Timeline
On 16 November 2026 the Bureau of Industry and Security (BIS) issued a rule. BIS is the US Commerce Department agency that runs export controls. The rule targets a gap that has existed since Washington first restricted AI chip sales to China in 2022. Chinese companies could not legally buy Nvidia's most powerful chips. They could, however, rent computing time on those chips from data centres in Malaysia, Singapore and elsewhere.
The new rule requires a US licence before advanced chips are supplied to any data centre, anywhere, that sells large amounts of computing power to companies headquartered in China or other embargoed countries. It also puts chip shipments to Indonesia, Malaysia, Singapore, Thailand and Brazil under licensing. It describes them as "countries which present a higher-than-normal risk of diversion."
Each earlier round of US controls closed one route to the chips and left another open. This is the first to treat renting the computing power as equivalent to receiving the hardware. Whether it works depends on enforcement. Politico, citing unnamed sources, reported that the department that must enforce it did not want to write it. The rule landed four days after Jensen Huang resigned as Nvidia's chief executive. It also came as President Donald Trump flew to Asia for a meeting with Xi Jinping.
What the rule actually restricts
In plain terms, the rule makes selling time on powerful chips to a Chinese-headquartered company legally similar to shipping those chips to China. BIS described its purpose this way: "The rule is designed to ensure that existing restrictions on access to sensitive American technology cannot be circumvented through remote computing arrangements involving data centers in third countries".
It has two main parts, according to the BIS announcement and analyses of the rule text:
- A data-centre control. Suppliers need a licence to provide advanced chips or servers to a data centre if they know it will give more than 12 million units of "aggregate TPP" to a company headquartered in China, Macau or another arms-embargoed country, or whose ultimate parent is. Capacity across linked facilities counts together.
- Licensing for five countries. Exports, re-exports and in-country transfers of advanced chips to Indonesia, Malaysia, Singapore, Thailand and Brazil need a licence, reviewed case by case. Companies headquartered in the US are exempt if exporters file monthly reports.
- Carve-outs and clean-up. Systems used only to rank or recommend content are exempt, as online retailers and social media platforms do. The rule also formally rescinds the Biden administration's 2025 "AI Diffusion" rule.
TPP, or total processing performance, is the measure BIS uses for a chip's raw computing throughput. One industry analysis of the text estimated that 12 million TPP equals about 758 of Nvidia's H100 or H200 chips, or about 222 of its newer B300s. BIS has not confirmed that figure.
Reports on timing differ because the rule runs two clocks. It takes effect on publication in the Federal Register. Politico and others reported that this would be Thursday 19 November. Data centres already serving covered customers then have 90 days to obtain a licence. Public comments are due 60 days after publication.
The authorship is unusual. BIS said the rule was written "by and at the direction of" the Treasury and the National Security Council. Chris McGuire is a senior fellow at the Council on Foreign Relations and a former Biden administration adviser. He wrote on X: "I have never seen a BIS rule that was not written by BIS".
How it compares with four years of chip controls
The table sets out the main US measures on AI chips and China since 2022, and the route each left open.
Measure | Date | What it restricted | Gap left open |
|---|---|---|---|
First China chip rule | 7 Oct 2022 | Advanced chips and chipmaking tools to China | Chips just below thresholds; third countries; cloud access |
Updated chip rule | 17 Oct 2023 | Wider thresholds; 22 embargoed countries plus Macau; Chinese-headquartered firms anywhere | Renting compute from non-US data centres |
Cloud customer-checks proposal | 29 Jan 2024 | Would require US cloud firms to verify foreign customers | US providers only; proposal, no effective date set |
AI Diffusion rule | 13 Jan 2025 | Country caps on chips worldwide; AI model-weight controls | Never enforced |
Diffusion rule dropped | 13 May 2025 | Guidance against using US chips to train Chinese models | Guidance only; replacement delayed |
H200 sales allowed | Dec 2025 | Loosening: H200s to approved Chinese buyers, 25% to US | Blackwell barred from China, but rentable abroad |
Headquarters guidance | 31 May 2026 | Licences enforced for Chinese firms' foreign subsidiaries | Data centres not required to cut off service |
Remote-access rule | 16 Nov 2026 | Data centres serving China-headquartered firms; five countries | Activity inside US; use below threshold; recommender systems |
Sources: Bureau of Industry and Security, Sidley Austin, The Register, CNBC, NPR, Fortune.
The pattern is clear from the right-hand column. The October 2023 update already required licences for chips sold to Chinese-headquartered companies wherever they were located. But that covered buying chips, not renting them. A January 2024 proposal would have made American cloud providers check their foreign customers. It did not reach non-US operators.
The diffusion rule tried a different approach: caps on how many chips each country could receive. The Trump administration announced in May 2025 that it would not enforce it. Under Secretary Jeffrey Kessler called it part of the previous administration's "ill-conceived and counterproductive" policies. BIS promised a replacement. It took 18 months to arrive.
Meanwhile, policy on older chips loosened. In December 2025 Trump said Nvidia could sell its H200 to approved Chinese customers, with 25 per cent going to the US government. Blackwell chips stayed barred. Liza Tobin is a former National Security Council China director. She told PBS that Chinese developers could still reach such chips remotely, leaving a gap between stated policy and practice.

An Nvidia H100 PCIe accelerator card, the class of AI chip at the centre of the rule, photographed in 2023. Photo: 极客湾Geekerwan / Wikimedia Commons, CC BY 3.0.
Why the gap stayed open, and why it closed now
The rule followed reports about Megaspeed International. It is a Singapore-based company that rents computing power from data centres including facilities in Malaysia. On 8 November The New York Times reported that Jensen Huang had a relationship with Alice Huang, a former Megaspeed executive. It also reported that some US officials suspected she was a Chinese intelligence asset. The two are not related. Jensen Huang resigned on 12 November after reportedly admitting an affair.
Beijing has called the spying allegations "entirely groundless, speculative and intended to smear and malign China". Alice Huang has not commented publicly. Megaspeed says she left the company in 2025. It says it has hired outside counsel for an internal investigation. "Megaspeed conducts legitimate and lawful business," it said.
Reported figures on Megaspeed vary. Coverage cited reports that Megaspeed and its Malaysian affiliate, Speedmatrix, bought about $4.8bn of Nvidia hardware over three years. Bloomberg reported that documents it reviewed showed Alibaba accessing Nvidia chips in Malaysia through Megaspeed. It said the access ran via a Singaporean shell company controlled from the Cayman Islands. None of these claims has been tested in court.
The Megaspeed reports exposed a split inside the administration:
- Treasury Secretary Scott Bessent said four agencies had wanted Megaspeed on the Entity List, the US trade blacklist. He said: "Commerce has the pen and BIS has the pen."
- Senator Maria Cantwell alleged to Commerce Secretary Howard Lutnick that his department had received a classified warning about Megaspeed in May 2026 and had not acted.
- Politico, citing three unnamed sources, reported that Bessent pushed the rule through over Commerce objections. Commerce did not respond to Politico's request for comment.
Trump has defended Lutnick and said he has confidence in him. Before leaving for Asia, he described the chips sold to China as close to obsolete. He added: "I don't think you want to insult China just before you get there."
The Asian hubs in the firing line
The five named countries have built data-centre industries partly on demand from Chinese technology companies. The table shows the exposure of each, as reported.
Hub | Data-centre exposure | Share moves, 17 Nov | Government position |
|---|---|---|---|
Malaysia | RM385.7bn (about $95bn) investment approved since 2021 | YTL Power down nearly 10%; YTL Corp down 8% | Own permit rule for re-exports since July 2025; probing Speedmatrix |
Singapore | Megaspeed's base; 2026 growth forecast raised to 4.5–5.5% | Frencken down 3.91%; AEM down 2.9% | Studying rule; seeking "trusted access" |
Thailand | 42 data centres; 3,400 MW; about 750bn baht (about $22.6bn) | none reported | Large new sites limited to industrial zones |
Indonesia | Market valued at $2.81bn in 2025 | none reported | not disclosed |
Brazil | not disclosed | none reported | Operators must reserve 10% of capacity for local users |
Sources: Reuters, The Register, Malay Mail, CNA.
Malaysia carries the largest exposure by investment, and its stocks fell hardest. Joshua Kurlantzick of the Council on Foreign Relations told Fortune: "For Malaysia, the controls could be disastrous." Malaysia's trade ministry said on 13 November that it had never cleared Megaspeed or Speedmatrix. It added that Malaysia does not enforce other countries' unilateral export controls. It expects companies to take them into account.
Singapore's trade and digital ministries said they were engaging US authorities and "exploring how to facilitate trusted access for Singapore companies". The rule creates two tiers there. American cloud providers such as Amazon Web Services, Microsoft and Google can keep receiving chips under the reporting exemption. Local and other foreign operators need licences. Fortune noted that the remote-access limits still apply to the US firms.
The diplomacy is awkward. Singapore was a founding signatory of Pax Silica, the US-led AI supply-chain pact launched in December 2025. It also joined a new protection framework with the US and eight other partners over the weekend before the rule. Trump was expected to press South-east Asian leaders in Manila to sign up to Pax Silica.
Markets had other reasons to fall. Nvidia's shares dropped 4.12 per cent on 16 November. They had lost more than 30 per cent of their value over the month. Hemera reported on 16 November that a wider chip sell-off was already hitting Asian markets.
Will it work?
The stakes are significant. The Institute for Progress, a Washington think tank, has estimated that 30 per cent or more of China's advanced AI computing power in 2027 could come from renting overseas data centres and smuggling chips into South-east Asia. Several limits are already visible:
- The threshold. If the 758-chip estimate is right, a Chinese customer renting fewer H100-class chips from one provider would fall below the line. Linked facilities are still counted together.
- Ownership checks. Enforcement depends on suppliers knowing who the end customer is. Data centres will have to audit clients' direct and indirect owners. Bloomberg's reporting on Megaspeed illustrates how shell structures could obscure them.
- Remaining loopholes. McGuire called the rule "a major step forward" but said gaps remain. The chief one is activity inside the United States, which he said needs a follow-up rule.
- The enforcer. McGuire told Politico that Commerce "doesn't want to do this at all". Christopher Padilla, a former Commerce official, said the public disagreement could give China openings for workarounds.
- Legal footing. The House passed the Remote Access Security Act on 12 January 2026 by 369 votes to 22. It would write remote access explicitly into export-control law. The Senate has not acted.
China also has alternatives. Beijing has largely withheld permission for its companies to import Nvidia chips. It is pushing them towards domestic suppliers such as Huawei. Neil Shah of Counterpoint Research told The Register he expects a short-term computing bottleneck for Chinese firms. He expects a longer-term shift to Huawei's Ascend chips.
Analyst Hanna Dohmen suggested to Fortune that losing overseas access could make roughly 600,000 H200s, reportedly sitting in Hong Kong, more attractive to Chinese buyers. The figures suggest the rule may raise China's costs and slow its access rather than cut it off.
What happens next
- 19 November: scheduled Federal Register publication, when the rule takes effect. Trump is due to meet Xi in Shenzhen the same day, at the APEC summit.
- Within 90 days of publication: data centres already serving Chinese-headquartered customers must obtain licences or stop. Comments close after 60 days.
- 10 December: Lutnick is due before the Senate Banking Committee. The House Foreign Affairs Committee has asked BIS head Jeffrey Kessler to appear. Huang has agreed to answer congressional questions in private, with no date set.
- Ongoing: Singapore and Malaysia are pursuing their own investigations into Megaspeed and related firms. Singapore is negotiating with Washington over "trusted access".
The open question is whether licences are granted freely to data centres that can prove clean customer lists, or used to choke the region's Chinese business. The rule leaves that choice to BIS's case-by-case review. The same department that reportedly resisted writing it will run that review.
Related coverage: Asian Stocks Slide as Chip Selloff Meets Weak Chinese Data · Jensen Huang Misled Nvidia Lawyer Over Reported Relationship, Report Says
Featured image: 2014-04-06-Herbert-C-Hoover-Building by Gunnar Klack, via Wikimedia Commons, CC BY-SA 4.0.




