The 760-H100 line: what the US rule on rented AI chips for China really covers
Everyone is saying it today: "US closes China's back door." We read all 132 pages of the new export rule and ran the numbers. Big training clusters like Megaspeed's are clearly caught. But small rental deals, recommender systems and the rule's legal footing remain weak spots.

What this is about
If you can't buy modern Nvidia chips in China, you rent them. A data center in Malaysia, Singapore or Thailand racks the GPUs, and the Chinese AI lab accesses them over the internet. Until now, that was largely legal. On November 16, the US Commerce Department's export agency, BIS, issued an Interim Final Rule, meaning a rule that takes effect immediately, with comments collected only afterward. It carries the identifier RIN 0694-AK50 and the title "Addressing Remote Access to Advanced Semiconductors." According to the department, it was drafted "at the direction" of the Treasury Department and the National Security Council, just days after Nvidia CEO Jensen Huang resigned over the Megaspeed affair.
We went through the 132-page pre-publication version on bis.gov ourselves. Our thesis: the rule deliberately targets large training clusters. But small rental deals, the special status of US companies, the carve-out for recommender systems and a shaky legal basis leave gaps.
The threshold: 12 million TPP, checked
The core of the rule is the new §744.24. You need a license if you export, reexport or transfer controlled chips or servers to a data center "worldwide" with "knowledge" that it provides, or will provide, companies headquartered in Macau or in Country Group D:5 with access to more than 12,000,000 "aggregate TPP." D:5 is the US list of arms-embargoed countries, including China, Russia, Iran and North Korea. For corporate groups, what counts is the headquarters of the ultimate parent. TPP (Total Processing Performance) is the performance metric used in US export controls. It's a chip's compute multiplied by the bit width of the operation. Per the definition in §772.1, "aggregate TPP" is the sum of all controlled chips and servers in a group.
Reports mention "around 750 H100s" in passing. We did the math ourselves, using vendor figures for dense compute, meaning without the artificially inflated sparsity numbers:
| Chip | Compute (dense) | TPP per chip | Maximum below the threshold |
|---|---|---|---|
| H100 / H200 (SXM) | 989.5 TFLOPS FP16 | 15,832 | 757 chips |
| B200 (HGX) | 4.5 PFLOPS FP8 | 36,000 | 333 chips |
| GB200 (in NVL72 rack) | 5 PFLOPS FP8 | 40,000 | 300 chips = just over 4 racks of 72 |
| H20 | 296 TFLOPS FP8 | 2,368 | ~5,067 on paper – counting unclear |
The 758th H100 puts you over the line. A GB200 NVL72 rack comes to 2.88 million TPP. Four racks are allowed, the fifth isn't. The H20, Nvidia's throttled China chip, has a TPP below 4,800, the threshold at which ECCN 3A090.a kicks in. In our reading, the text leaves open whether such chips count toward the total at all.
What does that mean in practice? For comparison: DeepSeek says it trained its V3 model on 2,048 H800s, about 2.8 million GPU hours. With 757 H100s of equal performance, that would have taken roughly five months instead of two. Our take: training a frontier model below the threshold is barely feasible anymore. Fine-tuning, experiments and serving finished models fit comfortably underneath it, though.
"Linked data centers" and the detour through third parties
Splitting a cluster across several buildings doesn't help. Under §744.24(d)(2), a "linked data center" is a group of sites that "share resources, distribute workloads and replicate data" over network infrastructure. The text explicitly states that this includes purely digitally connected infrastructure.
Straw-man setups are covered too. According to the preamble, the rule applies if a data center knows that a customer outside D:5 is training models for a D:5 company, helping with the training or providing models to it. Data centers must apply "know your customer" checks to find out who directly or indirectly owns their customers. If a customer refuses to disclose that, it's treated as a red flag that "likely establishes knowledge." If the end use changes after the fact, from a harmless customer to China training, the rule treats that as an in-country transfer that needs approval.
Plenty remains open. The text doesn't say whether the 12 million applies per Chinese customer or to all of a provider's D:5 customers combined. It also doesn't specify a time period. If you rent just under the threshold from several unlinked providers, you're not covered by the letter of the rule. That said, distributed training across different providers is technically painful.
Who's covered? Five cases
| Case | Consequence under the rule |
|---|---|
| Megaspeed-style neocloud: Singapore company, data centers in Malaysia, tens of thousands of Blackwell chips for Alibaba | Clearly covered: far above the threshold, license within 90 days, new chips for Malaysia require a license anyway |
| Small rental server in Malaysia with 512 H100s for a Chinese startup (about 8.1M TPP) | Below the threshold, no license needed under §744.24. But the operator will need a license for new chips going forward |
| European company rents 1,000 H100s in Germany and trains for a Chinese lab (about 15.8M TPP) | Covered as soon as the data center knows. Without customer checks, trouble and sanctions loom |
| US hyperscaler in Singapore | Gets chips without a license but must report monthly. Still needs a license to provide more than 12M TPP to Chinese customers |
| Chinese company runs a pure recommender system abroad | Exempt, with notification via the BIS portal |
The fourth case shows that the exemption for US companies is narrower than some headlines suggest. It only waives the new country license under §742.6. For §744.24, the text says explicitly: "No license exception may overcome this prohibition." The gap lies elsewhere. In Malaysia or Singapore, US companies can get as many chips as they like, and the agency only learns about it from monthly reports listing parties, destination, product, TPP and quantity.
The carve-out for "commercial recommender systems" is narrowly defined. It covers systems "exclusively designed and exclusively used" to rank, filter or suggest products, services or content. Anything that generates new text, images, video or code doesn't qualify, even as part of a larger system. According to BIS, recommender systems "do not present the same national security threat." Our take: China's platform giants in e-commerce, short video and delivery run exactly these systems at massive scale. Whether a compute farm really calculates "exclusively" recommendations is nearly impossible to verify from the outside.
Megaspeed as the test case
The rule looks tailor-made for this case. According to Bloomberg, Megaspeed and its subsidiaries Speedmatrix in Malaysia and PT Megaspeed International Indonesia imported Nvidia hardware worth at least $4.6 billion between 2023 and November 2025, about 136,000 GPUs. CBS puts the figure at $4.8 billion. According to Wikipedia, the biggest customer was Alibaba, which had also lent money to the Malaysian subsidiary in 2024. Bloomberg reported that Nvidia auditors initially found only a few thousand of the newest chips during a visit. Nvidia later said it had located the missing hardware in other warehouses. According to The Next Web, citing the Straits Times, Indonesian police searched the subsidiary's Jakarta office at the FBI's request. They found unopened server racks there, but no Nvidia GPUs. Megaspeed denies any smuggling and has not been charged.
A rough calculation: if the 136,000 chips were mostly B200s, that would add up to about 4.9 billion TPP, roughly 400 times the threshold. Going forward, a provider like that would need a license just for the chip delivery, let alone for renting to Alibaba. Applications from D:5 and Macau generally face a "presumption of denial," meaning they're rejected by default. For data centers in third countries, BIS reviews each case individually. One problem isn't solved: according to the analysis blog The Substrate, authorities slowed the investigation into Megaspeed, and the company still isn't on the Entity List, the US sanctions list.
Why Brazil – and no longer Norway?
Shipments to Indonesia, Malaysia, Singapore, Thailand and Brazil now require a license. The department justifies this only in general terms with an "above-average risk of diversion." The text gives no specific reason for Brazil, and Brazilian outlets like Times Brasil and Revista Fórum couldn't find one either. Some context: according to Bloomberg Línea, TikTok parent ByteDance is planning a data center worth more than 200 billion reais at the port of Pecém in Ceará. There's no evidence that this was the trigger. According to The Information (via ZeroHedge) and Chinese reports, a draft also included Norway and South Africa, but they're missing from the final version. No reason has been given. The Commerce Secretary can add more countries. The department also threatens countries that "tolerate circumvention" with import restrictions and downgrades.
What changes for Europe
At the same time, the rule formally rescinds the AI Diffusion Rule from January 2025, which BIS says was "never enforced." It had split the world into three tiers. Only ten EU states, including Germany and France, were in Tier 1. The other 17, including Poland, Portugal and the Czech Republic, would have been allowed about 50,000 H100 equivalents each through 2027, according to CSIS. Those tiers are gone, as are the export controls on model weights (ECCN 4E091), the AIA and LPP license exceptions and the program for certified data centers (Data Center VEU). For EU data centers, that removes a cap on chip purchases. But the rental rule applies to them too: if you rent out more than 12 million TPP to Chinese customers in Frankfurt, you need a US license.
Deadlines, costs and the legal question
The rule takes effect on the day it's published in the Federal Register. The pre-publication version still contains placeholders, and according to Blogspan, it couldn't be found there on the afternoon of November 16. From the effective date, existing providers have 90 days and commenters have 60 days. Shipments already in transit can be completed within 30 days. BIS itself expects only about 100 additional license applications and roughly 970 reports per year. That also suggests the rule is meant to hit only a handful of large providers.
Legally, the agency relies on sections 1753 and 1754 of the Export Control Reform Act. In our reading, the trick is not to regulate remote access itself but the delivery of chips to a data center with a known end use. Whether that holds up is disputed. A Baker McKenzie lawyer told The Information that it is "widely accepted" among export lawyers that the department can't enforce remote access under current law. BIS itself concluded in advisory opinions from 2009 to 2014 that cloud providers aren't exporters. The Remote Access Security Act could fix that. The House passed it 369 to 22 on January 12, 2026, and it's been sitting in the Senate Banking Committee ever since.
According to ZeroHedge, Nvidia lobbied against the rule, arguing that it would only push China toward its own chips faster. Backing that up: Meituan trained its LongCat model on more than 50,000 Chinese chips. On the stock market, AMD fell another 4 percent, according to TradingKey, and the semiconductor index dropped more than 5 percent. Chinese commentators on Sina expect Tencent, Alibaba and ByteDance to shift more heavily to domestic chips.
Our take
The rule is more precise than "back door closed." It targets clusters capable of training frontier models and puts large neoclouds in Southeast Asia under licensing requirements. That fits the evidence that Chinese labs have recently trained on Blackwell chips. But if you stay below 757 H100s, can make a credible case for a recommender system or buy through a US company that only files monthly reports, you're operating in gray areas. Enforcement depends on customer checks by private operators and on host countries like Malaysia, which is only now considering oversight of data centers. And before the 90 days are up, a court may have to decide whether the Commerce Department is allowed to do any of this at all.
Sources
- BIS: Interim Final Rule RIN 0694-AK50 (Vorabfassung, PDF) bis.gov
- TechCrunch: US tightens AI chip export controls techcrunch.com
- ZeroHedge: US Closes The Cloud Loophole zerohedge.com
- Trending Topics: U.S. Shuts China’s Back Door to A.I. Chips trendingtopics.eu
- CASRAI: BIS issues rule on remote access to AI chips casrai.org
- ai-tldr.dev: US remote-access chip rule ai-tldr.dev
- Blogspan: USA verlangen Lizenzen für Rechenzentren blogspan.net
- Boring Compliance (chinesisch): BIS发布临时最终规则 boringcompliance.com
- Tech Times: Commerce drafts AI chip rule (Baker McKenzie, BIS-Gutachten) techtimes.com
- Congress.gov: H.R.2683 Remote Access Security Act congress.gov
- Wikipedia: Megaspeed en.wikipedia.org
- The Next Web: Indonesian police found no Nvidia chips at Megaspeed unit thenextweb.com
- The Substrate: What's going on with Jensen Huang, Megaspeed, and export controls? the-substrate.net
- TradingKey: AMD verliert weitere 4 % tradingkey.com
- Sina (chinesisch): 芯片不出境,算力远程用? sina.cn
- Times Brasil: EUA impõem novas restrições ao Brasil timesbrasil.com.br
- Revista Fórum: Trump coloca Brasil em lista de risco revistaforum.com.br
- Bloomberg Línea: TikTok anuncia data center no Brasil bloomberglinea.com.br
- TrendForce: Breakdown of the tiered restrictions (AI Diffusion) trendforce.com
- Flopper: GB200 NVL72 Spec Sheet flopper.io
- Crypto Briefing: Nvidia 51%-Regel für Cloud-Kunden cryptobriefing.com
Ask the AI assistant
Questions about this article, installation or background? Our assistant knows this page and runs on our own machine, not in someone else's cloud.
The Assistant is an AI (Gemma 4) and can make mistakes – please check important details against the source. Questions are used only to generate an answer and are deleted after one hour. Please don't enter any personal data.
Comments
What do you think? Already tried it? Write it here, no sign-up needed.
Comments are public. Links aren't allowed, and we delete abusive posts. You can delete your own comment at any time in this browser. Learn more in the Privacy Policy.