US export controls on AI chips attach to the chip and to who ends up with it, not to the invoice. Since 7 October 2022, chips that pass a performance test in export classification 3A090 have needed a Commerce Department licence for China, and the Bureau of Industry and Security widened the licence requirement in its October 2023 rule to destinations in Country Groups D:1, D:4 and D:5 that are not also in A:5 or A:6.
Nvidia’s filings show why the invoice can mislead. Its 10-K and 10-Q tables put $51,976 million of revenue in Singapore by customer billing location across fiscal 2023, 2024, 2025 and the first half of fiscal 2026, by our arithmetic, and the 10-K said shipments to Singapore were less than 2% of fiscal 2025 revenue. More than $21,074 million billed to Singapore that year went elsewhere, by our arithmetic, as set out in our analysis of the Singapore filings.
Last updated: 12 November 2026. This explainer is revised when a rule, a filing or a court record changes; the date above shows the latest revision.
Business Pill · GPUs Explained
An evergreen Business Engineer explainer on why AI runs on GPUs, the chips the export rules classify. It teaches the general idea only and does not discuss export controls.
The key insight: As we read the rules and the filings, US controls attach to the chip and its end user, so the invoice, the shipment and the user can sit in three different countries: Nvidia billed $23,684 million to Singapore in fiscal 2025 while shipments there were less than 2% of its revenue.
The October 2022 rule added two export classifications: 3A090 for specified high-performance chips and 4A090 for computers and components that contain them. A chip was caught if it had an aggregate bidirectional transfer rate of 600 Gbyte/s or more and a processing score of 4800 or more, a threshold the rule says “can be met with 600 tera integer operations at 8 bits or 300 tera FLOPS at 16 bits”.
The October 2023 rule rewrote the test. Paragraph 3A090.a now covers chips with a “total processing performance” of 4800 or more, or of 1600 or more with a “performance density” of 5.92 or more, and a new 3A090.b catches somewhat less powerful chips. The rule said BIS “also seeks to further impair diversion channels through third countries, particularly those with AI commercial and research ties to the PRC”.
The AI Diffusion rule, published on 15 January 2025, would have sorted the world in three. Eighteen listed countries plus the United States could obtain the most advanced chips largely without a licence; China, Macau and the rest of Country Group D:5 kept the strict rules; every other destination faced country allocations, a cumulative maximum installed base of 790,000,000 TPP, the rule’s processing-performance measure, from 2025 to 2027. Compliance was due on 15 May 2025.
Two days before that date, on 13 May 2025, the Commerce Department announced a rescission, said enforcement officials had been instructed not to enforce the rule, and said BIS plans to publish a regulation formalising the rescission and will issue a replacement rule. That was 118 days after the rule was published, by our arithmetic. The same release imposed additional export licensing requirements on chips for Malaysia and Thailand, citing “known instances of diversion”.
One later change runs the other way for one chip. A rule published on 15 January 2026 moved licence reviews for the Nvidia H200 and less advanced chips to China and Macau from a presumption of denial to case-by-case review, if the exporter certifies, among other things, that the item undergoes independent third-party testing in the United States. Each change is dated in our Nvidia and China timeline.
Three paths a US-controlled AI chip can take. Paper path: Nvidia’s fiscal 2025 10-K put $23,684 million of revenue in Singapore by billing location and shipments to Singapore at less than 2% of revenue; the more than $21,074 million that went elsewhere is our arithmetic. Physical path: transit countries named in four Justice Department releases from August 2025 to October 2026, which are charges, not findings; Malaysia appears in three of the four, our count. Access path: BIS’s October 2023 rule and 13 May 2025 guidance; the Remote Access Security Act passed the House and is not law.US export-control actions on advanced AI chips: the October 2022 and October 2023 rules (87 FR 62186, 88 FR 73458), the AI Diffusion rule of January 2025 (90 FR 4544) and its rescission announced on 13 May 2025, Malaysia’s permit of 14 July 2025, the January 2026 H200 licence-review rule (91 FR 1684), the House vote on the Remote Access Security Act and the remote-access rule reported by The Information this month, which has not been published. The 118 days from the Diffusion rule to its rescission are our arithmetic.
Billed, Shipped and Used Are Three Different Places
Nvidia’s 10-K for fiscal 2025 counted revenue by where customers were billed. Singapore went from $2,288 million in fiscal 2023, 8.5% of revenue, to $6,831 million in fiscal 2024, 11.2%, and $23,684 million in fiscal 2025, which the filing put at 18%. By our arithmetic the Singapore line grew 10.4 times over those two years, while total revenue grew 4.8 times.
The filing’s footnote explains the gap: “Customers use Singapore to centralize invoicing while our products are almost always shipped elsewhere”. It put shipments to Singapore at less than 2% of fiscal 2025 revenue, which caps them below $2,610 million, by our arithmetic.
The 10-Q for the quarter to July 2025 used the export classifications themselves. Singapore was $19,173 million of the first half’s $90,805 million, which the filing put at 21%, and it said over 99% “of controlled Data Center compute revenue billed to Singapore was for orders from U.S.-based customers for the second quarter and first half of fiscal year 2026”. It defined that revenue as products that “meet the characteristics of Export Control Classification Numbers 3A090.a or 4A090.a”.
From the quarter to October 2025 Nvidia counts revenue by customer headquarters, with no Singapore row, as covered in our piece on the recast tables.
Nvidia revenue billed to customers in Singapore, from its fiscal 2025 10-K (fiscal 2023 to 2025) and its 10-Q for the quarter to 27 July 2025 (first half of fiscal 2026, 21% of revenue). The fiscal 2025 split uses the 10-K’s statement that shipments to Singapore were less than 2% of revenue. The fiscal 2023 and 2024 shares, the split and the $51,976 million total are our arithmetic.
The Three Diversion Paths
The first path is on paper: an invoice in one country and a shipment to another, which is what Nvidia’s footnote describes; as we read the rules, the licence turns on the destination and the end user, not the billing address. The second is physical: a chip is exported to a third country and moved on. The third is access: the chip stays abroad and someone elsewhere rents its computing power.
For the physical path, BIS’s guidance of 13 May 2025 lists red flags for exporters. They include a customer that received no exports of advanced chips before October 2022, consignees such as freight forwarders or logistics companies that do not usually need large numbers of chips, and a data centre that cannot affirm it has the power, cooling or space to run them. The guidance says data centres able to run servers at or above 10 megawatts merit additional scrutiny.
Court records show the route prosecutors describe. In four Justice Department cases we read, announced between August 2025 and October 2026, prosecutors allege exports through third countries. An August 2025 complaint describes shipments to freight forwarders in Singapore and Malaysia, and a November 2025 indictment alleges 400 Nvidia A100 GPUs were exported to China through Malaysia and Thailand. In March 2026 complaints, prosecutors describe an order of 750 servers for about $170 million in the name of a Thailand company, which was not completed, and an October 2026 indictment alleges more than $300 million of servers sent to China through countries such as Malaysia and Singapore.
By our count, Malaysia appears in three of the four cases. Every one is a charge, not a finding, and each release says the defendants are presumed innocent.
Host governments have moved too. Singapore’s home affairs minister said at a media conference on 3 March 2025 that a fraud case, which he called an independent Singapore investigation, involved servers that came into Singapore and then went to Malaysia, with possible false representations about their final destination, and that “We assessed that the servers may contain Nvidia chips”. He added: “we only know what we know in Singapore, we have no visibility beyond our borders”.
Malaysia’s trade ministry said on 14 July 2025 that all exports, transhipments and transits of high-performance AI chips of US origin need a Strategic Trade Permit, and its trade minister said on 12 November that Malaysia is considering a law under which data centre applicants would disclose the number of GPUs brought in, as reported in our piece on Malaysia’s proposed disclosure law.
For the access path, the October 2023 rule said open-source reporting had identified PRC companies “accessing and operating datacenters located outside of the PRC with the ICs subject to EAR controls”, and BIS said it “continues to evaluate how it may approach this through a regulatory response”. The 2025 guidance says a licence may be required under catch-all controls when an exporter has knowledge that a cloud provider will train models for parties headquartered in D:5 countries or Macau for weapons of mass destruction or military-intelligence uses.
A wider fix is pending. The Remote Access Security Act passed the House by 369 votes to 22 on 12 January 2026 and awaits the Senate, and The Information reported this month that a Commerce rule on remote access is close, as covered in our piece on the remote-access rule. Reuters reported that Chinese AI companies are trying to rent capacity abroad, as set out in our piece on the capacity rush.
Members of Congress have called for investigations and export-control bills, as counted in our tally of the calls, and our explainer on Megaspeed sets out what is on record about the Singapore-based data-centre company, which has denied wrongdoing.
Update, 12 Nov 2026, 18:56Z: an analyst estimate. SemiAnalysis, a chip research firm, estimates that more than 1.5GW of restricted Nvidia chips ended up in China. In an article published at 18:06 UTC, the firm writes that its Accelerator Model “tracks every AI chip’s production and destination, and our data shows more than 1.5GW of Nvidia chips, which are restricted, still ultimately ended up in China.” The figure is SemiAnalysis’ estimate from its own model, measured in power rather than units; the article does not publish the underlying data. It is not an official count, and SemiAnalysis does not say through which of the paths above the chips travelled.
The Structural Read
As we read the sequence, each rule moved the control point: from a chip’s performance in 2022, to the buyer’s headquarters and third countries in 2023, to country allocations in a rule rescinded before compliance, and now toward remote access, which the 2023 rule flagged and the House-passed bill would write into the export-control statute.
As we read the four court cases, the route prosecutors allege runs through the hubs the rules now name: Malaysia appears in three of the four, by our count, and the May 2025 release added licence requirements for Malaysia and Thailand.
Bureau of Industry and Security, 88 FR 73458, 25 October 2023
“BIS is also concerned regarding the potential for China to use IaaS solutions to undermine the effectiveness of the October 7 IFR controls and continues to evaluate how it may approach this through a regulatory response.”
Three Implications
FOR EXPORTERS BIS’s 2025 guidance asks for end-user certifications and a written attestation from the data centre that the end user is authorized to operate there and has the infrastructure to run the servers.
FOR READERS OF NVIDIA’S FILINGS Billing location, ordering customer and headquarters are different measures; only fiscal 2025 carries a number for shipments to Singapore, as a ceiling.
FOR HOST COUNTRIES Malaysia has required a Strategic Trade Permit for exports, transhipments and transits of US-origin AI chips since 14 July 2025, and its trade minister said on 12 November that a GPU disclosure law is under consideration.
The framework argues that the AI supply chain is “creating geopolitical leverage that operates independently of who has the best models”.
As we read it, export controls are where that leverage is written down: the rules decide which chips, which countries and now which kinds of access count. The framework predates this explainer and does not discuss it.
What Is Not Established
Nvidia’s filings measure billing, ordering and headquarters locations, and shipments to Singapore only for fiscal 2025 and only as a ceiling, so the $21,074 million is a minimum for that year. None of the measures shows where every chip was finally used, and nothing in the filings ties the Singapore line to any investigation or customer.
The four court cases are allegations that have not been proven; in each release, the people charged are exporters and intermediaries, not chip makers. The remote-access rule is reported, not published, and the Remote Access Security Act is not law. We have not named the people charged, and we did not contact Nvidia, the Commerce Department or the governments cited.
This Week In The AI Supercycle
The Great Convergence of Capital, Power, and Geopolitics
This week’s issue: for two years the AI build-out was priced as a supply story, and last week as a credit story. This week it was priced as a permission story, and the answer arrived as a cost, not a ban.
US chip export controls license the chip by its performance and by where it goes and who uses it. Nvidia billed $51,976 million to Singapore over three and a half fiscal years, by our arithmetic, while saying its products were almost always shipped elsewhere; prosecutors allege transit through Malaysia, Singapore and Thailand, and the dedicated remote-access rule is reported, not published, while the House-passed bill is not law.
A note on sourcing. We read the Bureau of Industry and Security’s rules of October 2022, October 2023, January 2025 and January 2026 in the Federal Register, its 13 May 2025 release and diversion guidance, Nvidia’s fiscal 2025 10-K and its 10-Q for the quarter to July 2025 on SEC EDGAR, Malaysia’s trade ministry statement of 14 July 2025, the opening statement of the Singapore minister’s 3 March 2025 media conference and four Justice Department releases.
The court cases are charges, and we do not name the people accused. The remote-access rule, the Reuters report and Malaysia’s proposed law come from our earlier pieces linked above. Shares, totals, multiples, day counts and the count of cases are ours. We did not contact Nvidia, the Commerce Department or the governments cited.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
