What is happening with Nvidia's H200 chips in Hong Kong?
About 600,000 Nvidia H200 chips are sitting in Hong Kong, waiting for permission to enter mainland China. Reuters reported the figure on 5 November 2026, citing two people familiar with the matter. The size of the stockpile had not been reported before.
Beijing has not approved large-scale imports of the chip, even though Chinese technology companies want it. Nvidia, Alibaba, ByteDance, DeepSeek and Tencent did not respond to requests for comment, and neither did two Chinese ministries. That leaves the figure resting on unnamed sources rather than an official disclosure.
If the figure is accurate, hundreds of thousands of chips are parked within a short drive of the mainland border rather than installed in Chinese data centres.
The H200 is not Nvidia's newest processor. It is a Hopper-generation data centre GPU with 141GB of HBM3e memory and 4.8TB/s of bandwidth, according to Nvidia. Blackwell chips are newer, and Nvidia's roadmap has moved on. Even so, the H200 remains useful for training large AI models, especially when paired with Nvidia's software.
Hong Kong is a free port with its own customs system, which makes it a convenient holding point for goods destined for the mainland. Some H200s have entered China in recent months. Several large internet companies have each received about 10,000 chips, one source told Reuters. Smaller quantities also arrived through intermediaries and indirect routes.
Why is Beijing holding back large-scale H200 imports?
The delay is not a simple ban. It reflects a conflict between China's need for computing power and its wish to reduce dependence on American technology. Reuters describes the stockpile as a sign of that reluctance.
Demand is not imaginary. Chinese internet companies are spending billions of dollars on data centres to build AI services and compete with US rivals, Reuters reported. That spending explains why the same government that limits imports also faces pressure to keep its AI labs supplied.
US Commerce Secretary Howard Lutnick described the dynamic at a Senate hearing in April 2026. Beijing “has not let them, as of yet, buy the chips,” he said. Officials were “trying to keep their investment focused on their own domestic industry.”
Security concerns add friction. Chinese authorities have worried about tampering or hidden vulnerabilities in imported hardware, Reuters reported in May. The State Council also issued two supply chain security regulations, prompting a government-wide effort to identify foreign dependencies in critical technology infrastructure.
Beijing has discussed requiring companies to buy a quota of domestic chips before approving foreign imports, Reuters reported. The Information reported in July that officials wanted Chinese processors used for inference, the work of running trained models. Imported H200s would be reserved for training.
That split has a practical logic. Training a frontier model requires thousands of chips working together for weeks. Inference can run on smaller clusters and is easier to shift to domestic hardware. Officials could therefore allow a limited number of H200s for training while pushing inference toward Huawei.
No public rule sets out volumes, eligibility or timelines. Companies apply, wait and receive little explanation. That opacity helps explain why reported approval totals and actual shipments have diverged so sharply.
How many H200 chips has China actually let in?
The paper trail runs from Washington to Beijing and back. The H200 became the test case because it is far more capable than the H20, the previous Nvidia chip cleared for China. Huawei now sells products that rival the H20, Reuters reported. Reuters put the H200 at roughly six times the H20's performance. Washington had barred H200 exports before Trump's December 2025 reversal, which followed US-China trade negotiations.
On 8 December 2025, President Donald Trump said the US would allow H200 sales to approved customers in China. The Commerce Department's Bureau of Industry and Security then issued a case-by-case licensing rule on 13 January 2026.
That rule requires buyers to show export compliance procedures, including customer screening. It also requires the hardware to undergo independent, third-party testing in the United States. Nvidia must certify that US customers will not lose production capacity. Reporting by Reuters described two further conditions: buyers must demonstrate sufficient security procedures and not use the chips for military purposes. The US takes a 25% share of revenue, and chips must pass through US territory, Reuters reported.
Beijing moved next. On 28 January 2026, China approved a first batch covering more than 400,000 H200 chips for ByteDance, Alibaba and Tencent, Reuters reported. The approval came during a visit by Nvidia chief executive Jensen Huang. Chinese firms had placed orders for more than two million H200s, far above Nvidia's available inventory.
Deliveries did not follow at that scale. By 14 May, the US had cleared about ten Chinese buyers, each allowed up to 75,000 chips, but not a single delivery had been made, Reuters reported. Lenovo and Foxconn were approved as distributors.
The timing was delicate. Huang joined Trump's delegation to Beijing that month after an invitation from the president, Reuters reported. The trip raised hopes that talks with Xi Jinping could unlock the stalled sales. It did not produce large deliveries.
In July, Under Secretary of Commerce Jeffrey Kessler told Congress that “very few shipments against licenses for H200s and equivalents have taken place.” “It's a very small quantity of chips,” he added. By November, major companies had still received only about 10,000 chips each. Beijing had not approved large-volume imports.
The numbers are hard to reconcile because they measure different things. An approval is not a customs clearance, a purchase order is not a delivery, and a licensed shipment is not an installed server. The figure of 600,000 chips in Hong Kong is the clearest measure of the gap between demand and permission.
What do gray-market prices say about demand?
Where official channels are slow, informal ones fill the gap. An eight-card H200 server was offered for more than 6 million yuan, about $900,000, in mainland China at the end of October. That was roughly three times its price a year earlier, one person told Reuters.
Servers built around Nvidia's newer Blackwell chips were offered for more than 20 million yuan each. Those chips remain subject to US export controls. The price gap shows how much buyers will pay to skip the queue.
Demand is not limited to H200s. In October, ByteDance, Alibaba and Tencent ordered more than three million Nvidia RTX Pro 5500 graphics cards. The product is less powerful than the H200 but complies with US export rules. The combined sticker price exceeded $35 billion, Reuters reported.
Those orders are due to ship from the first quarter of 2027, and Reuters reported that supply would be limited. For Chinese AI labs, the queue matters as much as the chip. Training runs are planned months ahead, and a delayed cluster can slip a model release.
The workaround has limits. The RTX Pro 5500 is a less powerful product, and Chinese buyers still want H200s for the largest training jobs. The two order books tell different stories: one shows demand being redirected, the other shows demand being deferred.
Gray-market purchases carry their own risks. Buyers may get no warranty, no vendor support and no certainty about the hardware's history. Reuters reported that some chips arrived through intermediaries and indirect routes, without saying how many.
Can Huawei's Ascend chips replace the H200?
Beijing's restrictions only make sense if domestic alternatives can carry more of the load. Huawei is the most important supplier. At its Connect event in September 2025, the company published a three-year Ascend roadmap.
The Ascend 950PR was scheduled for the first quarter of 2026, aimed at prefill and recommendation workloads. The 950DT is due in the fourth quarter of 2026, with 144GB of memory and 4TB/s of bandwidth for decode and training. The 960 is planned for late 2027, and the 970 for late 2028.
Huawei is also selling systems rather than standalone chips. Its Atlas 900 A3 SuperPoD links up to 384 Ascend 910C chips and delivers up to 300 PFLOPS, the company says. The 950 series raises interconnect bandwidth to 2 TB/s, which Huawei says is 2.5 times the 910C. That approach trades per-chip efficiency for scale, which can work for large training clusters if the software keeps up.
On paper, the 950DT's memory figures sit close to the H200's 141GB and 4.8TB/s. Raw capacity does not settle performance, though. Benchmark results depend on software, networking and how well a model is tuned for the hardware.
Software is the harder gap. Nvidia's CUDA ecosystem is mature, and Reuters notes that the H200's value comes partly from that stack. Huawei is opening its CANN software to outside developers. It said CANN interfaces would be open-sourced by 31 December 2025.
Chinese AI companies are testing the alternatives. DeepSeek and others have increasingly highlighted their use of domestic chips, Reuters reported. The H200 still has capabilities that Chinese chips have not matched, particularly for training, according to the same reporting.
The practical result is a split workload. Inference may move to domestic silicon first, because it is more tolerant of software gaps and easier to scale across many small jobs. Training stays on Nvidia where possible, because a failed run wastes weeks and millions of dollars.
What does the standoff cost Nvidia?
Less than the headlines suggest, at least in the near term. Nvidia's results for the quarter ended 26 July 2026 showed revenue of $96.2 billion and data centre revenue of $89.0 billion. For the following quarter, it guided to $108.0 billion and assumed no data centre compute revenue from China.
That guidance choice matters. It means the 600,000 chips in Hong Kong are not baked into Nvidia's forecast. Any large approvals would be upside, not a rescue. It also means investors get little direct read on Chinese demand from the company's numbers.
The longer-term picture is worse. Before US curbs tightened, Nvidia held about 95% of China's advanced chip market. China accounted for 13% of its revenue. Huang has said the company's share of AI accelerators in China has effectively fallen to zero.
Huang estimated in May that China's AI market would be worth $50 billion in 2026. The same month, he told investors to “expect nothing” from Chinese sales. Some China hawks in Washington argue that any resumption helps Beijing close the gap. “Any deal that allows Nvidia to sell more chips to China means fewer Nvidia chips for U.S. firms,” said Chris McGuire of the Council on Foreign Relations, who warns of a smaller US lead in AI.
Holding inventory abroad is not free. The chips tie up working capital, and they age as newer products ship. The H200 is already a generation behind Blackwell. The 25% revenue share to the US government also trims the economics of each licensed sale.
Nvidia's next quarterly report is scheduled for 17 November 2026, according to Wall Street Horizon's earnings calendar. Analysts will listen for any change in the China language, even if the guidance excludes the market.
What should you watch next?
The first signal is whether Beijing approves a large batch. Reuters' reporting suggests the decision is political, not technical. A published quota, or a formal rule on domestic-chip conditions, would end much of the guesswork.
The second is Huawei's 950DT. Huawei has said it will ship in the fourth quarter of 2026. If volumes arrive on schedule, Chinese buyers gain a credible alternative for some workloads. If they slip, pressure to allow H200 imports will grow.
The third is enforcement. US rules require third-party testing and end-use checks. Reports of indirect routes invite tighter controls, which could slow gray-market supply and raise prices further.
The fourth is Nvidia's results on 17 November. The company has told investors to expect nothing from China. Any hint that licences are turning into revenue would change that baseline.
Customs data and distributor disclosures are worth tracking too. Lenovo has confirmed that it is approved to sell H200s in China, which gives the market a public signal when shipments move.
A large approval would change prices quickly. Gray-market premiums would fall as licensed supply reached buyers, and Nvidia would convert inventory into revenue. A continued freeze would do the opposite, pushing more buyers toward Huawei and toward indirect routes.
For technology buyers outside China, the standoff is a reminder that AI capacity is allocated politically. Server prices, delivery dates and even model release schedules can move with a customs decision in Beijing or a licensing decision in Washington.