Nvidia has reduced the number of authorized Asian customers to purchase its most advanced AI accelerators by more than half. The company is implementing stricter controls over cloud providers, data center operators, and distributors to ensure that the GPUs do not end up in China through third countries.
This measure does not mean Nvidia has stopped selling any AI processors to the Chinese market. While increasing scrutiny on Blackwell systems intended for other countries, the US has permitted limited exports of H200 to previously authorized Chinese companies. The goal is to distinguish between licensed sales and diversion attempts to bypass US restrictions.
Key points of Nvidia’s new control in 20 seconds
- Nvidia has created a smaller list of authorized Asian buyers.
- More than half of previous customers did not pass the new checks.
- Excluded companies can reapply after correcting deficiencies.
- The controls mainly affect cloud companies specialized in AI infrastructure.
- Nvidia reviews contracts, data centers, end users, and hardware location.
- Malaysia, Singapore, and Japan are among the markets under increased scrutiny.
- The US Department of Commerce backs the tightening, according to published information.
- The priority is to prevent Blackwell systems from indirectly reaching China.
- Nvidia and the Department of Commerce have not publicly confirmed the list details.
- Authorized exports of H200 to certain Chinese customers follow a different process.
The new policy comes after Washington intensified pressure on manufacturers and distributors to know not only the direct buyer but also the end user of each server. Nvidia is reportedly performing onsite visits, interviewing clients, and checking hosting contracts, electrical supply, and data center construction before approving orders.
A whitelist for Blackwell GPU buyers
Demand for AI accelerators has led to many companies purchasing large server quantities to rent out computing capacity. Some operate their own data centers with identifiable clients. Others are newer entities relying on third-party facilities or cannot clearly demonstrate who will ultimately use the equipment.
This second group raises most of Nvidia’s and US authorities’ concerns. A company registered in Singapore or Malaysia can make a legal purchase, but the operation becomes non-compliant if the servers are later transferred to China or if remote access is granted to an entity that couldn’t purchase them directly.
The whitelist attempts to close this gap. To remain an authorized buyer, presenting a formal company and purchase order is no longer enough. Nvidia requests evidence of the data center’s existence, electrical capacity, customer contracts, and the specific destination of the systems.
Over half of the previously admitted Asian buyers initially failed the process. This does not necessarily mean Nvidia has permanently lost half of its regional customers, as might be inferred from a quick reading. These companies can reapply after modifying procedures or providing additional information.
The tightening primarily targets Blackwell processors, whose export to China is restricted and which offer more capacity than the GPU models specifically authorized for certain operations there. Nvidia’s risk isn’t limited to financial penalties: a distribution chain unable to prove where the hardware ends up could face further bans, license delays, and stricter controls on all clients.
The Department of Commerce is reportedly collaborating with Nvidia and backing the verification process politically. Still, the detailed existence and operation of the list come from sources cited by Financial Times. Nvidia and the department had not responded publicly to requests for comment when the information was leaked.
The Supermicro case showed how inspections could be fooled
US concern increased after reports emerged of a scheme to divert AI servers to China via intermediaries and fake documentation.
In March 2026, US prosecutors charged three individuals related to Super Micro Computer with participating in a scheme handling orders worth approximately $2.5 billion. The indictment states that at least $510 million worth of Nvidia technology servers reached China despite restrictions.
Prosecutors allege the group used a Southeast Asian transit company, manipulated documentation, and fictitious equipment to evade inspections. The accused reportedly moved labels and serial numbers from real servers to cases prepared to appear as if the hardware remained at the authorized site.
Supermicro was not charged as a company and claimed to be cooperating with authorities. Nvidia is not implicated in the alleged scheme and has stated it provides no support or services for illegally exported equipment.
The case illustrates the limitations of one-time inspections. Visiting a data center once does not guarantee the hardware remains there. Servers might appear installed while the actual units have been shipped elsewhere.
Therefore, new controls seem aimed at verifying a broader chain: who pays for the purchase, who operates the infrastructure, which clients rent the GPUs, the physical location of the data center, and whether electricity consumption matches the declared hardware quantity.
Access remotely complicates this further. A company in China doesn’t need to physically transport a GPU there for its use. It can contract legally hosted infrastructure in another country, provided US regulations are respected and the end customer isn’t restricted. Determining when a cloud service is legitimate versus when it’s a means of circumventing restrictions remains a key issue Washington is trying to clarify.
US authorities are also preparing new measures on semiconductors and AI. Jeffrey Kessler, head of export controls at the Department of Commerce, stated on 07/14/2026 that new regulatory actions are forthcoming but did not specify details during his congressional appearance.
US restricts Blackwell but still allows some H200 exports
The situation is less straightforward than the idea that Nvidia aims to prevent all AI GPUs from reaching China.
The same day the reduction in Asian buyers was announced, a high-ranking Commerce Department official confirmed Nvidia had begun shipping a limited number of H200 accelerators to China. These exports are under licenses granted to specific companies and do not constitute a general market opening.
Recipients include Alibaba, Tencent, ByteDance, JD.com, ZTE, and other tech or infrastructure firms. According to the US government, as of 07/14/2026, only a small amount of chips had been shipped despite several licenses being approved.
Washington is thus applying two policies simultaneously: on one hand, allowing certain Chinese buyers to acquire H200 under conditions and oversight; on the other, trying to prevent Chinese-linked companies from obtaining more advanced Blackwell systems via subsidiaries, intermediaries, or cloud providers in third countries.
China also controls the import of these systems. Beijing previously halted purchases while promoting domestic accelerators, but is now considering allowing its major AI companies to import fewer than 200,000 H200 units. This number may be less than half of what interested companies initially requested.
Therefore, it’s inaccurate to say China has definitively rejected 200,000 GPUs. That figure reflects the maximum potentially permitted, according to unconfirmed Chinese government reports. It also does not mean all those chips have been manufactured, shipped, or delivered.
The balancing act reflects conflicting interests: Chinese companies need computing power to train and run advanced models, but the government seeks to reduce US technological dependence. Nvidia aims to regain part of a market that was significant for its business, though it must avoid triggering new restrictions with sales.
Meanwhile, the US seems to prefer a controlled dependency: allowing some exports keeps Chinese developers within Nvidia’s hardware and software ecosystem while reserving the newest, most powerful products for authorized clients outside China.
Supply chain control will be costly and never fully airtight
The new whitelist can limit obvious diversion but cannot guarantee the disappearance of the black market.
GPUs are relatively small, valuable, and easy to resell. Servers are harder to conceal but can be disassembled, transported piece by piece, or misrepresented with false descriptions. Intermediary companies claiming independence from the end user can also be created.
Frequent inspections increase costs for Nvidia and its clients. Legitimate operators may be required to reveal contracts, ownership structures, usage projections, and customer details to retain access to accelerators. Startups, even if not tied to China, could be excluded for lacking sufficient facilities or documentation.
This policy benefits larger cloud providers with established legal teams, known data centers, and compliance procedures. Smaller suppliers may face greater hurdles in directly purchasing hardware, potentially consolidating AI capacity in fewer companies.
There’s also a risk that restrictions accelerate China’s domestic industry. Bans on advanced chips have motivated China to invest in Huawei, Cambricon, Biren, and other local manufacturers. Limiting foreign supply doesn’t automatically halt model development—companies can improve software efficiency, share infrastructure, or deploy more lower-performance units.
For Nvidia, the whitelist is both a business safeguard and a regulatory obligation. The company needs to control its sales network to continue exporting to Asian markets and maintain Washington’s trust. Simultaneously, it must sell enough hardware to prevent clients from seeking alternatives.
The result is a less straightforward policy than a complete shutdown. Nvidia carefully monitors Blackwell orders in Asia; the US authorizes limited H200 exports to China; and Beijing decides how much foreign hardware it allows in. Between these decisions, an international network of distributors, data centers, and cloud providers now must demonstrate much more precisely who is using each GPU.
FAQs
Has Nvidia stopped selling AI GPUs to China?
No. The most advanced Blackwell GPUs remain restricted, but the US has authorized limited exports of H200 to certain Chinese buyers.
Has Nvidia lost half of its Asian customers?
The company appears to have removed more than half of previous buyers from its authorized list. This isn’t necessarily permanent, as they can reapply after making adjustments.
What does Nvidia check before approving a sale?
According to the published information, they review data centers, contracts, electrical capacity, end users, and the intended location of the hardware.
Why are Malaysia and Singapore concerned?
Both countries host data centers and intermediary firms capable of legally acquiring servers. The US fears some transactions may hide end users in China under restrictions.
