AI Chip Export Rule Sets Cloud License Line at Roughly 750 Nvidia H100s for China-Linked Tenants
23 hour ago / Read about 28 minute
Source:TechTimes

Valentine Tanasovich/pexels.com

The US Commerce Department's Bureau of Industry and Security (BIS) released an interim final rule on Monday that requires a license to ship advanced AI chips to any data center that will give companies headquartered in China, Macau or other arms-embargoed Country Group D:5 nations access to more than 12 million aggregate "total processing performance" (TPP). TechCrunch estimates the threshold equals the maximum output of about 750 Nvidia H100 or H200 chips, an approximation that does not come from the rule itself.

The rule also requires licenses for advanced chip exports to Indonesia, Malaysia, Singapore, Thailand and Brazil, and it gives data centers that already serve covered customers 90 days from the effective date to obtain a license. That effective date remains open pending Federal Register publication. Operators and renters of GPU capacity in those five countries now face new demands to document who their tenants are.

The release follows a week in which Jensen Huang resigned as CEO of Nvidia after admitting an affair with Alice Huang, a former executive of Singapore data-center company Megaspeed whom US officials suspect of working for Chinese intelligence. Megaspeed denies wrongdoing, and Chinese officials have said China was not involved.

Rule Attaches Licenses to Chip Shipments Instead of Cloud Logins

BIS cannot license a cloud login directly, because under current law cloud computing is not an export. The rule works around that gap. It bars exporters from knowingly supplying chips to a restricted data center, and, according to BERI's analysis, it treats a later switch to restricted use as an in-country transfer needing authorization.

As posted, the rule does not declare remote access itself an export; it reaches remote access indirectly by attaching conditions to chip shipments, transfers and reexports. The BIS preamble calls the new control a "logical extension of existing controls."

Does 12 Million TPP Equal About 750 H100s?

BIS defines TPP as 2 times MacTOPS times bit length, where MacTOPS is a chip's theoretical peak multiply-accumulate rate, aggregated across all processing units on the chip. Because the formula scales with raw throughput, chips with higher per-chip performance reach the 12 million line with fewer units.

The threshold is not limited to one building. It applies to a single data center or to linked data centers, including digitally linked infrastructure, and the preamble treats a remote user as covered when it is the exclusive or primary user, meaning more than 50 percent of compute. BERI says the threshold will count compute across all restricted companies a data center serves, so one large tenant or several smaller ones can push a site over the line. It adds that tenants that train or supply models for a covered company may also face use attestations.

Final Rule Lists Five Countries as Reported Draft Named Seven

TechTimes reported a seven-country draft on Nov. 13, relaying reporting by The Information that the list would include Singapore, Malaysia, Thailand, Brazil, Indonesia, Norway and South Africa. The published text names five: Indonesia, Malaysia, Singapore, Brazil and Thailand. Commerce says those five carry a "higher-than-normal risk of diversion," according to The Hill.

Brazil is the least expected name. Its president signed the ReData data-center tax regime into law on Sept. 15, about two months before the rule. End users whose ultimate parent is headquartered in the United States fall outside the five-country license requirement, but only if the exporter files monthly reports under Section 743.10 listing each party and its parent, each product's export classification, its TPP and its quantity.

Operators Carry Burden Through Knowledge Presumption and Ownership Tracing

Exporters must be able to show they lacked "knowledge" that a data center would exceed the threshold. Under Note 3, a foreign data center that cannot supply enough ownership information is presumed to give the exporter that knowledge unless the red flag is resolved first. BERI's analysis says the burden therefore shifts to the data center, which will push it onto tenants through ownership and use attestations and change-of-use notice duties.

The same logic reaches chipmaking. Designers headquartered in D:1, D:4 or D:5 countries, or in the five listed countries, can no longer overcome the presumption that their chips are controlled 3A090.a items, and foundries and outsourced assembly-and-test providers need licenses to export such chips for those designers.

BIS estimates the paperwork at about 100 additional license applications and a net increase of about 970 reporting and exemption submissions a year. A narrow exclusion applies because commercial recommender systems are exempt when they are designed and used exclusively for ranking, filtering or suggesting products or content; systems that generate novel content are not. BERI flags that the preamble and Section 744.24(b)(2) read differently on the recommender timeline, and it advises against planning around the more generous reading.

BIS Calls Control Logical Extension While Export Lawyers Doubted Authority

TechTimes reported in August that an unnamed Baker McKenzie attorney told The Information it is "widely acknowledged" in the export-control bar that Commerce cannot enforce a remote-access regulation under existing law. The House passed the Remote Access Security Act (H.R. 2683) 369-22 on Jan. 12 to add that authority explicitly, and the bill awaits Senate action.

The published rule does not wait for that bill. It cites the Export Control Reform Act of 2018, relying on 50 U.S.C. 4812 for regulating exports, reexports and in-country transfers, on 4813 for control lists and licensing, and on 4821 to issue the rule without prior notice and comment. Chris McGuire of the Council on Foreign Relations called the rule "a major step forward" while saying gaps remain, including some covering activity inside the US, according to The Next Web. CFR senior fellow Joshua Kurlantzick told Fortune the controls could be disastrous for Malaysia, The Next Web reported.

Treasury and National Security Council Wrote Rule After Huang Resignation

BIS says the rule was written by, and at the direction of, the Treasury Department and the National Security Council. The Hill reports that The New York Times described Commerce Secretary Howard Lutnick as clashing with top officials and taking a softer line on Nvidia's China chip sales. The Times also reported that Commerce's Megaspeed probe was handed to the Justice Department last year.

Reuters reported last week that Chinese AI companies were rushing to secure overseas computing capacity for fear of tighter US restrictions.

Lutnick and Kessler Face Congress on Data Centers and H200 Exports

Lutnick has agreed to appear before the Senate Banking Committee, and the House Foreign Affairs Committee has asked BIS chief Jeffrey Kessler to testify. Both arrive with technology-policy records already under scrutiny. In December 2025, congressional Democrats led by Sen. Elizabeth Warren asked the Commerce inspector general to examine whether Lutnick's promotion of AI data centers conflicts with his family's financial interests through Cantor Fitzgerald and Newmark Group, which the letter describes as a real estate broker that facilitates leases for AI data centers. Kessler told the House Foreign Affairs Committee in July that H200 shipments to China were trivial despite an estimated $10 billion in approved licenses.

Chip Stocks Slid Monday Morning With Other Drivers in Play

The Nasdaq Composite was down 1.6% as of 9:51 a.m. ET, with Sandisk down about 7%, Micron down 6%, Intel down 5% and AMD down 4%, according to MarketSpeaker, which attributed the selloff mainly to chip-sector weakness and said the rule added uncertainty without a quantified revenue impact. AMD was down 4.4% as of 10:53 a.m. ET, and Singapore accounted for 12.4% of AMD's 2025 revenue. Nvidia fell as much as 2.9% and the Philadelphia Semiconductor Index as much as 5.5%, according to Bloomberg figures relayed by The Next Web.

GPU Renters Should Get Ownership and Use Terms in Writing

Non-US tenants get no exclusion from the five-country requirement, so chips shipped into their capacity in those countries would need a license. BERI advises tenants to expect ownership and use attestations and a contractual duty to notify providers before workload or customer mix changes. The 90-day window applies to data centers already serving D:5 or Macau companies and runs from the effective date, which has not been set.

Federal Register Publication, Nvidia Earnings and Trump-Xi Meeting Set Next Dates

Federal Register publication will set the effective date and start a 60-day comment period, and shipments already en route under prior eligibility may be completed within 30 days after publication. Some press reports say the rule takes effect in 90 days; the BIS text gives that window to existing data centers and ties the effective date to publication instead.

Nvidia will report results Tuesday, Nov. 17, at about 4:20 p.m. ET, with Colette Kress now serving as interim CEO, and investors will look for any comment on licensing effects in the five countries. Trump meets Xi Jinping in Shenzhen on Thursday, according to The Next Web. None of the five governments had issued a formal response in the coverage reviewed.

For operators and renters, the decision frame is concrete. A cluster in the range of 750 H100-class chips serving China-headquartered tenants could cross the license line, the burden of proving who the tenants are falls on operators, and the rule's reliance on chip shipments rather than cloud access has not yet been tested against its stated legal authority.

Frequently Asked Questions

Does the rule stop Chinese companies from renting Nvidia GPUs abroad?

Not outright. It requires licenses for chip shipments to data centers that will give D:5 or Macau-headquartered companies access above 12 million aggregate TPP, and existing operators have 90 days from the effective date to obtain one. Below that line, the new end-use control does not apply, though earlier BIS guidance still requires licenses for chips sold to entities whose ultimate parent is headquartered in China.

How many Nvidia GPUs is 12 million TPP?

TechCrunch estimates roughly 750 Nvidia chips, either H100 or H200, a figure that is its own approximation rather than a number in the rule. Because the count is aggregate across linked data centers and scales with each chip's throughput, newer chips can reach it with fewer units.

Which countries now need a US license for advanced AI chips?

Indonesia, Malaysia, Singapore, Thailand and Brazil. US-headquartered recipients are excluded only if the exporter files monthly reports under Section 743.10.

When does the rule take effect?

On Federal Register publication, which has not happened yet. Comments are due 60 days after publication, and existing data centers serving covered customers have 90 days from the effective date to get a license.