Nvidia Cuts Its Asia AI Chip List and Concentrates Supply

Nvidia has more than halved the Asian companies cleared to buy its AI chips after a new whitelist knocked most neo-cloud buyers off the list. The July 2026 cut, aimed at Singapore, Malaysia and Japan, followed U.S. pressure to stop Blackwell-class processors reaching Chinese-owned buyers through third countries.

The shorter list did not empty the region of GPUs. It packed remaining supply into a few inspectable operators, while Chinese demand moved toward rental access and freight routes the paperwork still does not fully cover.

A Shorter List for Singapore, Malaysia and Japan

People familiar with the process said Nvidia spent months building a new white list of approved regional buyers, then dropped more than half of its previous customers on the first pass. The reviews ran hardest in Singapore, Malaysia and Japan, the three markets that have shown up most often in chip-diversion cases.

THE CUT AT A GLANCE

• The share: More than half of Nvidia’s previous Asian customers were removed from the approved buyer list.
• The markets: Checks focused on Singapore, Malaysia and Japan.
• The method: Staff visit data centers, verify contracts and interview end users.
• The door back: Firms that failed can make changes and reapply.

Neo-cloud providers, the smaller AI clouds that rent GPUs rather than run giant public-cloud regions, failed in the greatest numbers. Nvidia did not comment on the buyer-list change when it became public on July 14, 2026. The U.S. Commerce Department sat in on the process, people familiar with it said, giving the audits political cover as well as paperwork.

Washington Forced the Ownership Test in May

The whitelist is Nvidia’s operational answer to a May 31, 2026 notice from the Bureau of Industry and Security. That guidance restated a rule first written in November 2023: exporters need a license for any D:5-parented buyer, including a Singapore or Malaysia subsidiary, when the ultimate parent is headquartered in China, Macau or another Country Group D:5 destination.

A license is required to export advanced computing items to entities headquartered in Country Group D:5 or Macau or with an ultimate parent company headquartered in Country Group D:5 or Macau, even if the entities themselves are located outside Country Group D:5 or Macau.

Bureau of Industry and Security, May 31, 2026 guidance

The test shifted from where a crate is addressed to who owns the buyer. A company registered in Singapore can no longer clear a Blackwell sale on local paper if its parent sits in China. Industry estimates circulating in May put the earlier leakage through Chinese-owned subsidiaries in the hundreds of thousands of chips. BIS presented the notice as a clarification, not a new ban, after months of questions about whether the old end-user control still applied once the AI Diffusion Rule stopped being enforced.

On the same July 14 as the whitelist news, Under Secretary of Commerce for Industry and Security Jeffrey Kessler told the House Foreign Affairs Committee that licensed H200 shipments to mainland China and Hong Kong had begun and were still “very few,” a “trivial” quantity against the licenses on file. The legal window into China is a trickle. The gray window through Southeast Asia is what Nvidia was told to shrink.

Why Neo-Clouds Failed the First Review

Smaller regional clouds built businesses on fast access to Nvidia’s best accelerators. The new checks treat that speed as a risk. A Malaysian or Singaporean operator that has done nothing wrong can still fall off the list because of a shareholder it cannot explain, and Nvidia has not published the reapplication bar.

WHY ACCOUNTS GOT CUT

• Parent company: A China or D:5 headquarters now requires a license even when the buyer is incorporated in Singapore, Malaysia or Japan.
• Empty floor space: Inspectors look for a real, running data center, not a lease and a logo.
• End user: Staff interview the people who will run the racks, not only the firm that signs the purchase order.
• Contracts: Paper trails have to match the hardware on the floor.

Singapore and Malaysia as transshipment routes were not a secret inside the trade. The whitelist arrived after years of that traffic, not at the start of it. Cutting the list changes who can place an order. It does not erase Chinese demand, and it turns GPU access into capacity planning for any cloud that is not already on the approved sheet. Parallel supply, including domestic Chinese accelerators and rental of racks that already sit in Southeast Asia, fills the gap the paperwork opens.

Hyperscalers with clean ownership files and U.S. headquarters were never the main target. The pain landed on the middle layer: neo-clouds that grew up as the easy way to rent a cluster without building a full region.

22,000 More GPUs Still Headed to Malaysia

The same company that shortened the buyer list is still putting large GPU counts into Southeast Asia through partners that clear the new bar. On Sept. 22 and 23, 2026, Nvidia ran AI Day in Singapore and used the stage to show who remains inside the tent.

WHO STILL GETS THE CHIPS

Aolani has been an Nvidia Cloud Partner since 2023. Singapore-founded Sea Limited is the first enterprise in the ASEAN region to adopt the NVIDIA Vera Rubin platform, according to Nvidia’s own event recap. The approved names get bigger orders. Everyone else waits on a reapplication whose scoring rules are private.

That split sits on top of a much larger build. An ISEAS paper on Malaysia’s data-center boom credits the country with attracting more than $34 billion in investments from major U.S. and Chinese cloud firms, with Johor going from 10 megawatts to 2.1 gigawatts. The whitelist does not freeze that construction. It decides which of those halls may take the newest Nvidia boards.

The Guidance That Left Racks Running

The May 31 notice closed the subsidiary purchasing trick and, in the next breath, left installed machines in place. BIS told bona fide data-center operators they are not required to stop the ongoing use, storage, disposal or servicing of advanced computing items because of that guidance, until further notice.

Chips that already landed in a Malaysian or Singapore hall can keep running. The remaining path for a restricted customer is not always a crate. It is a login. Remote access to a legal cluster in Southeast Asia was the workaround the ownership test does not, by itself, shut off. Commerce has been circulating a draft that would treat GPU rental to Chinese AI firms through servers in places such as Thailand and Singapore as a violation. Export-control lawyers have questioned whether the department can police network access under rules written for physical goods.

Until that draft becomes an enforceable rule, the whitelist polices who may buy the next shipment. It does not automatically police who may rent the last one.

Apex Logistics and the Freight Turn

Once buyers are on a short list, the next weak link is the people who move the boxes. Enforcement since spring has moved from chip designers and resellers toward landlords, airports and freight forwarders.

THE ENFORCEMENT CALENDAR

1. March 2026: U.S. prosecutors charge a Super Micro Computer co-founder and two employees over an alleged $2.5 billion Nvidia-chip scheme that used a Southeast Asian company as a proxy.
2. June 2026: Malaysian officers seize 72 servers worth nearly $13 million at the main airport, after a cargo described as computer parts.
3. Early July 2026: Singapore police seize a bungalow valued at S$55 million ($42.4 million) and about S$1 million, and charge four people with money laundering in a case tied to Nvidia-class servers shipped toward Malaysia.
4. July 14, 2026: Nvidia’s shortened Asian whitelist becomes public, with Commerce officials described as part of the checks.
5. August 2026: BIS investigates Apex Logistics, a Singapore-based Kuehne+Nagel unit, over 47 shipments handled in 2024.

Apex said it is aware of U.S. “concern for and interest in a small number of shipments that Apex handled in 2024 and which may have involved materials and equipment which were ultimately forwarded to prohibited locations,” and that it is cooperating. The suspected path in that probe runs from Taiwan through the United States and Southeast Asia to Hong Kong, then over land into the mainland. Two former employees are described as having applied a shipping code that treated the hardware as outside U.S. export jurisdiction. Kuehne+Nagel said the subsidiary is cooperating and that the parent had not been contacted.

If that case ends in a penalty, it would be among the first aimed at a transportation company rather than a vendor or a fake cloud. The buyer list can only catch the customer of record. A mislabeled bill of lading does not need a neo-cloud account.

Restricted Chips Still Appear in Trade Data

On Sept. 9, 2026, the Washington nonprofit C4ADS published Covert Compute, an open-source map of how restricted Nvidia parts still reach Chinese users. The group found 50 export shipments of restricted NVIDIA GPUs through Vietnam, India and Malaysia toward Hong Kong and China between 2023 and 2025, totaling about $13.4 million. That trade-data total is a separate figure from the nearly $13 million airport seizure in Malaysia. In Chinese government records from July 2025 through January 2026, C4ADS also counted at least 56 restricted chips worth about $1.7 million bought by universities and research institutes, often bundled inside larger contracts.

The largest number in the same report is corporate, not crate-level. C4ADS put Megaspeed International’s Nvidia-hardware imports at about $4.6 billion from 2022 to 2025 and said the company’s ultimate owners may have ties to China, without treating the full import total as proven illicit trade. U.S. and Singapore authorities have been examining Megaspeed. Bain Capital’s Bridge Data Centers removed the firm from a Malaysian campus in April 2026. C4ADS said its counts almost certainly understate the true volume, because the study only tallied chips named in public records.

Former BIS official Kate Koren has argued the practical limit is scale, because moving thousands of accelerators at once is hard to hide. The whitelist is built for that large, inspectable order. It is a weaker tool against a relabeled pallet, a fishing-boat run, or a rented rack that was already legal the day it was powered on. U.S. policy still wants Southeast Asia full of American AI infrastructure and empty of Chinese workarounds. Those two aims now share the same halls, and only some of the doors have locks.