The Commerce Department moved Monday to close the route that let Chinese firms rent American computing power abroad, issuing an interim final rule that requires export licenses for advanced artificial intelligence chips sent to five countries and for data centers that serve Chinese customers above a set threshold.

The rule, issued by the department's Bureau of Industry and Security, does not ban the sales. It makes them subject to case-by-case review, and it extends the reach of export controls from the chip itself to the building where the chip is installed. The bureau described the measure as addressing "remote access to advanced super intelligence (SI) chips and servers," using the administration's term for the most powerful AI hardware.

The timing is not incidental. The rule landed four days after Jensen Huang resigned as Nvidia's chief executive amid revelations about an affair with an executive at Megaspeed International, a Singapore-based data center firm suspected of diverting Nvidia chips to China. It also arrived as Congress prepares to question Commerce Secretary Howard Lutnick and Huang about why Megaspeed stayed off the entity list. The rule does not name Megaspeed.

How the route worked

Export controls have focused on the chip for four years. The United States restricts what performance level can be sold to China, and Nvidia responded by designing lower-performance processors for that market. The controls also cover chipmaking equipment and high-bandwidth memory, which makes it harder to build a substitute at home.

The system left an opening that did not require a chip to enter China. A Chinese company could rent computing capacity from a data center in Singapore, Malaysia or another third country, and the chips powering that rental stayed where they were. The sale was legal because the buyer was in a permitted destination. The end user was not.

Megaspeed is the case that made the opening visible. The Singapore-registered firm traded more than $4 billion of Nvidia chips over two years, according to prior reporting, and U.S. officials suspect that some of those processors were routed onward to China. Megaspeed has denied sending chips to China. It built data centers in Malaysia and sold capacity across the region, which put it at the center of the gap between the chip rules and the cloud business.

The rule published Monday addresses that gap directly. It requires a license when a data center provides more than 12,000,000 aggregate total processing performance to companies headquartered in China, Macau or Country Group D:5, the export-control category that includes China, Russia, Iran and North Korea. The threshold is measured across linked data centers, a structure meant to prevent an operator from splitting capacity among affiliated facilities to stay under the limit.

The license requirement applies to centers that already serve those customers and to centers that begin serving them after the chips are exported. That second part closes the argument that a facility was compliant when it bought the hardware. Existing data centers have 90 days from the effective date to obtain a license. BIS said facilities that do not apply will face future export restrictions, a phrase that leaves the agency room to decide later what the penalty is.

Interim final rules take effect when they are published in the Federal Register, and this one carries a 60-day comment period that begins at publication. As of Monday evening, the document had not appeared in the register's public listing, so the clock had not started. The rule formally rescinds the Biden administration's AI Diffusion Rule, which the Trump administration had declined to enforce but had left on the books.

The five countries

The rule also requires licenses for exports, reexports and in-country transfers of specified advanced chips to Indonesia, Malaysia, Singapore, Thailand and Brazil. BIS said those destinations carry a higher-than-normal risk of diversion, according to NBC News. The requirement is not a blanket prohibition; licenses are available case by case.

Singapore and Malaysia are established data center hubs, and both have drawn billions of dollars of investment from American cloud companies. Indonesia, Thailand and Brazil are newer entrants to the same buildout. BIS said businesses in the covered countries had bought billions of dollars of high-end chips in recent years, raising concern about where they ended up.

The rule includes an exemption for buyers headquartered in the United States, or owned by a US parent, if they report the parties to the transaction, the chips' final destination and the product details. That carve-out appears designed to keep the regional investments of Microsoft, Oracle, Amazon and other American firms outside the licensing queue while still restricting who can use the compute. The access restriction applies even when the hardware is owned by a US company, according to reporting on the rule.

There is also an exception for commercial recommendation systems, subject to reporting. Generative AI systems do not qualify.

For Malaysia, the exposure is direct. It has marketed itself as a neutral host for AI capacity, and its data center corridor has grown on the premise that global customers could use it without regard to Washington's China policy. Joshua Kurlantzick of the Council on Foreign Relations told Fortune that for Malaysia "the controls could be disastrous."

The pen and the case

The rule's provenance is unusual. BIS said the interim final rule was "written by and at the direction of" the Treasury Department and the National Security Council. Export-control rules are normally drafted inside BIS, and the attribution is a public acknowledgment that the policy was driven from outside the bureau.

It also lands in the middle of an argument about who controls chip policy. On Friday, Treasury Secretary Scott Bessent said four agencies had favored blacklisting Megaspeed for months and that Commerce, which controls the entity list, had not acted. "But Commerce has the pen," he said on CNBC. The new rule shows the pen being used, though not necessarily by the hand Bessent described.

Chris McGuire of the Council on Foreign Relations disputed the authorship account. "I have never seen a BIS rule that was not written by BIS," he told The Next Web. He called the rule a step forward while saying loopholes remain, including inside the United States.

The rule does not resolve the Megaspeed case. The entity list is a separate instrument, and a listing decision would require its own determination. Congress is moving on a parallel track. Three House committees have requested a closed-door briefing on Nov. 18, the Senate Banking Committee has scheduled a Dec. 3 hearing with Lutnick, and three chip bills in the Senate version of the defense authorization act would write tighter standards into statute. One of them, the AI Overwatch Act, targets Chinese access to American chips through overseas data centers, the same route the new rule addresses.

The rule can be read as an attempt to take the policy away from Congress. If BIS can show that licensing is working, the case for legislation weakens. If the first 90 days produce a backlog and a set of exemptions that look like the old system, the case for legislation gets stronger. That is the political test the agency now faces.

What the market heard

Nvidia shares were down 3.95 percent at $156.42 in afternoon trading Monday, and the Philadelphia Semiconductor Index fell as much as 5.5 percent, according to Yahoo Finance. The chipmaker is already under pressure from the leadership change. Colette Kress, the chief financial officer, is interim chief executive while the board searches for a permanent successor.

The rule is designed to constrain demand from Chinese customers without cutting off the Southeast Asian data center market that American suppliers and cloud companies have spent years building. Whether it can do both is the question analysts focused on. Hanna Dohmen of Georgetown's Center for Security and Emerging Technology said the regulatory path in Washington would be harder and bumpier as the bills and the agency process compete.

Ryan Fedasiuk of the Foundation for American Innovation took a different view of the rule itself, calling it "a long-overdue step in the right direction." The split is not about the goal. It is about whether a licensing regime administered case by case can move faster than the networks it is meant to close.

The enforcement test

Licensing rules depend on knowing who the end user is, and that is the weakest link in the chain. A chip sold to a data center in Malaysia can be installed in a rack leased to a Singapore intermediary that serves a customer in Shenzhen. The rule reaches linked data centers and indirect arrangements on paper. Proving those links in an application, and later in an audit, is the harder part.

BIS said it would monitor evasion and consider additional licensing requirements, import limits and Entity List additions. The agency did not announce new enforcement resources, and the 90-day window gives it a large batch of applications to review at once. The aggregation language is the main tool for stopping operators from splitting capacity among affiliates, but it depends on the agency's ability to map ownership and control across borders.

The first applications will also define how the agency treats ownership. The US-parent exemption turns on control, not on where the customer is incorporated, and the difference between a US-owned regional subsidiary and a joint venture with a Chinese partner will decide which queue a data center joins. Companies have an incentive to structure around that line. How BIS reads the first few structures will tell the market how far the rule reaches.

A license is only as strong as the information in the application. The rule creates no verification mechanism on the ground, and the agency has limited presence in the countries it covers. The chip-tracking provisions in Congress would add a technical tool, but they are not law yet.

What comes next

The comment period gives the five countries, chipmakers, cloud providers and trade groups a chance to argue about the threshold and the country list before the rule is finalized. Because the rule is an interim final rule, it takes effect before those comments are resolved. The agency can revise it later, but operators have to plan against the version that exists now.

The political calendar is tighter. The House briefing on Nov. 18 will cover the Megaspeed investigation, and the Senate Banking Committee has committed Lutnick to a Dec. 3 hearing and invited Huang. The NDAA conference will decide whether the chip bills survive, and Nvidia's earnings on Nov. 17 will give investors a first read on how the company is pricing the China channel.

For the data center industry, the practical question is narrower. A license is now part of the cost of serving a Chinese customer from anywhere, and the 90-day clock starts when the Federal Register publishes the rule. Operators that have been treating jurisdiction as a design choice are about to find out how much the choice costs.

Primary sources

  1. Bureau of Industry and Security for the interim final rule, its stated authorship and the description of remote access.
  2. NBC News for the five countries, the remote-access provision and Ryan Fedasiuk's comment.
  3. Fortune for the Megaspeed context, the US carve-out and Joshua Kurlantzick's comment.
  4. The Next Web for the 12 million aggregate TPP threshold, the 90-day window, the market reaction and Chris McGuire's comment.
  5. Yahoo Finance for the license mechanics, the recommendation-system exception and Nvidia's share price.
  6. Seoul Economic Daily for the five-country license scope and the US-parent reporting exemption.