About 600,000 Nvidia H200 Chips Await China Entry as Beijing Limits Imports

A vast stockpile waits while Chinese buyers seek computing power

About 600,000 Nvidia H200 artificial intelligence chips are sitting in Hong Kong awaiting shipment to mainland China, according to two people familiar with the situation. Beijing has allowed some small purchases but has yet to approve imports on a scale that would release the stockpile, even as Chinese technology companies seek more computing power.

The restrictions have helped drive prices sharply higher. By the end of October, a server containing eight H200 chips was quoted at more than 6 million yuan on China's black market, roughly three times its price a year earlier. Servers using Nvidia's newer Blackwell architecture, which faces tighter American export controls, were quoted at more than 20 million yuan each.

China is trying to meet two competing needs: give its AI developers access to powerful American processors and build a domestic chip industry capable of supplying them. The H200 is no longer Nvidia's newest technology, yet its computing performance and established software remain attractive to companies training large language models, the systems behind advanced chatbots.

Washington's decision to permit sales to approved Chinese customers removed one obstacle. Permission from Beijing remains another. Chinese authorities have reportedly told companies including Alibaba, ByteDance and DeepSeek to reserve limited H200 allocations for training models while giving domestic processors priority when running those models for users.

The resulting market has substantial demand, stock waiting nearby and buyers unable to obtain as much hardware as they want through approved channels. For Chinese AI companies, the immediate constraint is therefore access to chips as well as their technical capabilities.

How small approvals compare with the stockpile

Some major Chinese internet companies have reportedly received permission to acquire about 10,000 H200 chips each. Small quantities have also entered mainland China through intermediaries and indirect channels. Beijing's total approved import quota could be below 200,000 chips, according to the reported estimate.

A quota below 200,000 would amount to less than one third of the approximately 600,000 chips said to be waiting in Hong Kong. That comparison helps explain why permission for individual purchases has not cleared the accumulated inventory.

The Chinese government has not publicly confirmed those stockpile and quota figures. The transcript of the Ministry of Commerce briefing on March 19, 2026, records a Bloomberg journalist asking whether China had already approved Chinese companies' purchases of H200 chips. Responding to that question, Ministry spokesperson He Yongqian said:

I am not aware of the situation you mentioned.

No date has been announced for broader import approval. Companies seeking large installations therefore face uncertainty about when they can obtain enough processors to put their planned computing capacity into service.

How American permission and Chinese restrictions developed

The United States began restricting exports of advanced Nvidia chips to China in 2022. Those controls limited the processors Chinese companies could buy, making access to powerful AI hardware a recurring issue in relations between Washington and Beijing.

In December 2025, President Donald Trump announced that Nvidia could sell H200 chips to approved customers in China and other approved countries, subject to American national security conditions. Nvidia welcomed the decision to supply commercial customers.

At a December 9 briefing, Chinese Foreign Ministry spokesperson Guo Jiakun said Beijing had noted the announcement and supported cooperation between China and the United States for mutual benefit. That response expressed a diplomatic position without setting out an import allocation.

The main developments stretch across several dates:

• 2022: The United States begins restricting exports of advanced Nvidia chips to China.
• December 2025: Trump announces permission for H200 sales to approved Chinese customers under national security conditions.
• December 9, 2025: Nvidia welcomes the approach, while China's Foreign Ministry responds in support of cooperation.
• March 19, 2026: Asked whether China has already approved Chinese companies' purchases of H200 chips, Commerce Ministry spokesperson He Yongqian says she is unaware of the situation mentioned.
• By the end of October: An eight chip H200 server is quoted at more than 6 million yuan in China's black market.

The two governments control different parts of the transaction. American export permission lets an eligible seller supply an approved customer. Chinese import approval determines whether that customer can bring the hardware into mainland China.

Why Beijing separates training from running AI

Alibaba, ByteDance and DeepSeek have reportedly been told that H200 purchases should be limited to model training, while domestic chips should receive priority for inference. Those are two different stages of an AI system's working life.

Training is the process of teaching a model by using large amounts of data and repeatedly adjusting its internal settings. It can require many processors working together for long periods. Inference happens afterward, when the trained model answers a question, writes text or performs another task for a user.

Beijing's reported approach gives developers some access to Nvidia hardware for creating models while directing the continuing work of serving users toward Chinese suppliers. For a company operating a popular chatbot, that could mean using different families of processors for development and everyday service.

Nvidia's appeal also extends beyond the physical chip. Developers use software tools to divide work among processors, manage memory and make models run efficiently. Moving to another supplier can require changes to that software and testing to make sure the system works reliably. This helps explain why Chinese companies continue seeking H200 chips despite efforts to expand domestic supply.

Scarcity drives the price of an eight chip server

The quoted price of more than 6 million yuan for a server containing eight H200 chips puts a concrete cost on restricted access. At roughly three times the level a year earlier, the current quotation compares with an earlier price of about 2 million yuan.

A server is a complete computing machine, rather than a box containing processors alone. Its price can include memory, networking equipment, other components and the premium a seller charges for obtaining scarce hardware. Buyers planning an AI installation must also pay for power, cooling and the connections between machines.

Blackwell servers were quoted above 20 million yuan each, showing the much higher sums being sought for access to Nvidia's newer architecture under stricter export controls. The reported quotations do not specify matching server configurations, so buyers would need those details to compare the computing capacity available at each price.

For Chinese developers, expensive hardware makes expansion more costly. A company that cannot secure an approved allocation must decide whether to wait, use different processors or seek computing capacity through another route. Those choices affect the resources available for training and operating its services.

Earlier orders show the scale of unmet demand

Chinese companies had already placed orders exceeding 2 million H200 chips for 2026, against reported Nvidia inventory of about 700,000 units. Those figures, described in December 2025, represented an immediate inventory gap of more than 1.3 million chips.

The earlier inventory estimate included approximately 600,000 standalone H200 chips and 100,000 GH200 Grace Hopper products. Nvidia also approached Taiwan Semiconductor Manufacturing Co. about additional production, with manufacturing expected to begin in the second quarter of 2026.

The present account places approximately 600,000 H200 chips in Hong Kong awaiting entry to mainland China. The commercial problem has therefore expanded beyond how many processors Nvidia can supply: a substantial quantity is reportedly available near its intended market but still awaits permission to enter.

The earlier orders also show why small allocations matter to individual buyers without satisfying the wider market. An allocation of about 10,000 processors can support a substantial installation, yet it represents only a small fraction of demand measured in millions.

Domestic suppliers are gaining customers

China's campaign to expand domestic supply is already changing purchasing patterns. IDC figures discussed in October put Nvidia's shipments of AI accelerators in China at 678,000 during the first half of 2026, down 44% from a year earlier. Huawei shipped 572,000, bringing the companies much closer in volume.

The figures behind Huawei's narrowing shipment gap gave Nvidia a 34% market share and Huawei 28%. China received 2.008 million AI accelerators during that period, with nine of the 11 Chinese chipmakers tracked by IDC increasing shipments from a year earlier.

Domestic suppliers are also competing through product demonstrations. At an October 28, 2026 event in Shanghai, Iluvatar CoreX presented its Tiangai 300 processor and described a test in which it achieved throughput equivalent to 1.05 times an H200, or 5% more output, under a particular inference arrangement.

The company's Tiangai 300 demonstration focused on an arrangement that separates reading a user's request from generating an answer. Iluvatar CoreX had not published the detailed test method, model list or workload conditions behind that comparison. For buyers following Beijing's preference for domestic inference hardware, performance on their own services will be central to purchasing decisions.

Technology companies seek other Nvidia products

ByteDance, Alibaba and Tencent have reportedly ordered more than 3 million Nvidia RTX Pro 5500 cards that comply with American export restrictions. Their purchases show that demand for Nvidia products extends beyond the H200, even as domestic suppliers gain ground.

The scale of those orders is striking beside the approximately 600,000 H200 chips reportedly waiting in Hong Kong. The RTX Pro 5500 order count is more than five times as large, although the products have different capabilities and the number of cards alone cannot describe the computing work each group can perform.

Chinese technology companies are consequently assembling their computing resources from several directions: limited H200 purchases for training, domestic processors for running models and other eligible Nvidia products. Each route offers a different combination of availability, performance and software requirements.

Beijing's policy gives Chinese chipmakers a larger opportunity to supply AI services, while leaving leading developers seeking access to American processors for demanding training work. The Hong Kong stockpile captures that unresolved purchasing problem in physical form: chips ready for potential customers whose access remains restricted.

Key Points

• About 600,000 Nvidia H200 chips are reportedly waiting in Hong Kong for shipment to mainland China.
• Some major internet companies have received allocations of about 10,000 chips each; total approved imports could be below 200,000.
• An eight chip H200 server was quoted above 6 million yuan by the end of October, roughly three times its price a year earlier.
• Beijing reportedly wants limited H200 supplies used for training, with domestic chips given priority for inference.
• ByteDance, Alibaba and Tencent have reportedly ordered more than 3 million RTX Pro 5500 cards.
• No date has been announced for broader H200 import approval.