In the final days of the Biden Administration, the US has issued new sanctions that are intended to restrict the export of US processors and AI-based technologies to China. The incoming Trump administration should be on the same page, with regard to China, but the new rules will come into effect in 120 days, giving Trump’s team plenty of time to make alterations, should they want to.
The goal is to limit how much AI-focused compute power that nations can acquire, from the likes of Nvidia, AMD, and Intel. Cloud computing providers, like AWS and Microsoft, will be required to not deploy more than 50% of their AI computing power outside of the US, no more than 25% outside of twenty close allies (Tier One, described below), and no more than 7% in any non-Tier-One country.
US Commerce Secretary Gina Raimondo said that “as AI becomes more powerful, the risks to our national security become even more intense. The framework is designed to safeguard the most advanced AI technology and ensure that it stays out of the hands of our foreign adversaries but also enabling the broad diffusion and sharing of the benefits with partner countries.” Broadly, the US thinks it has an 18-month lead on China, in AI advancements, and the restrictions intend to preserve and hopefully extend that buffer.
Unsurprisingly, companies across the board are not happy about the decision. For MNOs that have been considering GPU-aaS projects, likely stemming from Nvidia’s recent marketing blitz, these new restrictions mean a lot more consideration needs to be given to which firms the MNOs commission. In general, global supply chains are going to get more complex.
Lawyers will be busy, but sifting through the hundreds of pages of primary documentation, never mind the secondary resources, will be a major headache. China will respond, in kind, and these geopolitical stratagems do not seem to be going anywhere anytime soon – so if you use a GPU anywhere in your stack (hello RAN engineers), buckle up.
Nvidia came out swinging, being the largest subject of these new rules, as most of the briefing discussions have made it clear that ‘advanced’ GPU-based accelerator cards are the prime subject of the regulations – with the H100 essentially used as the basis of the per-country calculations.
“Built on American technology, the adoption of AI around the world fuels growth and opportunity for industries at home and abroad. That global progress is now in jeopardy. The Biden Administration now seeks to restrict access to mainstream computing applications with its unprecedented and misguided ‘AI Diffusion’ rule, which threatens to derail innovation and economic growth worldwide,” said Ned Finkle, VP Government Affairs for Nvidia.
There are other vocal critics. The Semiconductor Industry Association (SIA) said that it was “deeply concerned by the unprecedented scope and complexity of this potential regulation, which was developed without industry input and could significantly undercut US leadership and competitiveness in semiconductor technology and advanced AI systems.”
Oracle EVP Ken Glueck said that the rules “do more to achieve extreme regulatory overreach than protect US interests and those of our partners and allies. It practically enshrines the law of intended consequences and will cost the US critical technology leadership.” Oracle is quite close to the Trump administration, and so these remarks are notable.
The restrictions (described as a “200+ page regulatory morass,” by Nvidia’s Finkle) are based on three tiers of countries. The first are allies, which face no new restrictions. These are: Australia, Belgium, Canada, Denmark, Finland, France, Germany, Ireland, Italy, Japan, the Netherlands, New Zealand, Norway, the Republic of Korea, Poland, Spain, Sweden, Switzerland, Taiwan, and the United Kingdom.
The second tier is those to which any exports are blocked, although most of these are already the subject of existing blocks – including December’s update to semiconductor manufacturing export controls. These countries are China, Iran, North Korea, and Russia.
The third tier is the largest, including many major economic and geopolitical partners, and the new rules will implement caps on the number of GPUs that can be exported to a country. The framework has been created in an attempt to prevent China purchasing these products via third-party countries.
For this tier, licenses will be required for orders larger than 1,700 units of GPU. National caps will be implemented, of up to 320,000 GPUs over the next two years, against which orders under that 1,700 unit threshold do not count.
The US will require that entities (companies) meet high security and trust standards (details are scant), but those that do and are headquartered in Tier One countries will be able to acquire ‘Universal Verified End User’ (UVEU) status. This lets them deploy up to 7% of compute in any country they like, bar the block-list. There is no limit on the number of GPUs.
The National Verified End User (NVEU) status is available for high-security entities that are headquartered in the Tier Three list. This will allow entities to purchase up to 320,000 GPUs over the next two years.
Any company that cannot secure either UVEU or NVEU status is essentially capped at 50,000 GPUs per country. However, the US is then using these restrictions as a bargaining tool, and will double the 50,000-unit caps for countries that align their security and clean energy efforts with the US. Precisely what that means is unclear.
The rules also restrict the export of the model weights, which are used in these AI-based models to set machine learning tolerances. These apply to closed-weight models, and new security requirements to protect these closed-weights are being introduced. Again, details are scant.
Nvidia’s Finkle added that “This sweeping overreach would impose bureaucratic control over how America’s leading semiconductors, computers, systems and even software are designed and marketed globally. And by attempting to rig market outcomes and stifle competition — the lifeblood of innovation — the Biden Administration’s new rule threatens to squander America’s hard-won technological advantage.”
“While cloaked in the guise of an “anti-China” measure, these rules would do nothing to enhance U.S. security. The new rules would control technology worldwide, including technology that is already widely available in mainstream gaming PCs and consumer hardware. Rather than mitigate any threat, the new Biden rules would only weaken America’s global competitiveness, undermining the innovation that has kept the U.S. ahead.”