The threshold is 12 million TPP. That stands for total processing performance, Nvidia’s own benchmark for a chip’s raw compute, and it is the number the Bureau of Industry and Security picked to decide which data centers now need an export license before they can serve a customer tied to China. Using Nvidia’s published datasheets, 12 million TPP works out to roughly 758 H100 or H200 GPUs, 333 B200s, or 222 B300s, according to an analysis of the rule. Cross that line, in aggregate, across any data centers that are physically or digitally linked, and the operator needs Commerce’s permission to keep selling compute to that customer.
What the rule actually closes is a gap that was legal by construction. From 2021 to 2024, the US barred its own cloud providers from letting restricted customers rent time on advanced chips remotely. No equivalent rule applied to a cloud operator sitting in Singapore, Malaysia, Indonesia, Thailand or Brazil. A Chinese firm could not buy an H100 and ship it home, but it could rent one by the hour from a data center abroad and run its training job over the network, never touching US export control because no physical chip crossed a restricted border. Ars Technica reported that this is the loophole the Commerce Department says it has now closed.
How the old workaround ran Remote-compute routing before the new BIS rule
| From | To | How |
|---|---|---|
| US chipmaker (Nvidia, AMD) | Data center, 3rd country (Singapore, Malaysia, Indonesia, Thailand, Brazil) | legal chip sale |
| Data center, 3rd country (Singapore, Malaysia, Indonesia, Thailand, Brazil) | China-parented customer (no physical chip crosses the border) | rented remote access |
| China-parented customer (no physical chip crosses the border) | AI model trained/run remotely | training/inference job |
The interim final rule, issued Sunday and described by BIS as written “by and at the direction of” the Treasury Department and the National Security Council, does three things. It bans remote access to covered GPUs for parties in Country Group D:5, the US list that includes China and Macau, once a data center’s aggregate capacity to those customers passes 12 million TPP. It adds a separate, lower bar for the five named countries: any sale of advanced AI chips or servers into Indonesia, Malaysia, Singapore, Thailand or Brazil now needs a license, full stop, because BIS has designated them places that “present a higher-than-normal risk of diversion.” And it aggregates: a provider serving ten small China-parented tenants, none individually over the line, still crosses it once their usage is added together. The rule takes effect 90 days after it is posted to the Federal Register, under RIN 0694-AK50.
One carve-out matters for the industry’s bottom line: infrastructure “exclusively used for commercial recommender systems,” the kind of ranking engine that powers e-commerce and social-media feeds, is exempt. A system that generates new text or images is not. That line is the rule’s attempt to spare routine cloud business while still catching anything that looks like frontier-model training.
The scandal that forced the question #
The immediate trigger is Megaspeed, a Singapore-based cloud and data-center firm that was, until this year, one of Nvidia’s largest customers in Southeast Asia. Nvidia’s then-CEO Jensen Huang resigned after disclosing a relationship with Alice Huang, a former Megaspeed executive US investigators suspect of acting for Chinese intelligence; China has denied she was a spy, and Megaspeed has denied any smuggling, saying she left the company in 2025. US investigators allege she used industry contacts and shell companies to move 240,000 export-restricted chips, mostly H100s and H200s, into China, according to CBS News, cited by Ars Technica. The Information put the value of that flow at roughly $7.5 billion; the Wall Street Journal separately tabulated close to $12 billion in top-tier chips routed to Chinese buyers between March 2024 and June 2025. The two figures come from different outlets using different transaction sets, and neither has been confirmed by a regulator; nobody has shown that every chip in either count entered China illegally.
What makes the Singapore connection more than circumstantial is Nvidia’s own disclosure history. For the quarter ended July 27, 2025, Nvidia reported $10.2 billion in revenue billed to Singapore, 22 percent of total revenue and up 81 percent year on year, against 8 percent two years earlier. Nvidia’s explanation was that Singapore served as a billing address for US-headquartered customers, with the hardware shipped elsewhere; the company said more than 99 percent of that revenue was for orders placed by US-based buyers. The next quarter, Singapore vanished from Nvidia’s geographic breakdown entirely, after the company switched its reporting basis from customer billing address to parent-company headquarters. Treasury Secretary Scott Bessent has since cited reports that only 1 percent of chips billed to Singapore stayed in the country.
Bessent has been the loudest internal voice for tightening the rule, and he has said so on camera. “The real story here should be: Did Nvidia know their client, and was there a robust effort at Nvidia to make sure that there was no leakage and that they were selling to what were obviously pass-through shell companies?” he said on CNBC, adding that he wanted answers “in the next 10 days from the board and from the outside counsel,” a deadline that landed two days before Nvidia’s next earnings report. Commerce Secretary Howard Lutnick, who holds the actual statutory authority to blacklist a company, had resisted adding Megaspeed to the entity list and reportedly favored looser controls Nvidia argued would preserve US technology’s global lead. Trump has backed Lutnick publicly: “if you didn’t delay it… you could have a blow-up with China,” he told reporters, and he said at the same gaggle that “nobody’s proven that she’s a spy. She may not be a spy,” about Alice Huang.
What the rule costs, and who pays it #
The compliance burden lands on data-center operators, not chipmakers. Firms in the five flagged countries now face monthly reports naming every party in a transaction, its parent company, and the chips’ final destination — modeled on bank anti-money-laundering checks rather than anything the chip-export regime has required before. Neil Shah, vice president of research at Counterpoint, called it a “massive ‘Know Your Customer’s Workload’” burden and said it creates “a severe short-term compute bottleneck for Chinese tech giants” and, over a longer horizon, pushes Beijing toward what he called “full-stack sovereign alternatives” such as Huawei’s Ascend chips — the same dynamic Chinese analysts cited to the South China Morning Post and the People’s Daily, who told those outlets the policy would “accelerate Chinese companies’ shift to domestic AI chips” and that “the cost of switching back to US platforms will rise” the longer the curbs hold. Sun Chenghao, a senior fellow at Tsinghua University’s Center for International Security and Strategy, put the US risk in blunter terms to the People’s Daily: “what US companies stand to lose could be customers and technological influence for years to come.”
Joshua Kurlantzick, a senior fellow at the Council on Foreign Relations, said the fallout is uneven across the five named countries: “For Malaysia, the controls could be disastrous,” given how much of the country’s recent data-center investment rode on exactly the remote-rental model the rule now licenses. Brazil, by contrast, already requires data-center operators to reserve at least 10 percent of capacity for local users, a rule that predates this one and that may insulate it somewhat. Singapore has limited new data-center construction on its own land constraints for years, independent of US policy; Indonesia’s exposure is smaller still, since Commerce does not count it as a major data-center hub yet.
The rule also quietly buries an older one. BIS’s announcement formally rescinds the Biden administration’s AI Diffusion Rule, written five days before Trump’s inauguration and never enforced under Trump, whose Under Secretary for Industry and Security, Jeffery Kessler, said the administration would instead pursue what BIS called a “bold, inclusive strategy” for sharing US AI technology with “trusted foreign countries” while excluding adversaries. The new rule is narrower and more mechanical than Diffusion’s country-tier system: it does not classify nations by trust level, it counts TPP.
Trump was in China this week for the summit covered in an earlier dispatch from Shenzhen, and a senior administration official told The New York Times that Trump “did not expect countermeasures from China in response” to the new chip rule. Hong Yuan, secretary-general of the Center for American Military Studies at the Chinese Academy of Social Sciences, offered the People’s Daily a different read: “Before any judicial causality was established, Washington handed down a political conviction, substituting suspicion for proof, espionage for personal contact, and an industrial embargo for private conduct. The issue is not the cause of tighter controls, but a political catalyst.” Whether Beijing treats the rule as routine enforcement or as retaliation for the Jensen Huang affair will show up first in whether China moves against rare-earth exports or Nvidia’s market access, not in anything either government says this week.
The number to watch over the next 90 days is not political. It is whichever data-center operator in Kuala Lumpur, Jakarta, Bangkok or São Paulo is first to either file for a license, cut off its China-parented tenants, or get caught doing neither.