Nvidia chips pile up in Hong Kong as China withholds import approval

The H200 is an advanced Nvidia processor designed for demanding artificial intelligence work. The article highlights its value for training AI models, rather than calling it Nvidia’s newest chip. Training requires substantial computing power, so capable processors can affect how quickly companies develop and improve AI systems.

The H200 is especially useful because Chinese alternatives have not yet matched its capabilities. Nvidia’s software also strengthens its usefulness. Chinese technology companies therefore want access even as Beijing worries about dependence on US technology and encourages domestic processors for some tasks.

The chip is not freely available to Chinese buyers. US export controls restricted advanced Nvidia products, while Beijing has withheld approval for large-volume H200 imports. Some shipments have still entered China through approved purchases, intermediaries, and indirect routes. The result is a valuable but constrained resource for companies facing surging AI demand.

About 600,000 Nvidia H200 chips are reportedly stored in Hong Kong while awaiting shipment to China. This unusually large stockpile shows the gap between strong Chinese demand for advanced AI computing and Beijing’s reluctance to approve large imports.

China has allowed limited quantities to enter in recent months. Several major Chinese internet companies have each obtained about 10,000 H200 chips over the past few months. Other small shipments moved through intermediaries and indirect routes, although the article does not give their total volume.

The stockpile does not mean all 600,000 chips have been approved for Chinese use. Beijing has still not approved large-volume imports, according to people familiar with the matter. The contrast is striking: companies are receiving limited batches, while hundreds of thousands remain nearby in Hong Kong, waiting for permission and shipment.

China has allowed some H200 shipments because its technology companies need powerful computing capacity for surging AI demand. The H200 remains useful for training models because Chinese alternatives have not matched its capabilities, especially alongside Nvidia’s software.

At the same time, Beijing is concerned about relying more heavily on US technology. The government has therefore withheld approval for large-scale imports and told companies to prioritise Chinese processors for inference, the task of running AI models. Reported permissions have been limited.

This creates a controlled compromise rather than an open market. Approved Chinese customers can receive some H200 chips, and small quantities have also moved through intermediaries and indirect routes. But large volumes remain blocked, leaving about 600,000 chips in Hong Kong and forcing companies to balance immediate AI needs with China’s longer-term technology goals.

US export controls have restricted China’s access to advanced Nvidia chips and the equipment needed to make them since 2022. China has added its own limits by withholding approval for large H200 imports. Together, these restrictions have narrowed authorised supply.

Scarcity has pushed Chinese buyers toward intermediaries and indirect routes. That has increased prices in the unofficial market. By the end of October, an eight-card H200 server was offered in mainland China for more than 6 million yuan, or about US$900,000.

That price was about three times higher than a year earlier. Newer Nvidia Blackwell servers, which fall under US export controls, were offered for more than 20 million yuan each. Companies can still obtain some chips through approved or indirect channels, but limited supply and uncertain permissions make advanced Nvidia computing far more costly.

Chinese technology companies want H200 chips because they provide advanced computing power for a rapidly growing AI market. The article says the H200 is not Nvidia’s newest processor, but Chinese alternatives have yet to match its capabilities for training AI models.

The key mechanism is the combination of hardware and software. H200 chips supply the computing power needed to train models, while Nvidia’s software makes the chips particularly useful. Companies have therefore sought limited approved supplies and, in some cases, used intermediaries or indirect routes.

China is still trying to reduce dependence on US technology. Beijing has directed companies to prioritise Chinese processors for inference, which means running AI models. But that policy does not remove the immediate need for strong training hardware. As AI demand rises, companies face a practical trade-off between domestic priorities and Nvidia’s current advantages.

If companies cannot obtain enough H200 chips, the clearest alternative is to use Chinese processors. Beijing wants firms to prioritise domestic processors for inference, the task of running trained AI models. The article does not identify which Chinese processor models are available or compare their performance.

Another option is Nvidia’s RTX Pro 5500 graphics card, a less powerful product designed to comply with US export rules. Reuters reported that ByteDance, Alibaba, and Tencent ordered more than 3 million of these cards. Companies have also obtained small quantities of H200 chips through intermediaries and indirect routes, although those routes are not a true substitute for broad supply.

These alternatives may help companies continue building AI capacity, but the article presents them as constrained choices. H200 chips remain more capable for training, while domestic processors are being prioritised for inference. Chinese firms must therefore combine permitted Nvidia products, domestic hardware, and limited H200 access.

Training is the process of building an AI model by processing large amounts of data and adjusting the model so it can produce useful results. Running the model for users is called inference. It applies the trained model to new requests, such as generating an answer or prediction.

The two tasks differ in purpose, but both require computing power. Training repeatedly performs large calculations while the model learns, so advanced chips can help handle the workload. Inference also performs many calculations, often repeatedly for many users, so efficient processors matter there too.

The article directly connects the H200 with training AI models and says Beijing wants Chinese processors prioritised for inference. This division helps explain the policy tension. Chinese companies want H200 capability for training, while Beijing is trying to reserve more of the model-running work for domestic processors. The article does not compare their exact performance.