What is the US remote-access rule? New license requirements target large clusters of Nvidia (NVDA) and other AI chips serving Chinese companies.

Department of Commerce seal with the text: What is the remote-access rule?

Key points

  • What Commerce's remote-access rule requires
  • How Nvidia's H100 compares with the limit
  • Five countries added to the US chip license list
  • When the Commerce rule takes effect

The Commerce Department's Bureau of Industry and Security (BIS) released a rule on Nov. 16 that adds US license requirements for advanced AI chips used in data centers, wherever they're located, that give companies based in China, Macau, and other countries of concern more than 12 million "aggregate TPP" of computing power. Exporters will need a license to ship chips to those data centers. A data center that already has the chips will need a license if it starts serving those companies.

TPP is the government's measure of a chip's processing power, and the limit is set in TPP, not in a number of chips. For comparison, 12 million TPP equals the combined processing power of about 758 of Nvidia's (NVDA) H100 chips, using the FP8 specifications for the H100's SXM version.

BIS posted the interim final rule on its website and said in a press release that it was "written by and at the direction of the Department of Treasury and the National Security Council." BIS said the rule is meant to make sure existing chip restrictions "cannot be circumvented through remote computing arrangements involving data centers in third countries." We reported in August that BIS was drafting a rule on remote access.

What the remote-access rule requires

Remote access means using chips in a data center over a network, usually by renting computing time, without owning the chips or having them shipped to you. The rule targets companies headquartered in Macau or in a group of countries the export rules call Country Group D:5, which includes China. It also covers companies whose ultimate parent is headquartered in one of those places.

Under the rule, exporters need a license to ship advanced AI chips when they know the chips will go to a data center that gives those companies access to more than 12 million aggregate TPP. The limit counts computing power at one data center or across data centers linked by a network. It also applies when a data center serves a customer outside D:5 that is training AI models for a D:5 company or providing models to one.

If a data center that already has the chips starts serving those companies, the owner or operator needs a license. Data centers that were already providing that much computing power to those companies before the rule takes effect have 90 days to get a license "or face future export restrictions," the rule says. BIS will review applications case by case.

The license requirement doesn't apply to what the rule calls commercial recommender systems, "such as those used by online retailers and social media companies." The rule uses the term "Super Intelligence" in place of "artificial intelligence," citing an executive order titled "Inaugurating the Era of Super Intelligence."

How Nvidia's H100 compares with the limit

The rule defines TPP as 2 times a chip's peak multiply-accumulate operations per second, in trillions, times the bit length of the operation. It also says to use the speed without sparsity, a shortcut that skips some calculations, and to use the highest TPP value the chip can reach. Aggregate TPP is the sum of the TPP of every chip in a group.

Nvidia's H100 spec page lists the SXM version at 3,958 teraflops for 8-bit FP8 math with sparsity, and half that, or 1,979 teraflops, without it. That works out to 1,979 times 8, or 15,832 TPP per chip. The H100's 16-bit and INT8 figures give the same 15,832 once sparsity is removed, and its other formats, such as TF32 and FP64, give lower values, so 15,832 is the H100's highest TPP under the rule. Divide 12 million by 15,832, and the limit equals the combined TPP of about 758 H100s.

That figure is a comparison. Other chips have different TPP values, so the number of chips that adds up to 12 million TPP depends on which chips a data center uses.

Five countries added to the US chip license list

The rule also requires a license to ship advanced AI chips to Indonesia, Malaysia, Singapore, Thailand, and Brazil. BIS says those countries present "a higher-than-normal risk of diversion," meaning chips sent there could be passed on to buyers the US has restricted. The Commerce secretary can add more countries.

Chips going to companies headquartered in the US, or whose ultimate parent is headquartered in the US, don't need the new license for those five countries if the exporter meets new reporting requirements. The rule says Commerce will watch for countries that tolerate evasion and could respond with import restrictions, new license requirements, or additions to its Entity List, among other steps.

Megaspeed, the Nvidia customer at the center of the scandal that ended Jensen Huang's run as Nvidia's CEO, is based in Singapore. Treasury Secretary Scott Bessent said on Nov. 13 that four agencies had wanted Megaspeed on the Entity List for months.

Chipmakers and the AI Diffusion Rule

The rule also changes the presumptions that chip factories and chip packagers face when they export advanced logic chips. Under existing rules, those chips can be presumed to be controlled AI chips unless certain conditions are met. The rule reduces the ways that chip designers in countries that need a license for AI chips can overcome that presumption.

Finally, the rule formally rescinds the AI Diffusion Rule, a Biden-era regulation published on Jan. 15, 2025, that set chip limits country by country. BIS says that rule was never enforced, and it announced the rescission in May 2025.

When the remote-access rule takes effect

The rule takes effect on the day it's published in the Federal Register, and comments are due 60 days after that. It hadn't appeared in the Federal Register as of 2 p.m. ET on Nov. 16. The 90-day window for data centers already serving D:5 companies starts on the effective date.

BIS estimates the rule will bring about 100 more license applications a year, along with about 970 more filings under its license exceptions and other authorizations.

Frequently asked questions

What is the US remote-access rule for AI chips?

It is an interim final rule the Commerce Department's Bureau of Industry and Security (BIS) released on November 16, 2026. It adds license requirements for advanced AI chips used in data centers, wherever located, that give companies headquartered in Macau or Country Group D:5, which includes China, more than 12 million aggregate TPP of computing power, either at one data center or across linked data centers. Exporters need a license to ship chips to those data centers, and a data center that already has the chips needs a license if it starts serving those companies. BIS said the rule was written by and at the direction of the Treasury Department and the National Security Council.

How does Nvidia's H100 compare with the 12 million TPP limit?

As a comparison, 12 million TPP equals the combined TPP of about 758 Nvidia H100 SXM chips. The limit is set in TPP, not in a number of chips. The rule defines TPP as 2 times peak multiply-accumulate operations per second, in trillions, times the bit length, using dense (non-sparse) performance. Nvidia lists the H100 SXM at 1,979 teraflops of FP8 without sparsity, which works out to 15,832 TPP per chip. The H100's FP16, BF16, and INT8 figures give the same value and its other formats give less, so that is its highest TPP. 12,000,000 divided by 15,832 is about 758.

Which countries were added to the US AI chip license requirement?

The rule requires a license to export advanced AI chips to Indonesia, Malaysia, Singapore, Thailand, and Brazil, which BIS says present a higher-than-normal risk of diversion. The requirement doesn't apply to chips going to companies headquartered in the US, or whose ultimate parent is, if the exporter meets new reporting requirements.

When does the remote-access rule take effect?

On the day it is published in the Federal Register, with comments due 60 days later. Data centers already providing that much computing power to D:5 companies have 90 days from the effective date to get a license. The rule hadn't appeared in the Federal Register as of November 16, 2026.

Are there exceptions to the remote-access rule?

Yes. The license requirement doesn't apply to commercial recommender systems, such as those used by online retailers and social media companies.

More on NVDA

Latest stories

Dennis Singleton
Dennis Singleton

Dennis Singleton was born in Australia and later moved to the United States. He has spent years following the markets, but what keeps his attention is how AI is built. He writes about the companies behind the technology, from semiconductor designers and advanced packaging to photonics, memory, networking, and the hardware powering modern AI. His approach starts with filings, earnings, and industry research, then translates the important details into clear, straightforward analysis without unnecessary hype.