries into three groups.
Trusted allies and partners: includes Japan, South Korea, Australia, the European Union and Taiwan. Companies in these nations are granted exceptions under an AI Authorization framework allowing access to advanced GPUs without individual licences if they meet certification standards, such as verifying end-user compliance and avoiding unauthorised transfers by either reselling items or providing them via cloud services to entities prohibited by BIS.
BT in your inbox
Start and end each day with the latest news stories and analyses delivered straight to your inbox.
Countries of concern: include China, Russia, North Korea, and Iran and 24 US Arms Embargoed Countries. Exports to these countries are strictly prohibited, with Hong Kong and Macau subject to the same restrictions.
Intermediate countries: about 150 countries, including Singapore, Vietnam and Saudi Arabia face quotas and additional restrictions to limit their ability to train the most advanced AI models, allowing them to benefit from AI large models without posing a threat to US national security.
For 2025 to 2027, these countries can cumulatively access up to 7.9 billion TPP (equivalent to about 50,000 advanced GPUs), with the potential to double this if they align with US security objectives.
Exporters must follow stringent guidelines, including verifying end users and preventing unauthorised transfers.
What about AI model weights?
The regulations extend beyond hardware to include restrictions on the export of AI model weights trained with computational efforts exceeding 10 to the power of 26 operations. These “next-generation AI models” require licences for export or transfer. US approval is also needed for foreign-developed AI models that use US-origin technologies, such as GPUs or servers.
How will cloud service providers and data centres be affected?
The rules also impose limitations on data centres and cloud providers, structured under two licensing frameworks:
Universal validated end users: Companies with this status can deploy no more than 25 per cent of their AI computational capacity outside the US or its allies. No more than 7 per cent of their total AI capacity may be deployed in a single country outside this group.
Users headquartered in the US are prohibited from transferring more than 50 per cent of their total AI capacity outside the US.
National validated end users: Intermediate countries can apply for these licences but are capped at 50.64 billion TPP (about 320,000 advanced GPUs) per company for the period 2025 to 2027.
Cloud providers in allied nations are barred from deploying data centres in restricted regions, such as China, and must adhere to limitations on AI computational power to prevent unauthorised training of advanced AI models.
What are the broader implications?
The rules extend extraterritorial jurisdiction, applying to items developed outside the US that incorporate US-origin technologies. For instance, a French company using Nvidia-made chips must seek US approval to export AI model’s weights to China. The BIS-issued red flag warnings for cloud service providers to monitor risks when enabling foreign entities to train advanced models.
Open-source AI models are unaffected as their computational thresholds are lower than the restricted levels, although the BIS indicated this could change as the technology evolved. Currently the economic and social benefits of open-source models outweigh their potential threats.
What is the international response?
The Chinese Ministry of Commerce condemned the measures, accusing the US of overreach that disrupts global trade and stifles technological innovation. They promised to take necessary countermeasures to safeguard China’s interests.
The rules are expected to affect global technology supply chains, particularly in high-tech industries such as AI, semiconductors and quantum computing.
Nvidia expressed strong opposition to the measures, describing them as “hasty and misguided” and warning that they could undermine US technological leadership and global competitiveness.
The Semiconductor Industry Association condemned the Biden administration’s late move, saying the regulations could cause lasting economic harm and result in the US ceding strategic markets to competitors.
The Information Technology and Innovation Foundation, a US think tank, described the measures as “overdesigned, yet underinformed”, cautioning that they could undermine US technological leadership and competitiveness. CAIXIN GLOBAL
