600,000 Nvidia H200 Chips Wait in Hong Kong as China Holds Back Bulk Imports

A vast chip stockpile waits at China's doorstep

About 600,000 Nvidia H200 artificial intelligence chips are being held in Hong Kong while China withholds approval for large shipments into the mainland, according to two people familiar with the matter. The stockpile puts a number on a costly standoff: Chinese technology companies want powerful American processors, while Beijing wants to limit dependence on American technology.

Some chips have entered mainland China in recent months. Three people familiar with the situation said authorities had allowed limited shipments, and one said several major Chinese internet companies had each obtained about 10,000 H200 chips. Three people also described small shipments through intermediaries and indirect routes, without giving a total.

The difference between those deliveries and the inventory waiting in Hong Kong is substantial. The stockpile contains 60 times the roughly 10,000 chips that individual companies reportedly received. That comparison shows the scale of the supply waiting outside the mainland, although the number of companies receiving those smaller allocations was not disclosed.

China's Ministry of Industry and Information Technology and the National Development and Reform Commission did not respond to requests for comment. Nvidia, Alibaba, ByteDance, DeepSeek and Tencent also did not respond. No date has been announced for a decision allowing imports on a large scale.

For businesses building AI systems, the immediate problem is practical. Hardware can be manufactured, purchased or positioned close to customers, yet still remain unavailable for installation in their computing centres. The Hong Kong inventory sits at precisely that point between potential supply and usable computing power.

Why American permission is only half the process

Washington and Beijing control different sides of the transaction. The United States decides whether Nvidia can export restricted American technology to particular customers. China decides whether those products can enter its market and under what conditions.

American export controls introduced in 2022 restricted China's access to advanced Nvidia processors and equipment used to manufacture sophisticated chips. President Donald Trump's administration later permitted H200 sales to approved Chinese customers following trade negotiations. Beijing continued to impose its own limits on imports.

The American approval process also carries technical and security conditions. A final rule issued by the US Commerce Department's Bureau of Industry and Security on January 13 required relevant chips to undergo performance testing at an independent laboratory in the United States before export. The process also requires confirmation that the chips will not be used for military purposes.

Trump signed a related presidential proclamation on January 14, following an earlier announcement that the United States would collect a 25% fee on the transactions. For buyers and Nvidia, permission therefore comes with compliance requirements and a financial charge, rather than unrestricted access to the Chinese market.

Beijing's concern runs in the opposite direction. People familiar with its thinking described reluctance to deepen reliance on American technology. Buying processors can help Chinese companies build better AI systems now, while making those companies more dependent on a supplier whose access to China has repeatedly changed with government policy.

Orders, approvals and deliveries tell different stories

Nvidia chief executive Jensen Huang gave a more cautious account of commercial progress during a January 29 visit to Taipei, after travelling in China. He said Nvidia had received no H200 purchase orders from Chinese companies and was still waiting for Beijing's decision.

Huang's statement conflicted with reporting that ByteDance, Alibaba and Tencent had received approval to buy more than 400,000 H200 chips. Other reporting described orders for more than 2 million chips. Those figures concern different stages of a transaction: regulatory permission, customer orders and physical delivery.

A company may seek permission before submitting a firm order. An approved purchase may still require export checks and import clearance before delivery. The later account of 600,000 chips in Hong Kong adds another stage, inventory positioned for shipment but still awaiting permission to enter the mainland.

The figures consequently need their dates and descriptions attached. Huang's January 29 statement was a company account of orders at that time; the Hong Kong figure describes a subsequently reported stockpile. A complete public breakdown identifying its owners, intended customers and contractual status has not been released.

How the restrictions developed

The dispute has unfolded through several policy decisions and reports of limited access. The dated developments show why expectations of a broad reopening have repeatedly run ahead of confirmed deliveries.

• 2022: The United States introduced controls restricting China's access to advanced AI chips and chip manufacturing equipment.
• January 13: The Bureau of Industry and Security issued a final rule setting conditions for H200 exports, including independent testing in the United States.
• January 14: Trump signed a related presidential proclamation.
• January 29: Huang said Nvidia had received no Chinese H200 purchase orders and Beijing was still considering whether to allow imports.
• July: Reporting described possible Chinese approvals for fewer than 200,000 H200 chips, mainly for training AI models.
• By the end of October: An H200 server containing eight cards was being offered in mainland China for more than 6 million yuan.
• The latest stockpile disclosure: About 600,000 H200 chips were reported waiting in Hong Kong while approval for large shipments remained on hold.

Chinese regulators have also reportedly instructed major technology companies to submit H200 purchase requests for approval. Proposed conditions included buying a specified share of domestic processors alongside Nvidia products. The percentage was not disclosed.

The July account described possible permission for companies including Alibaba, ByteDance and DeepSeek to buy limited quantities for training. Its estimate of fewer than 200,000 chips concerned potential approvals, while the much larger Hong Kong figure concerns inventory awaiting entry.

Why an older Nvidia processor still matters

The H200 belongs to Nvidia's Hopper generation. It is more capable than the H20 designed for the Chinese market, but sits behind Nvidia's newer Blackwell generation. Its appeal comes from the computing capability Chinese buyers can actually obtain, together with the software used to run it.

Training an AI model means processing large quantities of examples so that a system learns patterns. A language model, for instance, learns relationships among words and ideas through repeated calculations. Powerful processors allow developers to perform more of that work in a given period.

The H200 has capabilities that Chinese alternatives have yet to match for some demanding training tasks, particularly when paired with Nvidia's software. Companies therefore have a reason to pursue the chips even though they are no longer Nvidia's newest product.

Chinese officials have reportedly urged companies to favour domestic processors for inference. Inference is what happens after training, when a model uses what it has learned to answer a question, generate an image or carry out another task.

Beijing's reported approach would give companies some access to Nvidia hardware for training while directing more everyday operation of AI services toward Chinese suppliers. Training and inference both need computing power, but separating purchasing priorities gives domestic chipmakers a larger role as AI services attract more users.

Scarcity sends server prices sharply higher

An H200 server containing eight cards was offered in mainland China for more than 6 million yuan, approximately $900,000, by the end of October, according to a person familiar with the market. That was about three times its price a year earlier.

The comparison points to an earlier price of roughly 2 million yuan for a similar server. Moving from around 2 million to more than 6 million yuan represents an increase of about 200%. For a business buying many servers, that change can add heavily to the cost of building computing capacity.

Servers using Nvidia's newer Blackwell chips were offered for more than 20 million yuan each, the same person said. Blackwell products also fall under American export controls. The quoted prices cover complete servers, so their costs include more than the processors alone.

Limited authorised supply has pushed buyers toward intermediaries and indirect routes. That creates a market in which access itself commands a premium: a company may pay substantially more to obtain hardware sooner rather than wait for a regular supply channel to open.

The delivery bottleneck also affects planning. A company preparing to train a model needs servers, electricity, networking equipment and staff available together. Uncertain chip delivery can leave the rest of that investment waiting or force the company to choose different hardware.

Chinese buyers are pursuing another Nvidia route

ByteDance, Alibaba and Tencent have also reportedly ordered more than 3 million Nvidia RTX Pro 5500 graphics cards. These are less powerful products that comply with export rules, with a combined sticker price exceeding $35 billion.

The volume is more than five times the approximately 600,000 H200 chips reported in Hong Kong. The two products have different capabilities, so the larger card count describes a different purchasing strategy rather than an equivalent amount of computing power.

The purchases show that major Chinese companies are pursuing several routes at once: seeking access to H200 processors, buying Nvidia products permitted under export rules and responding to pressure to use domestic chips. Each route offers a different balance of performance, availability and regulatory exposure.

For Chinese suppliers, requirements to buy domestic processors could secure customers even while some American chips enter the country. For technology companies, the result could be computing systems built around several kinds of hardware, with different processors assigned to different tasks.

Billions in possible sales, with major conditions attached

Investor estimates illustrate the commercial stakes for Nvidia. A reported demand figure of 2 million H200 chips at $27,000 each produces a potential transaction value of $54 billion. Huang has described Chinese demand as very high and suggested possible annual sales exceeding $50 billion.

Applying a simple 25% deduction to $54 billion leaves $40.5 billion. An investment analysis rounded that figure to about $40 billion and used it to model higher Nvidia earnings and a possible share price of $385, assuming a 56% net profit margin and a price to earnings multiple of about 46.

Those calculations depend on purchases being approved and completed, as well as assumptions about the charge, costs, profit margins and market valuation. Huang's January statement that no orders had been received also makes timing essential when considering the reported 2 million chip figure.

American political resistance adds another constraint. Representative John Moolenaar, chairman of the House Select Committee on the Chinese Communist Party, wrote to Commerce Secretary Howard Lutnick urging strict enforcement of the military use restrictions attached to H200 exports.

Moolenaar alleged that Nvidia's technical assistance to DeepSeek had helped develop capabilities subsequently used in Chinese military systems. His letter argued for enforcing the certification requirements even if doing so prevented H200 exports to China. The allegation and policy demand came from the committee chairman, while the export rules place responsibility on the approval process to assess permitted uses.

Nvidia's opportunity therefore depends on decisions in both capitals. Chinese customers want more computing power, Beijing wants domestic suppliers to gain ground, and Washington wants controls over where advanced American chips go. For now, the most concrete result is a large inventory in Hong Kong and limited deliveries across the border.

Key Points

• About 600,000 Nvidia H200 chips are reportedly waiting in Hong Kong for permission to enter mainland China.
• Several major Chinese internet companies have reportedly received about 10,000 H200 chips each.
• US export permission includes testing and security conditions; China separately controls import approval.
• Beijing reportedly wants domestic processors prioritised for inference while allowing limited H200 access for training.
• An H200 server with eight cards was offered for more than 6 million yuan by late October, about three times its price a year earlier.
• ByteDance, Alibaba and Tencent have reportedly ordered more than 3 million less powerful RTX Pro 5500 cards.
• No date has been announced for approval of H200 imports on a large scale.