Skip to content
TEN Brief Ten verified stories a day 2026.11.18 KO

이 기사는 한국어로도 읽을 수 있습니다 →

Tech · 20 min read · Breaking

AI chip remote access rule — US licenses cloud rentals, five countries need permits

The interim final rule that the US Commerce Department's Bureau of Industry and Security (BIS) released on November 16, 2026 (US time) closes the route by which Chinese companies rented advanced AI chips in third-country data centres through the cloud. It has two parts. First, a licence is required to send chips to any data centre, anywhere, that provides more than 12 million aggregate TPP of compute to a company headquartered in China, Macau or another arms-embargoed country, or whose ultimate parent is. That is about 760 Nvidia H100s. Second, advanced chips going to Indonesia, Malaysia, Singapore, Thailand or Brazil need a licence whatever the end use. US-headquartered end users are exempt if the exporter reports monthly. The rule takes effect on Federal Register publication, scheduled for November 19. Comments run 60 days and existing data centres get 90 days. South Korea is not among the five

A data center campus lined with palm trees in Southeast Asia, with a truck and wooden crates at the loading dock and two workers in safety vests seen from behind

The three lines

  • What it requires — a licence for chips going to any data centre worldwide that gives China-linked companies more than 12 million aggregate TPP, and a destination-based licence for advanced chips to four Southeast Asian countries and Brazil (released November 16, 2026, publication scheduled for November 19)
  • What changed — earlier controls tracked the physical movement of chips. This one asks who uses the compute, wherever the chip sits. The same document formally rescinds the Biden administration's AI Diffusion Rule
  • Who is outside the list but still touched — South Korea, Japan, Europe and others face no new destination licence, but the data-centre test is global and the rule tightens checks on foundries and packagers. No official position from Seoul or from Samsung or SK hynix was found

Key questions

what is the ai chip remote access rule
**It is an interim final rule the US Commerce Department released on November 16, 2026 that puts cloud rental of US advanced AI chips to Chinese companies, through data centres in third countries, under the same kind of licensing as exports.** | Item | Detail | |---|---| | Released | Monday, November 16, 2026 (US time), by the Bureau of Industry and Security (BIS) | | Form | Interim final rule. In force first, comments afterwards | | Whose access is restricted | Companies headquartered, or with an ultimate parent, in China, Macau or an arms-embargoed country (Country Group D:5) | | Threshold | More than 12 million aggregate TPP across one data centre or linked data centres | | Effective | On Federal Register publication, scheduled for November 19, 2026 | | Comment period | 60 days after publication | | Existing data centres | 90 days from the effective date to obtain a licence | | Exclusion | Compute used only for commercial recommender systems. Generative AI is not excluded |
which countries need a license for ai chips
**Indonesia, Malaysia, Singapore, Thailand and Brazil. Advanced AI chips and servers going to these five now need a US Commerce Department licence regardless of end use. It is case-by-case review, not a ban.** | Country | Reported background | |---|---| | Singapore | Home of Megaspeed, the cloud company under US criminal investigation. Long suspected as a transit point for servers | | Malaysia | A leading data-centre hub built on demand from Alibaba and ByteDance. Several customs seizures in 2026 | | Thailand | The region's other main hub. Smuggling suspicions have been reported | | Indonesia | An early-stage market. An investigation into a local Megaspeed subsidiary is under way | | Brazil | An emerging data-centre market drawing Chinese tech investment | The only official reason Commerce gave is that all five present a higher-than-normal risk of diversion.
does the ai chip remote access rule apply to south korea
**South Korea is not one of the five licensed destinations. But the data-centre test applies worldwide, and the rule's foundry and packaging provisions and its effect on GPU demand reach Korean companies indirectly.** | Area | Where the rule bites | Status | |---|---|---| | Chip exports to Korea | Five-country licence requirement | Does not apply | | Korean cloud and data centres | A licence is needed if a facility gives China-linked customers more than 12 million TPP | Operators must check headquarters and ultimate parents | | Foundry and packaging | Chips from designers in restricted countries are presumed to be controlled items | Industry says no immediate effect because it has no such customers | | HBM memory | No provision bans HBM exports | Indirect, through weaker overseas GPU spending | | Government and company positions | Official statements | None found |

Washington has stopped asking only where a chip is shipped and started asking who uses it. On Monday, November 16, 2026 (US time), the Commerce Department's Bureau of Industry and Security (BIS) released an interim final rule controlling remote use of advanced AI chips. Chinese companies have been unable to buy Nvidia's top chips since 2022, but renting the same chips through the cloud from a data centre in Malaysia or Thailand was legal. The rule puts that route under licence.

It has two parts. One is an end-use control aimed at any data centre, anywhere in the world, that provides more than a set amount of compute to China-linked companies. The other requires a licence for advanced chips going to five countries, Indonesia, Malaysia, Singapore, Thailand and Brazil, on the basis of destination alone. Countries off that list are not untouched. The data-centre test applies in Seoul, Tokyo or Frankfurt exactly as it does in Johor, and the rule also tightens the checks that foundries and packaging houses must run on their customers. This piece sets out what the rule requires, why these five countries, how it differs from earlier controls, and where it reaches companies in South Korea, a US ally that makes the memory inside most AI accelerators. The rule text itself could not be opened. What follows is cross-checked from several reports that summarize it.

1. What the rule requires — one test for data centres, one for countries

The rule is titled "Addressing Remote Access to Advanced Semiconductors". An interim final rule takes effect first and collects comments afterwards, the reverse of ordinary rulemaking. According to TechTimes, BIS relied on provisions of the Export Control Reform Act of 2018 to skip prior notice and comment.

ItemDetailConfirmed through
ReleasedMonday, November 16, 2026 (US time)Multiple reports
EffectiveOn Federal Register publication, scheduled for Thursday, November 19, 2026Converge Digest, Politico, Seoul Economic Daily
Comment period60 days after publicationThe Next Web, TechTimes, CASRAI
Existing data centresMust obtain a licence within 90 days of the effective date to keep serving covered customersTechCrunch, The Next Web, Etoday
Shipments in transitGoods already en route under prior eligibility may be delivered within 30 days of publicationTechTimes
Rescinded in the same documentThe Biden administration's AI Diffusion Rule: country tiers, the validated end-user programme for data centres, and controls on model weightsNBC News, The Daily Brief
AuthorshipBIS said the rule was written by and at the direction of the Treasury Department and the National Security CouncilPolitico, NBC News, Seoul Economic Daily

If publication happens on November 19, the comment deadline falls around January 18, 2027 and the licence deadline for existing data centres around February 17, 2027. Those are dates worked out from the scheduled publication, not dates printed in the Federal Register. TechTimes notes that some press reports wrongly said the rule takes effect in 90 days. The 90 days is the grace period for data centres already serving covered customers.

The first part is the data-centre test. It adds a new Section 744.24 to the Export Administration Regulations. An exporter, re-exporter or anyone transferring goods within a country needs a licence if it knows that the advanced chips or servers it supplies are going to a data centre that meets the following description.

ElementDetail
Location of the data centreAnywhere in the world. No country is exempt
Who receives the computeA company headquartered in Macau or in Country Group D:5, the US arms-embargo list that includes China, Russia, Iran and North Korea, or a company whose ultimate parent is headquartered there
ThresholdMore than 12 million aggregate total processing performance (TPP)
AggregationNot one building. All data centres linked so as to function as one system are added together
Indirect useCovered too: a customer outside the restricted group that trains models for a covered company or supplies models to it, and rental through brokers
Change of useChips that needed no licence when first shipped become licensable if they are later turned to this use. The switch is treated as an in-country transfer
RestructuringSpin-offs and corporate restructurings are named as no way around the control
ReviewCase by case

TPP is the unit US export controls use for computing performance, calculated from a chip's operations per second and its bit length. The rule does not say how many chips 12 million TPP amounts to. News outlets did the conversion.

ChipTPP per chipNumber equal to 12 millionConverted by
Nvidia H100, H20015,832About 758The Daily Brief (TechCrunch says about 750)
Nvidia B200About 36,000About 333The Daily Brief
Nvidia B300About 54,000About 222The Daily Brief

Set against training clusters that run to tens of thousands of chips, that is a low line. It leaves small rentals alone and sends anything large enough to train a serious model to the licensing desk.

Not knowing is no longer much of a defence for the seller. The rule tells chip suppliers to obtain from the data centre the information needed to check its customers' direct and indirect ownership and control. If the data centre cannot or will not provide it, that is a red flag, and a supplier that goes ahead without resolving it is presumed to have known. This is Note 3 to Section 744.24, as relayed by TechTimes and The Daily Brief.

The second part is the country test. Advanced computing items going to, or moving within, Indonesia, Malaysia, Singapore, Thailand and Brazil now need a licence regardless of end use. It is not a ban. According to IT Daily and The Daily Brief, applications are reviewed case by case in light of national security and foreign policy interests and of how well each country enforces its own export controls.

Exclusion or exemptionConditionWhat to watch
US-headquartered companiesIf the ultimate consignee or end user is headquartered in the US, or has a US-headquartered ultimate parent, the five-country licence does not applyThe exporter must file monthly reports under Section 743.10: every party and its ultimate parent, each product's export classification, TPP and quantity, by the first business day of the following month
Companies headquartered elsewhereNo exemptionA facility in one of the five countries owned by a Korean, Japanese or European company needs a licence
Commercial recommender systemsCompute designed and used exclusively to rank, filter or suggest products or content is outside the data-centre test, subject to submitting informationA system that generates text, images, video or code is not excluded, even if it also recommends

The US-headquarters carve-out has a plain consequence. Microsoft, Google and Amazon facilities in Singapore or Malaysia keep receiving chips as long as the reports are filed. A local operator or a third-country operator next door waits for Washington's decision each time. The Daily Brief warns that the recommender exclusion reads differently in the preamble and in the regulatory text, and advises against planning around the more generous reading.

A third strand, on foundries and assembly-and-test providers, is taken up in section 4.

2. Why these five — one official line, and a different story in each country

Commerce's stated reason is that the five present a higher-than-normal risk of diversion. TechCrunch, Bloomberg, AFP and The Next Web all carry the same phrase. Whether the rule explains what is wrong in each country is not confirmed. The background in the table below is assembled from press commentary and is not the US government's stated reasoning.

CountryReported backgroundGovernment response as of November 17, 2026
SingaporeNot a large data-centre location, but long suspected as a transit point for servers headed to mainland China. Megaspeed, the cloud company under investigation, is based thereA joint statement from the trade and digital ministries: studying the details, talking to US authorities about scope and implementation, and exploring how to facilitate trusted access for Singapore companies
MalaysiaWith Thailand, one of Southeast Asia's two main data-centre hubs, built largely on demand from Alibaba and ByteDance. Bloomberg reported that Alibaba used Nvidia chips located in Malaysia through MegaspeedDid not immediately respond to Bloomberg
ThailandThe other main hub. Newspim, citing Bloomberg, says US authorities suspect a company tied to a Thai AI project of involvement in server smugglingNone found
IndonesiaAn early-stage market. According to AFP, an investigation into a local Megaspeed subsidiary is under wayThe Government Communication Agency said it is reviewing the measure, seeking industry input and open to technical dialogue, and that a strategic trade management framework is being developed
BrazilAn emerging data-centre market drawing investment from Chinese tech companies. The only one of the five outside Southeast AsiaSaid it is closely monitoring developments and declined to assess the impact

Brazil was the surprise. Latin Times reads the grouping as countries sharing an enforcement weakness, not a political alignment, and says Washington is not accusing Brazil's government of deliberate diversion. TechTimes reports that a draft listed seven countries, with Norway and South Africa also on it. Why those two were dropped has not been explained.

Three of the five are familiar from the smuggling story. As this site's October 30, 2026 explainer on transshipment set out, Malaysia, Singapore and Thailand are the places through which chips barred from direct shipment to China have passed under changed paperwork, and a report in the same edition covered a Malaysian customs seizure that was the fifth in 2026 tied to one syndicate. Transshipment is the problem of moving chips quietly. The new rule addresses the problem of using chips without moving them at all. Both grew in the same region.

Figures for the scale are scarce. Bloomberg cites an earlier estimate by the Institute for Progress, a Washington think tank, that at least 30% of China's advanced AI computing power in 2027 would come from renting overseas data centres and from chips smuggled through Southeast Asia. NBC News says businesses in the five countries bought billions of dollars of high-end chips over the past several years.

The outlook from the region is grim. Robert Walker of the Lowy Institute told Bloomberg the rules would be a major headwind to data-centre investment, with Singapore, Malaysia and Thailand most exposed, though he added that the region's place in semiconductor supply chains is more secure because exports to the US are booming. Joshua Kurlantzick of the Council on Foreign Relations told Fortune that for Malaysia the controls could be disastrous, as relayed by The Next Web. Deborah Elms of the Hinrich Foundation said the timing would make for a very uncomfortable meeting between Donald Trump and Southeast Asian leaders. Trump left Washington on November 16 for ASEAN-related meetings in Manila before heading to Shenzhen.

One sentence in Singapore's statement deserves attention. It says Singapore does not condone companies using their association with Singapore to circumvent other countries' laws. The statement does not dispute the transit-point label, and in the same breath promises to protect legitimate companies' access to chips.

3. How it differs from earlier controls — from moving goods to using compute

Until now, AI chip export controls caught goods at the moment they crossed a border. As this site's August 24, 2026 explainer on AI chip export controls described, the performance line has moved several times since 2022, but the object of control was always which chip could go to which country. Once a chip arrived somewhere permitted, nobody asked who was using its compute over the internet. BIS calls the new rule a logical extension of existing controls.

Earlier controlsAI Diffusion Rule (January 2025, never enforced)This rule (November 2026)
What is watchedPhysical movement of chips and serversChip volumes per country, and model weightsThe headquarters of whoever uses the compute after the chip is installed
GeographyExports and re-exports to China and other countries of concernThe whole world, sorted into three tiers with capsData-centre test worldwide, country test for five
Cloud rental in third countriesNot controlled. LegalPartly managed through a data-centre certification schemeLicensable above 12 million TPP
Customer checksConsignee and end userValidated end usersUltimate parent, and direct and indirect ownership and control
Change of use after exportAn issue only on re-export or transfer-The change of use is itself treated as a transfer
StatusIn forceFormally rescinded by this ruleDue to take effect on publication

The AI Diffusion Rule, issued in the last days of the Biden administration in January 2025, divided the world into tiers and capped chip imports country by country. The Trump administration said in May 2025 that it would not enforce the rule and would rescind it, but never formally did. In the meantime nothing addressed remote access. The new rule drops the wide net of country caps for a narrow target: China-linked users and five transit countries. For most countries, South Korea included, no new quantity limit appears.

The legal footing is the weak point. Export control law governs transfers of goods, software and technology, and renting compute through the cloud moves nothing across a border. The Daily Brief puts it in five words: cloud computing is not an export. The rule therefore takes a side road. It does not license the cloud service. It licenses the export of chips to the data centre that will provide it. The Seoul Economic Daily reports that some lawyers say Commerce lacks authority under current law to regulate remote computing access.

A bill meant to fill that gap is waiting in Congress. The Remote Access Security Act would amend the Export Control Reform Act to name remote access by foreign persons as something Commerce can control. It passed the House 369 to 22 on January 12, 2026 and has not moved in the Senate. The Seoul Economic Daily expects the rule to give the bill momentum and counts at least four related bills in Congress.

Nvidia has been worrying about that bill for some time. CNBC reported on November 17, 2026, citing two people with direct knowledge, that Nvidia warned customers in recent months that the act could expose both Nvidia and them to US legal action. According to the same people, Nvidia's stance for now is that it will not sell chips to customers that rent more than 50% of their GPUs to Chinese companies. Nvidia did not respond to CNBC's request for comment. In a securities filing in August 2026 the company said it aims to comply strictly with export control rules but has no physical control of its products after sale and relies on the compliance programmes of customers and partners. Note 3 of the new rule narrows exactly that reliance: a seller that cannot find out whom its customer rents to is presumed to know.

How the rule came about is unusual as well. Politico, citing three people familiar with the matter, reported that Treasury Secretary Scott Bessent pushed the rule through and that Marco Rubio, the secretary of state and national security adviser, backed it. For BIS to state that a rule issued under its own name was written by and at the direction of Treasury and the National Security Council is rare in itself. Chris McGuire of the Council on Foreign Relations, a technology and security official in the Biden administration, wrote that he had never seen a BIS rule that was not written by BIS. Christopher Padilla, a Commerce official under George W. Bush, said he could not recall an agency publicly distancing itself from a rule issued under its own authority, and warned that the split could give China an opening to look for workarounds.

Behind this is the criticism of Commerce Secretary Howard Lutnick covered in this site's November 13, 2026 report on the congressional probe into Nvidia's China export approvals. Lutnick has been faulted for delaying action against Megaspeed. On Friday, November 13, Bessent said on CNBC that the authority to write such rules sits with Commerce and that his own deputy had been pushing on the issue for months. Speaking on November 16 before leaving for Asia, Trump acknowledged that his cabinet holds two different views on China and said he was not considering replacing Lutnick. Politico later noted that it had changed its headline to remove a reference to Commerce Department objections. Late the same day, the Commerce and Justice departments confirmed in a joint statement a criminal investigation into Megaspeed. According to Al Jazeera, Commerce began examining the company in 2024 and Justice later took the lead.

Jensen Huang, Nvidia's co-founder, stepped down as chief executive on November 12, 2026, and several outlets read the allegations involving a former Megaspeed executive that prompted his exit as the thing that forced this rule out. What has happened since around Huang's position is covered in a separate report in this site's November 18, 2026 edition.

Assessments mostly point one way. McGuire called the rule a major step forward, the first serious attempt in about eighteen months to close gaps in the controls, while noting that gaps remain, including some covering activity inside the United States. As relayed by Asia Economy and the Seoul Economic Daily, his forecast is that if the rule is enforced as written, the business of Southeast Asia-based neoclouds such as Megaspeed is effectively over. He also told Politico it is dangerous to trust Commerce's leadership to enforce it.

4. South Korea — off the list, touched in three places

South Korea is a useful case for any country that is not one of the five. It is a US treaty ally, home to Samsung Electronics and SK hynix, which make the high-bandwidth memory (HBM) that goes into AI accelerators, and to Samsung's contract chipmaking business. It is not on the five-country list and has nothing to do with the arms-embargo group. Nvidia chips entering Korea face no new licence. Korea is also a member of Pax Silica, the US-led AI supply chain grouping this site covered on August 24, 2026. Bloomberg points out that Singapore is in Pax Silica too, and membership did not keep it off the list.

The rule reaches Korea in three places.

AreaWhich part of the ruleLikely burdenBasis
Cloud and data centres in KoreaThe data-centre test (Section 744.24), which ignores locationWhen offering compute above the threshold to foreign customers, operators must establish headquarters, ultimate parent and actual use. Chip suppliers will ask for itAnalysis by Dealsite and IT Daily
Korean-owned facilities in the five countriesThe country test (Section 742.6). Only US-headquartered companies are exemptAdvanced chips bound for a Korean company's data centre in Malaysia or Singapore need a case-by-case licenceInference from the rule's structure. No instance confirmed
Foundry and packagingA tighter presumption about chip designersChips from designers headquartered in restricted countries or the five listed countries are presumed to be controlled high-performance items, with fewer ways to rebut it. More customer vetting for Samsung's foundry and for packaging housesTechTimes, IT Daily, Seoul Economic Daily
HBMNo direct provisionIndirect, if GPU orders for overseas data centres serving Chinese customers shrink. Expansion elsewhere offsets part of itAnalysis by Dealsite

First, cloud. The data-centre test names no country. If a data centre in Korea provides more than 12 million TPP to a company headquartered in China, or to one whose ultimate parent is in China, whoever ships chips to that facility needs a licence. It does not matter whether the contracting entity is a Singapore company or a Cayman one. The ultimate parent is what counts. The Alibaba arrangement Bloomberg described had exactly that shape: a Singaporean shell company, controlled by a Cayman Islands entity, with Alibaba as the ultimate holding company. Whether any Korean operator crosses the threshold is not known. But any operator that expects to keep receiving Nvidia chips will need to be ready to explain the ownership behind its customer list.

Second, foundry. The rule revises a presumption that turns on where a chip's designer is headquartered. According to TechTimes, designers headquartered in the countries-of-concern groups or in the five listed countries can no longer rebut the presumption that their chips are controlled items, and foundries and assembly-and-test providers need a licence to export chips made for them. The Seoul Economic Daily describes this as the revival of a Biden-era measure the Trump administration had shelved, and says Samsung and TSMC now have to screen out orders from Chinese customers placed through shell companies. An unnamed Korean industry official told Asia Economy there would be no immediate effect because Korean firms were not doing this business with Chinese customers anyway, but that the industry is watching the effect on overall chip demand. IT Daily notes that not every foundry or packaging deal needs a new licence. It depends on performance and on the designer's status.

Third, HBM. Nothing in the rule bans exports of HBM. The memory goes into AI accelerators made by Nvidia and others, so any effect arrives through demand for those accelerators. Dealsite's analysis is that weaker GPU investment at overseas data centres built for Chinese customers is a variable for memory demand, but that expansion in other regions and existing long-term supply contracts cushion it, so a GPU restriction does not translate directly into lower HBM demand. Add the situation this site reported on November 9, 2026, in which about 600,000 H200 chips sit in Hong Kong because Beijing has not approved their import, and the paths by which Chinese companies reach Nvidia chips have narrowed from both sides at once: direct import and overseas rental. Analyst Paul Triolo, quoted by Asia Economy, expects Chinese companies cut off from overseas compute to lean harder on domestic chips such as Huawei's Ascend.

Markets reacted broadly on the day. On November 16, 2026 Nvidia fell more than 4% in New York and the Philadelphia Semiconductor Index dropped by more than 5%. GPU rental companies CoreWeave and Nebius lost between 6% and 8%. According to Al Jazeera, South Korea's Kospi fell 2.6% on November 17. The Nvidia scandal and a wider slide in technology shares were at work on the same days, so the rule's share of the move cannot be isolated.

No official position on the rule from South Korea's trade ministry, Samsung Electronics, SK hynix or Korean cloud operators was found.

5. What is still open — China's response, the summit, and enforcement

China responded within a day. On November 17, 2026 its commerce ministry said it strongly opposes the measure and that if the US persists, China will take countermeasures to safeguard its sovereignty, security and development interests. The ministry argued that the curbs go against understandings reached by the two leaders and endanger global chip supply, and it called for the dispute to be settled through talks. At a regular briefing the same day, foreign ministry spokesperson Guo Jiakun said the US had continuously overstretched the concept of national security and was infringing on the legitimate rights of Chinese companies. Neither said what the countermeasures would be.

The timing is delicate. November 19, the day the rule is due in the Federal Register, is also the day Trump and Xi Jinping are due to meet at the Asia-Pacific Economic Cooperation summit in Shenzhen. Lutnick and Bessent are travelling with the president on the same plane. The US-China trade truce that lowered tariffs in 2025 expires in January 2027. Whether the rule turns into a bargaining chip or survives the meeting intact, the first sign will be whether it is published on schedule.

Open questionStatus
Federal Register publication and effective dateReported as scheduled for November 19, 2026. Actual publication not confirmed
Comment deadline and the 90-day deadlineRun from publication. On a November 19 basis, around January 18 and February 17, 2027
China's countermeasuresThreatened only. No content or timing
Effect of the Trump-Xi meetingDue November 19 in Shenzhen. Whether the rule is on the agenda is not known
How a country gets off the listNo published criteria. Singapore is discussing trusted access
Additions to the listThe Commerce Secretary can reportedly add countries. Whether Norway and South Africa return is unknown
How often licences will be grantedNothing published beyond case-by-case review
Scope of the recommender exclusionPreamble and text reportedly differ. Likely to come up in comments
Commerce's will to enforceThe rule was written in one place and will be enforced in another. Former officials have voiced doubt
Remote Access Security ActPassed the House, stalled in the Senate. It bears on the legal-authority question
Data centres inside the USMcGuire says loopholes remain for activity within the United States
South Korean government and companiesNo position found

What came out on November 16, 2026 does not redraw the border for chips. It asks who owns the computation running on them. The line was set low, at roughly 760 H100s. Five countries will be reviewed on destination alone. And sellers now carry responsibility for knowing their customers' customers. Countries such as South Korea sit outside the list. But the question the rule poses, who the final user of this compute is, has to be answered by anyone who brings in Nvidia chips and rents them out, wherever they are.

How chips detour through third countries is covered in "What is transshipment?" and "Malaysia seizes 30 AI chip servers" (both October 30, 2026), and the three years of shifting performance lines in "What AI chip export controls are" (August 24, 2026). The congressional inquiry into Lutnick is in "Nvidia China export probe in Congress" (November 13, 2026), and Beijing's own block on Nvidia imports is in "Nvidia H200 China exports" (November 9, 2026).

Sources

  1. TechCrunch — US tightens AI chip export controls to deter Chinese companies from accessing compute remotely (2026-11-16)
  2. NBC News — U.S. closes loophole that let Chinese companies use top AI chips through foreign data centers (2026-11-16, updated 11-17)
  3. Politico (via Yahoo) — Bessent pushed new AI chips rule (2026-11-17)
  4. CNBC — Nvidia warns chip buyers of legal risks tied to China AI access, sources tell CNBC (2026-11-17)
  5. Al Jazeera — US moves to block China's remote access to AI chips amid Nvidia scandal (2026-11-17)
  6. Bloomberg (via Yahoo News Singapore) — Asia data centre hotbed hit by US chip curbs as Trump visits (2026-11-17)
  7. AFP (via Yahoo News Singapore) — Singapore seeks 'trusted access' to AI chips after new US curbs (2026-11-17)
  8. The Next Web — US curbs China's remote access to AI chips through overseas data centres (2026-11-16)
  9. The Next Web — China warns US of countermeasures over new AI chip remote access curbs (2026-11-17)
  10. TechTimes — AI Chip Export Rule Sets Cloud License Line at Roughly 750 Nvidia H100s for China-Linked Tenants (2026-11-16)
  11. The Daily Brief (Rajesh Beri) — BIS Requires Licenses for AI Chips Shipped to Five Countries (2026-11-16)
  12. Converge Digest — U.S. Targets Remote Access to Advanced AI Chips in Overseas Data Centers (2026-11-16)
  13. CASRAI — BIS issues rule on remote access to AI chips (2026-11-16)
  14. Latin Times — America Closes the AI Cloud Loophole, and Brazil Must Now Ask Washington's Permission to Buy Advanced Chips (2026-11-16)
  15. TokenPost — report on Nvidia's customer warning over the Remote Access Security Act (2026-11-17)
  16. Seoul Economic Daily — rule blocks rerouting of AI chips to China through third-country data centres (Korean, 2026-11-17)
  17. Seoul Economic Daily — renting US chips banned outside China too, Samsung and TSMC within reach (Korean, 2026-11-17)
  18. Asia Economy — US shuts China's AI chip detours, Samsung and TSMC watch the impact (Korean, 2026-11-17)
  19. IT Daily — US also controls change of use after export, tighter foundry and packaging checks (Korean, 2026-11-17)
  20. Dealsite — limits on Chinese firms' overseas GPU use a variable for Samsung and SK hynix HBM demand (Korean, 2026-11-17)
  21. Newspim — US blocks China's access to AI chips through third-country data centres, tightens rules on five countries (Korean, 2026-11-17)
  22. Financial News — US cuts off China's remote access to AI chips (Korean, 2026-11-17)
  23. Etoday — US blocks remote detours for AI chips, tightens export rules for Southeast Asia and Brazil (Korean, 2026-11-17)
  24. Korea Economic Daily — US blocks Chinese firms' indirect use of AI chips (Korean, 2026-11-17)

Verification

Published
Last modified
Cross-check
Checked against 24 independent sources.
Unverified
  • The rule text (Addressing Remote Access to Advanced Semiconductors, RIN 0694-AK50) and the BIS press release were not read directly. No working bis.gov address for the release was found and the Federal Register public-inspection copy did not surface in searches. Section numbers (744.24, 742.6, 743.10), the threshold, the exclusions and the transition periods come from TechTimes, The Daily Brief, Converge Digest, CASRAI and IT Daily, which summarize the text, checked against one another.
  • Reuters' first report of November 16, 2026 could not be opened. Reuters material is used only where another outlet cites it.
  • The effective date is the date of Federal Register publication. The scheduled date of November 19, 2026 is consistent across Converge Digest (citing the public-inspection notice), Politico (effective Thursday) and several Korean outlets. TechTimes and The Daily Brief, however, say the text BIS posted left the publication date, effective date and comment deadline as placeholders. Actual publication had not been confirmed when this was written. The comment deadline of January 18, 2027 and the licence deadline of February 17, 2027 for existing data centres are this site's own arithmetic from a November 19 publication, not dates printed in the Federal Register.
  • The length of the rule is given as 166 pages by the Seoul Economic Daily and 132 pages by Bloomberg. The difference is probably between versions but was not confirmed.
  • How many Nvidia chips equal 12 million TPP is a conversion that does not appear in the rule. TechCrunch says about 750 H100 or H200 chips. The Daily Brief uses 15,832 per H100 or H200 to get about 758, and estimates about 333 B200s and about 222 B300s. The B200 and B300 TPP values come from The Daily Brief alone.
  • That the covered items are those classified under ECCNs 3A090 and 4A090 comes from Latin Times and one secondary report. TechTimes and The Daily Brief name only 3A090. The full list of covered items was not checked against the text.
  • It is not confirmed that Commerce explained the choice of the five countries one by one. The phrase several outlets quote is a higher-than-normal risk of diversion. Al Jazeera renders it as an unacceptable risk of diversion. The country-by-country background in this piece is commentary from Newspim, Bloomberg, AFP and Latin Times, not the rule's stated reasoning.
  • That a draft listed seven countries, including Norway and South Africa, comes from TechTimes and a secondary report citing The Information. Why the two were dropped is not known.
  • That the Commerce Secretary may add countries after consulting the Secretary of State or the national security adviser was found only in Financial News and one secondary report.
  • The Daily Brief and TechTimes say the recommender-system exclusion reads differently in the preamble and in Section 744.24(b)(2). Which reading holds was not checked against the text. The 30-day allowance for shipments already en route comes from TechTimes alone.
  • BIS's estimate of about 100 additional licence applications and about 970 additional reporting and exemption submissions a year comes from TechTimes alone.
  • That Treasury Secretary Scott Bessent pushed the rule through is Politico's reporting from three unnamed people. Politico later added a correction saying it had changed the headline to remove a reference to Commerce Department objections. The White House, Treasury, State and Commerce did not respond to Politico's requests for comment. The account in which Commerce Secretary Howard Lutnick favoured a lighter version while Bessent, Marco Rubio and trade representative Jamieson Greer pressed for broad restoration comes from Asia Economy.
  • Nvidia's warning to customers, and its stance of not selling for now to customers that rent more than 50% of their GPUs to Chinese companies, is CNBC's reporting from two unnamed people, repeated by TokenPost and the Seoul Economic Daily. Nvidia has not published the policy and did not respond to CNBC. The Financial Times report of a verbal 51% instruction was seen only as cited by CNBC. No official Nvidia statement on the rule itself was found.
  • The House vote on the Remote Access Security Act (January 12, 2026, 369 to 22) was confirmed through TechTimes and a secondary source. The House roll call itself was not read. The view that Commerce lacks statutory authority to regulate remote access is attributed to a Baker McKenzie lawyer whose name and original remarks were not confirmed.
  • The Chinese commerce ministry's statement was not read in the original. The wording is as relayed by The Next Web from Bloomberg and Reuters, and The Next Web says it had not seen the statement either. Foreign ministry spokesperson Guo Jiakun's remarks were confirmed through NBC News and the Seoul Economic Daily. China has not announced any specific measure.
  • A reported statement by Malaysia's trade minister that Kuala Lumpur does not intend to negotiate and expects operators to comply, and a reported fall of nearly 20% in some Malaysian data-centre stocks, appeared only in search summaries. The originals were not opened and neither is used in the body. No reaction from the Thai government was found.
  • Market figures for November 16, 2026 differ by outlet. Nvidia's fall is 4.1% in Al Jazeera and more than 4% in Newspim and Asia Economy. The Philadelphia Semiconductor Index fell 5.3% according to Newspim and as much as 5.5% intraday according to The Next Web citing Bloomberg. CoreWeave down more than 6% and Nebius down nearly 8% come from the Seoul Economic Daily alone.
  • No official position from South Korea's trade ministry, Samsung Electronics, SK hynix or Korean cloud operators was found. The statement that there is no immediate effect is from an unnamed industry official quoted by Asia Economy. Whether any data centre in Korea gives China-linked customers compute above the threshold is not known. The effects on HBM and foundry are analysis by Dealsite, IT Daily and the Seoul Economic Daily, not established fact.
  • The estimate that at least 30% of China's advanced AI computing power in 2027 would come from overseas rental and smuggling is from the Institute for Progress as cited by Bloomberg, a single source. The January 10, 2027 end date for the US-China trade truce comes from Asia Economy alone.
Authoring
Reviewed by a person before publication. The full process is described in the Editorial.

Ten stories, once each morning

We send the three-line summaries only; the full pieces stay on the site. One-click unsubscribe, any time.

Related