U.S. Commerce Department Reviewing Chinese Firms' Remote Access to Nvidia Chips in Southeast Asian Data Centers
Chinese companies including Alibaba and ByteDance have rented Nvidia computing power in Malaysia, Thailand and Singapore, a practice current U.S. export rules do not clearly cover.
The Chips Never Left the Building
A Chinese company can't legally buy Nvidia's most advanced AI chip. It can, apparently, rent one for the afternoon.
That's the gap the U.S. Commerce Department is now examining. On August 19, 2026, CNBC reported that the export enforcement arm of the Bureau of Industry and Security is reviewing how Chinese AI firms reach Nvidia hardware sitting in data centers in Malaysia, Thailand and Singapore[1]. Alibaba, ByteDance and Tencent have all been named in reports describing the practice[1]. Chinese engineers log in from abroad, or fly in with hard drives, and train their AI models on machines they were never allowed to own[1][6].
Here's the collision at the center of it: U.S. law bans the sale of these chips to China. It does not clearly ban renting time on them. Both of those things are true at once, and they're why this story exists.
Why a Login Isn't a Shipment
Export control law was built to stop physical goods from crossing borders — chips packed in a crate, put on a plane, landing somewhere they shouldn't[1]. Renting computing power doesn't work that way. No box moves. No border gets crossed. A Chinese company pays a Malaysian data center for access to a machine, and that machine never leaves the building[1].
Commerce has read its own rules to mean that arrangement isn't an export at all. If a cloud provider is just selling a service, not shipping an item, then no export event has happened — and no rule has been broken[1]. That reading is the entire reason the practice has been legal.
Washington has already closed one nearby gap. On May 31, 2026, Commerce issued guidance saying license requirements follow the buyer's ultimate parent company, not the country where a subsidiary happens to be registered[2][3]. That means a Malaysian shell company owned by a Chinese firm can no longer just buy the chips outright and ship them home[2][3]. But renting is a different transaction, and it isn't clearly covered.
Congress has a bill that would close that piece too. The Remote Access Security Act would let the U.S. restrict remote access to controlled items, the same way it restricts physical exports[11]. As of August 19, 2026, it remains a bill. It has not become law[11]. So the honest description of where things stand is: no rule bans this yet, and the review is a first step toward possibly writing one.
Not All Nvidia Chips Are Even Banned
It helps to know which chips are actually at issue, because not all of Nvidia's China-restricted lineup is treated the same way. A rule that took effect January 15, 2026 moved chips like the H200 from a near-automatic denial for China and Macau to case-by-case review, so long as the seller follows know-your-customer and remote-access safeguards[4][5]. That's a real loosening, not a ban.
The chips at the center of the Southeast Asia rental story are different: Nvidia's newer, top-end Blackwell processors, which remain restricted outright[7]. One trade-press report, not confirmed by ByteDance itself, describes the company arranging access to roughly 500 Blackwell systems in Malaysia — about 36,000 individual chips[7]. That's the scale investigators are looking at.
The Justice Department Cases Are a Different Crime
U.S. security hawks in Congress don't see a paperwork gap. They see the same outcome — a Chinese firm training frontier AI on banned hardware — achieved by a workaround instead of a purchase[12][13]. Senators Banks and Warren wrote to Commerce Secretary Lutnick urging the agency to suspend or reconsider some of Nvidia's export licenses, citing a Justice Department indictment over Supermicro chips diverted through Malaysia and Thailand[13].
It's worth separating that criminal case from the rental story, because they're not the same activity. The diversion cases involve chips illegally smuggled across borders — an actual crime under current law[12][13]. Legal cloud rental involves no border crossing at all. Hawks argue the two produce the same military and AI capability for China either way, which is why they want the rules to treat access, not just possession, as the thing being controlled[11]. Nvidia CEO Jensen Huang has publicly said there's no evidence of chip diversion and that the company's market share in China has fallen to zero[12].
Nvidia's Case: The Rules Already Backfired
Nvidia's argument runs the opposite direction. Huang has said publicly that export controls have already pushed Chinese developers toward homegrown chips and open-source Chinese AI models, which the rest of the world is now building on[10]. In that telling, tighter U.S. rules don't slow China down — they hand Chinese firms a captive market and erode America's lead in the software layer that runs on top of the hardware.
Nvidia also raises a practical objection: policing who logs into a foreign data center is a much harder ask than policing what crosses a border. A rule reaching "remote access" would require American and allied cloud operators to verify the ownership and nationality of every customer signing in — a burden Chinese-run cloud services wouldn't face[1]. The company discloses the licensing regime as a material risk to its business in its own SEC filings[16], a reminder that Nvidia's revenue and Washington's policy goals are pulling in opposite directions almost by design.
Malaysia and Singapore Didn't Sit Still
The countries actually hosting these data centers reject the idea that they're a "loophole." Malaysia introduced a Strategic Trade Permit and a 30-day advance notice requirement for U.S.-origin AI chip exports back in July 2025, after Washington pushed it to start tracking Nvidia shipments[14]. Singapore has prosecuted server-fraud cases, and in July 2026 froze about $42 million in assets tied to a local reseller accused of misleading U.S. suppliers about where servers were really headed[15].
Regional commentary, including from the China-Global South Project, frames Southeast Asia as a "hinge" the global chip supply chain depends on — not a weak link that let something through[15]. The structural squeeze on these governments is real regardless of framing: they want the data-center investment and jobs that both American and Chinese money bring, and every new U.S. rule makes hosting Chinese customers riskier without making determined buyers give up[14][15].
Coverage of this story splits in ways that track those interests. CNBC's own headline calls the gap a "crucial loophole," a phrase that treats Washington as already closing something, even though no remote-access rule yet exists[1]. Tom's Hardware describes the leases as "skirting" U.S. restrictions, a verb that implies evasion of a rule the conduct may not actually violate[6]. The Diplomat argues in its own headline that expanding controls to remote access "may backfire" on U.S. ambitions, arguing the industry's enforceability concerns as settled fact rather than one side's case[10]. Al Jazeera stuck to attributed language — "US says" — throughout its coverage of the May guidance[3].
What nobody disputes is where things stand today. The chips are already installed in Johor and Bangkok, and they aren't going anywhere[6][8]. Whatever Commerce decides next will have to catch up to hardware that's already running.
Summary
Chinese technology companies have been getting Nvidia's most advanced AI chips without importing them. Instead of buying the hardware, they rent time on it. The chips sit in data centers in Malaysia, Thailand and Singapore. Chinese engineers log in remotely, or fly in hard drives, and train their AI models there[1][6]. Alibaba, ByteDance and Tencent have all been named in reports on the practice[1]. On August 19, 2026, CNBC reported that the enforcement arm of the U.S. Commerce Department is reviewing how Chinese AI firms reach Nvidia hardware overseas[1].
The legal question is narrow but important. U.S. export controls govern the shipment of physical items across borders. Renting computing power is a service, not a shipment. Commerce has taken the position that a cloud provider selling compute is not an 'exporter,' so no export event occurs[1]. That reading is why the practice has been legal. Washington has already closed a different gap: on May 31, 2026, Commerce said license requirements follow the buyer's ultimate parent company, so a Singapore or Malaysia subsidiary of a Chinese firm can no longer simply buy the chips outright[2][3].
The sides disagree about what the dispute is really about. U.S. security hawks say the point of the rules is to slow China's military and AI capability, and that renting the same chips achieves the same result as buying them[12][13]. Nvidia and much of the chip industry say the controls are already pushing customers toward Chinese alternatives, and that policing who logs into a foreign data center is close to unenforceable[10]. Southeast Asian governments say they are being asked to police American rules on their own soil, at a cost to their own data-center industries[14][15].
One thing to keep straight: not all Nvidia AI chips are banned from China. A rule effective January 15, 2026 moved H200-class chips from near-automatic denial to case-by-case review for China and Macau[4][5]. The chips at issue in the rental story are the newer, top-end Blackwell processors, which remain restricted[7].
The Event
On August 19, 2026, CNBC reported that the export enforcement arm of the U.S. Commerce Department's Bureau of Industry and Security is reviewing how Chinese AI companies access Nvidia hardware located outside China[1]. Reports have named ByteDance, Alibaba and Tencent as having used compute capacity in Malaysia, Thailand, Singapore and Japan to train AI models[1][6]. The review follows guidance BIS issued on May 31, 2026, which extended export license requirements to any buyer whose ultimate parent company is headquartered in China or Macau, regardless of where that buyer is registered[2][3]. No new rule covering remote access had been issued as of August 19, 2026, and the Remote Access Security Act remained pending legislation[11].
Undisputed Facts
- U.S. export controls restrict the shipment of Nvidia's most advanced AI chips to China; the rules operate on physical items crossing borders[4][5].
- On May 31, 2026, the Commerce Department issued guidance stating that license requirements apply based on the ultimate parent company of the buyer, not the country where the subsidiary is registered[2][3].
- A BIS final rule effective January 15, 2026 changed the license review policy for certain advanced computing chips, including the Nvidia H200, from a presumption of denial to case-by-case review for China and Macau, subject to conditions including know-your-customer and remote-access safeguards[4][5].
- The Trump administration rescinded the Biden-era Framework for Artificial Intelligence Diffusion in May 2025; that framework had included trusted-data-center and end-user verification provisions[9].
- Malaysia's Ministry of Investment, Trade and Industry began requiring a Strategic Trade Permit and 30 days' advance notice for exports, transshipments and transit of U.S.-origin advanced AI chips, effective July 14, 2025[14].
- The Remote Access Security Act, which would authorize U.S. restrictions on remote access to controlled items, has been introduced in Congress and has not been enacted[11].
- Senators have written to Nvidia and to Commerce Secretary Howard Lutnick seeking information on export-control compliance and asking that certain Nvidia licenses be reconsidered[12][13].
- Nvidia CEO Jensen Huang has publicly stated there is no evidence of AI chip diversion and that Nvidia's market share in China has fallen to zero[12].
The Pressure
Strip away the moralizing and blame. What structural realities persist regardless of which narrative wins?
- Compute is the scarce input
- Training a frontier AI model needs tens of thousands of top-end chips running for weeks. Nothing else in the stack is as hard to substitute. That is why the U.S. controls it and why Chinese firms will pay to reach it by any lawful route[6][7].
- The law follows goods, not services
- Export control law was built for things in crates. A rented login moves no crate. Commerce's own reading has been that a cloud provider is not an exporter[1]. Changing that likely needs Congress, which is why RASA exists and why the agency has moved on ownership rules it can already enforce[2][11].
- Nvidia's revenue and Washington's policy point opposite ways
- Every restriction shrinks Nvidia's largest potential market and gives Chinese chipmakers a captive customer base. That tension is disclosed in Nvidia's filings and is the honest reason its public arguments run against tightening[16].
- Host countries are caught in the middle
- Malaysia and Singapore want the data-center investment and the U.S. relationship at once. They have added permits and prosecuted cases, but every added rule makes their capacity less attractive to the Chinese customers who helped fill it[14][15].
Material realityThe chips exist, they are installed in Southeast Asia, and physics does not care who logs in. Restricting remote access requires verifying customer identity and ownership at the data-center level, in countries where U.S. agencies have no direct enforcement power. That is achievable only through the host governments' own permit systems, which Malaysia and Singapore have been building since 2025[14][15]. Meanwhile the underlying capability gap is narrowing on its own: Chinese domestic accelerators are improving, and Chinese open-weight models are widely used, so a rule that lands in 2027 governs a different world than the one that produced the 2022 controls[10]. Whatever Washington decides, the compute already sitting in Johor and Bangkok stays there.
Narrative as a weaponThree groups are actively shaping how this reads. Security hawks in Congress want you to see a loophole being exploited, so they foreground criminal diversion cases involving physical smuggling — real cases, but a different activity than legal cloud rental, and blending the two makes the case for a rule feel more urgent. Nvidia and the chip industry want you to see a policy that has already failed on its own terms, so they emphasize lost market share and the rise of Chinese open-source AI, which is true and also happens to argue for selling more chips. Chinese and regional commentary wants you to see American rule-making reaching into other countries' economies, which is a fair description of what ownership-based and access-based controls do, and also conveniently skips why Washington started. The one fact all three tend to blur: as of August 19, 2026, no U.S. rule restricts remote access to chips abroad. The story is about a review and a pending bill, not a change already made.
How Each Side Sees It
Each major actor’s view — how it frames things, its underlying incentive, and how it’s materially affected. Tap a side to read it.
Frames it asTheir argument is about effect, not paperwork. The purpose of the controls is to deny China the compute needed to train frontier AI models with military and surveillance uses. If a Chinese firm can rent the exact same Blackwell cluster it is barred from buying, the rule has been satisfied on paper and defeated in fact. They also point to the criminal record, not just theory: the Justice Department has brought diversion cases involving Nvidia hardware routed through Malaysia and Thailand, and senators have cited attempted exports of H100 and H200 chips and diverted servers in pressing Commerce to act[12][13]. Their second argument is precedent — controls on technology have long covered access, not just possession, and treating a login as untouchable would let any control be dissolved by renting instead of buying[11].
WhyTo keep a measurable U.S. lead in AI compute and to avoid being blamed later for a gap they were warned about[1][13].
Impact on themPolitically, chip enforcement is one of the few bipartisan issues in Washington; Senators Warren and Banks have signed jointly on it[13]. A visible failure is a hearing risk for Commerce; an aggressive rule is a fight with U.S. industry.
Frames it asNvidia's strongest case is that controls have overshot and are now shrinking the American ecosystem rather than China's. Huang has argued publicly that export controls largely backfired, and that Chinese open-source models are becoming the backbone other developers build on[10]. The company's position is that it sells to lawful customers under U.S. licenses and cannot police who later rents time from an independent foreign cloud operator. It also notes the practical problem: a rule reaching 'remote access' would require American and allied cloud firms to verify the nationality and ownership of every customer logging in, a burden Chinese cloud providers would not face. Nvidia discloses the licensing regime as a material business risk in its SEC filings[16].
WhyTo protect the largest addressable market in computing and to keep the world building on CUDA, Nvidia's software layer, rather than on domestic Chinese alternatives[10][16].
Impact on themDirectly exposed. Huang has said Nvidia's China market share fell to zero[12]. Senators have asked Commerce to suspend or reconsider some of its export licenses[13]. Its legal and compliance costs rise with every rule change.
Frames it asTheir position is that renting foreign computing capacity is ordinary, lawful commerce. No controlled item crosses a border. A company in Malaysia sells a service; a company in China buys it. Beijing's broader argument is that the U.S. is using national security language to slow a commercial competitor, and that the shifting rules — banned, then case-by-case, then ownership-based — show the goal is containment, not a stable legal standard[3][15]. They also argue the controls are self-defeating: restricted access has pushed Chinese firms toward domestic chips and open-weight models released free to the world, which erodes U.S. leverage rather than building it[10].
WhyTo keep training frontier models on the best available hardware while domestic chip production catches up[6][8].
Impact on themReal and immediate. If remote access is restricted, training runs move onto slower domestic silicon. ByteDance was reported to be arranging access to roughly 500 Nvidia Blackwell systems in Malaysia, around 36,000 B200 chips[7]. That report is from trade press and has not been confirmed by the company.
Frames it asThey argue they are sovereign states, not enforcement subcontractors, and that they have in fact tightened up. Malaysia imposed a Strategic Trade Permit and a 30-day notification rule on U.S.-origin AI chips in July 2025 after Washington asked it to track every Nvidia shipment[14]. Singapore has prosecuted server-fraud cases and, in July 2026, froze about $42 million in assets tied to a local reseller accused of misleading U.S. suppliers about the true recipient of servers[15]. Their stronger structural point: the region is a hinge of the global chip supply chain, not a weak link, and rules that make hosting compute legally risky will push data-center investment elsewhere without stopping determined buyers[15].
WhyTo attract data-center investment from both the U.S. and China without being forced to pick a side[15].
Impact on themData centers are a major growth industry in Malaysia and Singapore. Tighter U.S. rules raise compliance costs and can strand capacity built for Chinese customers[8][14].
Like this article?
The Bias Ledger average rating 4.1
The same story, as framed by outlets across the spectrum, ordered least to most biased. The bias score (1 = straight, 10 = heavily spun) is an AI assessment of that framing — click an outlet to see its track record. The tell is the word choice or omission that reveals the angle.
| Outlet | Vantage | Bias | How they frame it | The tell |
|---|---|---|---|---|
| Al Jazeera | Qatari state-funded | 2 | 'US says ban on AI chip shipments applies to Chinese firms outside China.' | Attributes throughout ('US says') and does not adopt the loophole frame; the omission is any account of what Chinese firms were doing that prompted the guidance. |
| CNBC | U.S. center, business press | 3 | 'The U.S. banned Nvidia's best chips from going to China. Now it's trying to close a crucial loophole.' | Calls the gap 'crucial' and frames Washington as already acting, which understates that no remote-access rule exists yet and that RASA is still only a bill. |
| Tom's Hardware | U.S. tech trade press | 4 | Says Alibaba and ByteDance 'allegedly' train Qwen and Doubao on Nvidia chips despite controls, via Southeast Asian leases that 'skirt' U.S. restrictions. | 'Skirt' implies evasion of a rule, when the reported conduct is legal under the current reading; the hedge 'allegedly' does the work the verb undoes. |
| China-Global South Project | Independent outlet covering China's engagement with the Global South; sympathetic to non-Western vantage points | 4 | 'Southeast Asia Is the Chip War's Hinge, Not Its Weak Link.' | Frames the region as a strategic partner both powers need rather than a security gap to be closed, and leads with regional agency (permit regimes, prosecutions) before any U.S. security rationale. |
| Asia Times | Hong Kong-based, Asia-focused business | 5 | 'Nvidia GPU crackdown hits China-linked Southeast Asia data centers.' | Centers the harm to regional operators rather than the security rationale; 'crackdown' casts U.S. policy as the disruptive actor. |
| The Decoder | German AI trade publication | 5 | Says ByteDance 'secures access to Nvidia Blackwell cluster in Malaysia, circumventing US export ban on China.' | States 'circumventing' as fact and reports a specific chip count sourced to trade reporting, without noting the company has not confirmed it. |
| The Diplomat | U.S.-based Asia-Pacific policy magazine, analysis | 6 | 'Expanding Export Control to Remote Access May Backfire on US AI Ambitions.' | Leads with the conclusion in the headline. Argues the industry case well but treats enforceability problems as decisive rather than as one side's contention. |
References
- The U.S. banned Nvidia's best chips from going to China. Now it's trying to close a crucial loophole — CNBC · U.S. business news network owned by Comcast/NBCUniversal; market-oriented, center
- U.S. takes step to halt Nvidia AI chip shipments to Chinese firms outside China — CNBC · U.S. business news network owned by Comcast/NBCUniversal; market-oriented, center
- US says ban on AI chip shipments applies to Chinese firms outside China — Al Jazeera · Funded by the government of Qatar; generally critical of U.S. foreign policy framing
- Revision to License Review Policy for Advanced Computing Commodities — Federal Register · U.S. government primary source; official record of agency rulemaking
- Department of Commerce Revises License Review Policy for Semiconductors Exported to China — Bureau of Industry and Security · U.S. Commerce Department agency; primary source, presents its own policy favorably
- China's top AI firms shift model training overseas to access Nvidia GPUs — Tom's Hardware · U.S. consumer/enterprise hardware trade press owned by Future plc
- ByteDance secures access to Nvidia Blackwell cluster in Malaysia, circumventing US export ban on China — The Decoder · German independent AI-industry trade publication
- Nvidia chip curbs turn Singapore into AI hub for China — Asia Times · Hong Kong-based business and geopolitics outlet; Asia-centric, often skeptical of U.S. controls
- Nvidia GPU crackdown hits China-linked Southeast Asia data centers — Asia Times · Hong Kong-based business and geopolitics outlet; Asia-centric, often skeptical of U.S. controls
- Expanding Export Control to 'Remote Access' May Backfire on US AI Ambitions — The Diplomat · U.S.-based Asia-Pacific policy magazine; analysis, not straight news
- Remote Access or Remote Possibility? RASA and the Future of Cloud Export Controls — Freshfields · International corporate law firm; advises exporters and tech clients on compliance
- Letter to Nvidia regarding export control compliance, June 1, 2026 — U.S. Senate Committee on Banking, Housing, and Urban Affairs · Primary source; document authored by the committee's Democratic minority staff
- Senators Banks and Warren Urge Commerce Secretary Lutnick to Suspend or Reconsider NVIDIA Export Licenses Amid Supermicro Chip Diversion Indictment — U.S. Senate Committee on Banking, Housing, and Urban Affairs · Primary source; bipartisan senators' press release, advocacy for a policy position
- Malaysian gov't introduces export license and 30-day notification requirement for export of US AI chips — DataCenterDynamics · UK-based data-center industry trade publication
- Southeast Asia Is the Chip War's Hinge, Not Its Weak Link — The China-Global South Project · Independent outlet covering China's engagement with the Global South; sympathetic to non-Western vantage points
- NVIDIA Corporation Form 10-Q, FY2026 (quarter ended April 26, 2026) — U.S. Securities and Exchange Commission · Primary source; company-authored filing subject to federal disclosure rules