A regulation proposed by the US government will require US citizens or permanent residents to notify the government about transactions with countries of concern related to building supercomputers that exceed 100 petaflops of LINPACK performance. Supercomputing is just one aspect of the broad regulation that covers semiconductors, AI, and other computing technologies considered essential to US national security interests.

The regulation, called “Provisions Pertaining to US Investments in Certain National Security Technologies and Products in Countries of Concern,” was proposed last month.

The law will also ban US citizens or permanent residents from making supercomputers exceeding 100 petaflops for China and require them to notify the US government about transactions related to the technology with China.

This proposed law could affect many executives of chip companies involved in building supercomputers and AI chips. It is also the US’s latest attempt to tighten control over high-technology exports to countries of concern, which for now is primarily China.

Already Some Intrigue

Chip companies with international partnerships could be accountable under this law. For example, AI chip maker Cerebras has an extensive partnership with UAE-based G42, which is building data centers in the US and internationally with Cerebras’ chips.

However, G42’s partnerships with Chinese companies attracted the interest of the US government, according to a report by the New York Times in November. The report also stated that Nvidia has a partnership with G42.

The New York Times report cited Cerebras’ CEO Andrew Feldman as being caught by surprise at the Biden administration’s concerns and that no one from the US government had contacted him about the issue.

G42 stated in August that it is working to meet US export control regulations, according to a report by Reuters.

Intel CEO Pat Gelsinger has publicly stated that the shifting government policy on export controls is disrupting business. He also said the government needed to be clear about its policy stance.

What Is the Proposed Regulation?

The regulation has a separate component specifically for supercomputing. It prohibits US persons from conducting specific transactions or requires them to report specific transactions within a supercomputing context with countries of concern, which include China, Macau, and Hong Kong.

The supercomputing context defines the prohibited activities to “include developing, installing, selling, or producing any supercomputer enabled by advanced integrated circuits that can provide a theoretical compute capacity of 100 or more double-precision (64-bit) petaflops or 200 or more single-precision (32-bit) petaflops of processing power within a 41,600 cubic foot or smaller envelope.”

To be sure, the regulation also covers semiconductors, AI, and other computing technologies. Supercomputing intertwines with many of those areas, which is a cause of concern for many policy hawks, venture capitals, and lobbying organizations tracking the regulation.

Also, US persons will be required to notify the government of certain supercomputing transactions if they are involved, hold a controlling interest, or are partners, managers, or investment advisors in a foreign entity. US citizens must also track and prevent foreign entities from conducting transactions within the supercomputing context with China.

Past Export Controls

The US government has already incorporated regulations that ban sales or exports of high-end AI chips such as GPUs. The regulations, passed on Oct. 7, 2022, also covered specific chip technologies in 100-petaflop supercomputers.

The initial regulations were meant to choke China’s emerging AI strategy by halting exports of AI chips, such as Nvidia’s H100 and A100, to the country. However, chip companies found creative ways to circumvent the regulations by constructing new chips that are below the bandwidth thresholds.

For example, the controls banned exports of chips with a bidirectional transfer rate of 600 GByte/s and “one or more processors… having a processor performance rate of 4,800 TOPS” (Trillions Operations per Second).

However, Nvidia got creative and developed the H800 and A800 chips that fell under those thresholds.

With that in mind, the government further tightened rules in October 2023 that banned chips such as Intel’s Gaudi2 and Nvidia’s A800 and H800 chips.

For its part, China is already building supercomputers with its homegrown chips. The country isn’t submitting systems for the Top500 list of the world’s fastest supercomputers.

Concerns

The proposed law is still in the comment period, and some in the computing industry haven’t taken it well, saying the restrictions were arbitrary. Others didn’t welcome the expanded government oversight, saying it would stifle innovation in AI.

Commenters also said the law is rooted in national security, with no heed paid to global market competitiveness for US technology companies.

Andreessen Horowitz’s a16z company, which has stockpiled GPUs and is renting them to companies, asked the government to remove computing power from the regulation. The venture capital firm has invested in hundreds of AI startups that require enormous computing power to sell AI products.

“The use of computational power to train AI models is not static and has changed rapidly over the last few years. Any restrictions imposed now may be irrelevant in a few years because the relevant computing power thresholds have been dramatically exceeded,” a16z said.

The most cutting-edge models are trained using lower levels of computing power; “therefore, basing the regulations on the quantity of computing power used to train the model has the potential to be significantly, and simultaneously, over and underinclusive,” a16z said.

Instead, a16z recommended that “Treasury move away from any restrictions based on computing power and instead focus the restrictions to be based on the type of covered foreign person in question,” a16z said.

The Semiconductor Industry Association (SIA) commented, “US semiconductor companies will be forced to cede their market share to foreign competitors.”

Foreign investors “may even be incentivized to forge partnerships within the targeted sectors in countries of concern in the absence of US investment activity,” SIA said.

That could “erode US leadership and strategic advantage in critical technology sectors like semiconductors, and other strategic industries that rely on semiconductors,” the SIA said.

The National Venture Capital Association (NVCA) said the proposed rules will impose a significant cost burden on US venture investments across the board.

The costs of meeting the regulation could add up to $100 million, much more than the US Department of Treasury’s estimate of $10 million annually, which is “woefully low,” NVCA said.

AI is a significant portion of many startups’ strategies, and the burden of compliance will add extra expenditure to conducting business, “but undoubtedly much higher in some – to each of the fifteen thousand or so venture investments made in the United States,” NVCA said.

Commenters said that the broad definition of US persons was vague and needed to be narrowed down.

In certain circumstances, the definition of people involved in building or using AI systems beyond a certain processing threshold is vague and could implicate innocent individuals.

“Currently, it is not clear whether the Department is suggesting that any AI system capable of image recognition or text mining could be used for government intelligence and, therefore, is a notifiable transaction,” said the Center for AI Policy in its comment.