600,000 Nvidia H200 Chips Wait in Hong Kong as China Limits Imports

A vast stockpile, but limited access for Chinese buyers

About 600,000 Nvidia H200 artificial intelligence chips are being held in Hong Kong awaiting shipment to mainland China, while Beijing has yet to approve imports on a large scale, according to people familiar with the matter. The stockpile sits alongside sharply rising prices: by the end of October, a server containing eight H200 chips was being offered in the mainland's unofficial market for more than 6 million yuan, about three times its price a year earlier.

Alibaba, Tencent, ByteDance and DeepSeek have reportedly each received permission to import only about 10,000 H200 chips in recent months. Small quantities have also entered through intermediaries and indirect routes, although the number moved that way was not disclosed.

The bottleneck puts Chinese AI developers between two governments making different decisions. Washington controls whether advanced American chips can be exported. Beijing controls whether Chinese companies can import them, while also encouraging those companies to buy processors developed at home.

The H200 belongs to Nvidia's older Hopper generation, rather than its newer Blackwell family. Even so, Chinese developers want it for demanding work such as training AI models, especially when it is used with Nvidia's software. Training means processing large amounts of data to teach a model to recognize patterns and produce useful responses.

For buyers, the immediate problem is getting hardware into computing facilities. Hundreds of thousands of processors waiting in Hong Kong represent potential computing capacity. Companies on the mainland can use that capacity only after the necessary permissions and deliveries are completed.

How much has China actually approved?

The reported allocations to Alibaba, Tencent, ByteDance and DeepSeek amount to roughly 40,000 chips combined. The 600,000 chips waiting in Hong Kong are about 15 times that total, illustrating the gap between the stockpile and the limited permissions described by people familiar with recent imports.

An earlier account described Chinese approval for several hundred thousand H200 chips, with the initial allocation concentrated among three major technology companies during a visit to China by Nvidia chief executive Jensen Huang. That larger figure contrasts with the more recent account of restricted allocations and a stockpile still awaiting permission. Chinese authorities have not publicly explained the scope or timing of those approvals.

The earlier permission was described as targeted approval for selected buyers. Companies seeking further access therefore still face decisions about their own purchases rather than an unrestricted reopening of the market.

Washington has opened a narrow licensing route

The United States began restricting China's access to advanced Nvidia processors and semiconductor manufacturing equipment in 2022. President Donald Trump announced on December 8, 2025, that authorized Chinese customers would be allowed to buy H200 chips under conditions that included a proposed 25% payment to the US government.

The Commerce Department's Bureau of Industry and Security subsequently revised its licensing policy for certain processors, including Nvidia's H200 and AMD's MI325X. The policy described in January 2026 covers chips with total processing power below 21,000 under the regulatory measure and DRAM memory bandwidth below 6,500 gigabytes per second. Memory bandwidth describes how quickly data can move between a processor and its memory.

Qualifying applications can receive individual review instead of facing a presumption of denial. The conditions include ready availability for American buyers and a limit on aggregate shipments to China and Macau equal to 50% of total use by US customers. The revised policy did not mention Trump's proposed 25% government payment.

US officials describe only limited shipments

Jeffery Kessler, the US under secretary of commerce for industry and security, told a congressional hearing that shipments under the licensing system had been small. His statement covered China and Hong Kong.

"The bottom line is very few shipments against licenses for H200s and equivalents have taken place. It's a very small quantity of chips," Kessler said.

Kessler said applicants must satisfy national security requirements and accept inspections to check compliance. He also told lawmakers that some license applications are denied.

Earlier, on February 24, 2026, David Peters, the assistant secretary of commerce for export enforcement, had told a House Foreign Affairs Committee hearing that H200 sales to Chinese customers stood at zero, to his knowledge. Kessler's subsequent description indicates that some licensed shipments had begun, while the scale remained limited.

The licensing process involves review by the US State Department, Defense Department and Energy Department. Disagreement over the security conditions had reportedly held up applications earlier in the year. For Chinese customers, an American export license is one step in a process that also requires permission from Beijing.

The decisions behind the bottleneck

The policy changed through several announcements and reviews, leaving customers and suppliers to make commercial plans while permissions were still being settled. The main dated developments were:

• December 8, 2025: Trump announced conditional permission for H200 sales to authorized Chinese buyers, including a proposed 25% US government payment.
• January 6, 2026: Huang said at the CES technology show in Las Vegas that China's position on H200 imports could become visible through purchase orders.
• January 8, 2026: Nvidia was reported to be requiring full payment in advance from Chinese H200 customers.
• January 13, 2026: Details of the revised US licensing policy described individual reviews and conditions protecting supply for American customers.
• February 24, 2026: Peters told lawmakers that H200 sales to Chinese customers were still at zero, to his knowledge.

Subsequent accounts describe both limited licensed shipments and a much larger stockpile awaiting mainland entry. No date has been announced for a decision allowing those 600,000 chips to move in bulk.

Scarcity pushes server prices higher

An eight chip H200 server offered for more than 6 million yuan costs roughly three times the quoted price a year earlier. That comparison suggests an earlier price around 2 million yuan. The figures are asking prices in the unofficial market, where buyers seek equipment outside ordinary authorized supply channels.

Servers using Nvidia's newer Blackwell processors were being offered for more than 20 million yuan each. Blackwell chips remain subject to US export controls and offer greater performance than the older Hopper generation.

The H200 and Blackwell prices cover complete servers, rather than individual processors. The quoted Blackwell systems also lack enough configuration detail for a comparison between equivalent machines. Buyers must consider the processors, memory, connections and other equipment included in a server when deciding how much useful computing capacity they are purchasing.

Intermediaries and indirect routes have become more attractive as authorized supply has tightened. Those purchases also carry enforcement risks. Peters told lawmakers in February that advanced semiconductor smuggling was occurring and proposed extending the time allowed to bring cases to 10 years, with penalties exceeding four times the value of illegal transactions.

Large demand comes with difficult payment terms

Chinese customers previously sought more than 2 million H200 chips, representing a potential sales channel valued at about $54 billion. Nvidia's reported inventory of roughly 700,000 processors was far smaller. The company pursued additional H200 production as customer demand exceeded existing stock by more than 1.3 million units.

Production discussions included Alibaba and other large Chinese buyers. Nvidia reportedly planned to expand H200 output after demand surpassed capacity, even while developing newer processor generations.

Payment terms shifted some of the regulatory risk toward customers. In January, Nvidia reportedly required Chinese H200 buyers to pay the full amount in advance, without the right to cancel, obtain refunds or change configurations. In exceptional cases, commercial insurance or pledged assets could replace cash payment.

Nvidia had previously sought advance payments, but the demand for full payment was described as unusual. A customer accepting those terms would commit money while Chinese permission to receive the equipment remained uncertain. Later accounts described Huang saying Nvidia had received no new H200 orders while import assessments continued, and some suppliers pausing production of key components.

China's alternatives are growing, but tasks matter

Beijing has urged companies to reduce reliance on imported processors while Huawei and other domestic suppliers expand their AI hardware businesses. China was also reported to be considering semiconductor industry incentives worth up to $70 billion. That figure concerned a proposal, rather than a confirmed spending allocation.

Chinese suppliers have already gained ground in shipment volumes. IDC figures described in China's changing chip market put Nvidia's deliveries at 678,000 AI accelerators in the first half of 2026, against Huawei's 572,000. Their market shares were 34% and 28%, respectively, a gap of six percentage points.

Performance comparisons depend on the job being done. Chinese alternatives have been described as lagging the H200 for demanding model training, particularly when Nvidia's software is included. Meanwhile, Iluvatar CoreX claimed a 5% advantage over the H200 in a particular test of inference, the process of running a trained model to generate answers.

At its October 28 demonstration in Shanghai, the company presented the result as part of its Tiangai 300 demonstration. The claimed comparison used a deployment that separated the stage when a model reads a request from the stage when it produces an answer. Buyers choosing hardware need performance and software suited to their own training or inference work.

Buyers seek other Nvidia products too

ByteDance, Alibaba and Tencent reportedly ordered more than 3 million Nvidia RTX Pro 5500 graphics cards, a less powerful product described as compliant with export restrictions. Their combined sticker price exceeded $35 billion.

Those purchases show the scale of demand for processors that Chinese companies can obtain through permitted channels. A graphics card that is easier to purchase can offer another route to computing capacity, even when it delivers less performance than the hardware a developer originally wanted.

Nvidia has kept potential Chinese AI chip revenue out of its forecasts because access to the market depends on regulatory decisions. Huang told investors in May to "expect nothing" from Chinese sales.

The immediate pressure falls on companies planning AI training and services. They must decide whether to wait for H200 access, purchase other available Nvidia products or commit more work to domestic processors. The Hong Kong stockpile makes the central problem concrete: hardware exists nearby, but government permissions determine when mainland developers can put it to work.

Key Points

• About 600,000 Nvidia H200 chips are reportedly waiting in Hong Kong for permission to enter mainland China.
• Alibaba, Tencent, ByteDance and DeepSeek have reportedly each received permission for about 10,000 chips in recent months.
• US rules allow individual licensing reviews under conditions, while China separately controls imports.
• An eight chip H200 server was offered for more than 6 million yuan by the end of October, about three times its price a year earlier.
• Chinese developers are also purchasing less powerful Nvidia products and expanding their use of domestic processors.
• No date has been announced for approval of bulk shipments from the Hong Kong stockpile.