In its final days, the Biden Administration has rushed through new rules designed to prevent AI developments that could threaten U.S. national security. However, the new rules have faced strong criticism from the chip and cloud industries, who argue that the revised rules are unnecessarily strict and will eventually favor Chinese chip competitors.
The Department of Commerce’s Bureau of Industry and Security (BIS) announced tighter controls on AI chips, and caps on the amount of compute that can be shipped on a per-company, per-country basis. The new rules, called the regulatory framework for the responsible diffusion of advanced artificial intelligence technology, also limit access to model weights for closed AI models over a certain size that are trained using U.S. technology.
The new default position is that AI chips roughly comparable with Nvidia’s previous-gen A100, or anything more powerful than 600 INT8 TOPS, will require a license to ship to any country. There are 18 exceptions, including most of northern Europe plus Taiwan, Canada, Australia, New Zealand, Japan and South Korea. Exports of these chips to U.S. arms-embargoed countries, which includes China and Russia, remains banned. Licenses will be considered on the basis of the compute power of the transaction, and in some cases, security measures agreed to by the recipient, according to the paperwork of the new rules.
These rules would apply to Nvidia previous-gen A100, current-gen H100 and next-gen B200 GPUs, as well as training-capable chips from companies like AMD, Cerebras, SambaNova and others. While the rules are designed to limit training large-scale models, current-gen inference-only chips with processing power above the 600 INT8 TOPS limit also fall under the restrictions. This would include those from companies like Groq, Untether and Tenstorrent.
Chips between 200-600 INT8 TOPS will require licenses only for certain countries (including most middle east countries) but are OK to ship elsewhere (except U.S. arms-embargo countries). Chips below 200 INT8 TOPS are not within the scope of the new rules, unless they fall foul of additional compute density limits.
Exceptions will apply for the supply chain (i.e. the development, production and storage of chips in countries that would otherwise require an export license, like Singapore, Vietnam, Malaysia and Mexico). There is also an exception for certain low volume deployments, though the BIS must still be notified; this applies to per-company deployments up to the equivalent of 5300 Nvidia A100s, or about a tenth of the resources of the cluster Meta used to train Llama3-405B.
The U.S. is also imposing new rules on who gets access to AI model weights for closed models that required more than 1026 computing operations to train. This scale is comparable with Meta’s Llama3-405B model, which required 3.8×1025 FLOPs to train (though these rules do not apply to Llama3 because its weights are open). The new rules also apply to models trained abroad using U.S. chips. There are no restrictions on who can access models of any size whose weights are open.
“This policy will help build a trusted technology ecosystem around the world and allow us to protect against the national security risks associated with AI, while ensuring controls do not stifle innovation or U.S. technological leadership,” said U.S. Secretary of Commerce Gina Raimondo in prepared remarks. “Managing these very real national security risks requires taking into account the evolution of AI technology, the capabilities of our adversaries, and the desire of our allies to share in the benefits of this technology. We’ve done that with this rule, and it will help safeguard the most advanced AI technology and help ensure it stays out of the hands of our foreign adversaries, while we continue to broadly share the benefits with partner countries.”
The BIS had introduced the first rules for the export of AI chips in 2022, which also apply to manufacturing equipment used to make the most advanced chips. The previous situation was seen as workable by companies including Cerebras and SambaNova, who previously told EE Times that while licenses were required to export to nascent but fast-growing markets in the middle east, these licenses had so far been granted by the BIS.
Market leader Nvidia had gone as far as developing cut-down versions of its chips, which fell just below the export limit, specifically for the Chinese market (the H20, for example).
“The first Trump Administration laid the foundation for America’s current strength and success in AI, fostering an environment where U.S. industry could compete and win on merit without compromising national security,” Nvidia’s VP of government affairs, Ned Finkle, wrote in a blog on the company’s website.
“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,” he added.
The blog also accused the Biden Administration of drafting rules in secret and without proper legislative review, and described the new rules as “sweeping overreach [that] would impose bureaucratic control over how America’s leading semiconductors, computers, systems and even software are designed and marketed globally.”
A statement from the Semiconductor Industry Association (SIA) said the group was “deeply concerned by the unprecedented scope and complexity of this potential regulation, which was developed without industry input and could significantly undercut U.S. leadership and competitiveness in semiconductor technology and advanced AI systems.”
Cloud service providers
CSPs also face additional rules for AI compute cluster deployments.
CSPs can apply for universal validated end user status (UVEU), intended for U.S. hyperscalers who will be allowed to deploy whatever chips they want, at whatever scale they want, provided it is not in a U.S. arms-embargoed country. Some companies will be able to apply for national validated end user (NVEU) status, which will get them permission for deployments capped by company and by country.
NVEU caps are cumulative—that is, they apply to whatever an NVEU company has already deployed in a particular country. They increase by calendar quarter, in a move designed to limit the rate at which these deployments can grow. For example, the cap for Q1 2025 is equivalent to around 127,000 Nvidia A100s (or 40,000 Nvidia H100s) per company, per country.
There is no regulation on end-user applications or API usage.
In a blog posted on the company’s website, Oracle executive VP Ken Glueck was just as scathing as his Nvidia counterpart.
“What Congress accomplished by passing the CHIPS Act (a mere $280 billion), the Biden Administration takes away with the Diffusion Framework, because in one [Interim Final Rule] it has managed to shrink the global chip market for U.S. firms by 80% and hand it to the Chinese,” he wrote.
Glueck argues that the new rules retroactively regulate global cloud GPU deployments, shrink the global market for U.S. cloud and chip suppliers, establish volume restrictions, and tell 20 countries they can be trusted only if they agree to new requirements—and likely pushes the rest of the world to Chinese technology.
The UVEU program will have strings attached, Glueck said in the blog, since it is tethered to the U.S. government’s FedRAMP high data and physical security standards. Bringing domestic and overseas data centers up to this strict standard “would fundamentally change the economics of data center deployment around the world,” he wrote.
“For the first time, we are applying draconian new regulations to largely unregulated public, commercial cloud,” Glueck wrote. “We are stifling innovation and strangling emerging business models. Worse, without fully contemplating the rule’s effects, we are likely handing most of the global AI and GPU market to our Chinese competitors.”
National security
In seeking to restrict, or at the very least, gain visibility into at-scale deployments of the latest AI chips, the U.S. government may be drastically handicapping the market for U.S. semiconductors, industry players argue. The new regulations are strict enough to apply to 5-year-old technology while inadvertently (and unnecessarily) applying to current-gen inference-only chips. The changes in these rules are significant, effectively giving the U.S. government insight into—and control over—all significant scale training clusters worldwide.
While the new rules stop shipments of the most advanced chips to all except 18 countries in the world, examples of specific national security threats given in the paperwork only reference China by name. The fear is that AI chips sold to entities in other countries, particularly those in the middle east, could find their way into China. There is no doubt that these rules are intended specifically to curb China’s development of frontier AI models, as well as encouraging companies to deploy AI infrastructure in the U.S.
While Nvidia’s Chinese chip competitors are currently a couple of generations behind Nvidia in terms of performance (that is about 4 years in semiconductor years), bigger and better chips are surely under development despite similar restrictions on the export of U.S. semiconductor manufacturing equipment.
The new rules will certainly change the market dynamics for U.S. and Chinese chip companies, but whether they will have any effect on U.S. national security remains to be seen.
From EETimes