NVIDIA's 600,000 H200 AI Chips Are Stuck in Hong Kong — and China Is Blocking Them

Roughly 600,000 Nvidia H200 AI chips are sitting in warehouses in Hong Kong, waiting for permission to enter mainland China. The United States has already cleared the exports. It is China — not Washington — that is refusing large-scale customs clearance, as Beijing pushes its semiconductor industry toward self-reliance. The chips are worth an estimated $20 billion to $30 billion at list prices, and the standoff is now the single biggest swing factor in Nvidia's China business.

That is the opposite of how this story is usually told. For four years, the familiar version has been Washington blocking advanced chips from reaching China. This time Washington opened the door, and Beijing bolted it shut. The reversal matters because it changes who is responsible, what the fix would look like, and what happens to Nvidia's earnings if the door stays closed.

What is stuck, and where

The H200 is Nvidia's previous-generation flagship AI accelerator — a Hopper-architecture chip with 141GB of high-bandwidth memory, built for training and running large AI models. It is not Nvidia's newest product; the Blackwell generation has succeeded it. But for Chinese AI companies, it remains the most capable chip they can plausibly obtain, because Blackwell is subject to US export controls that bar it from China altogether.

According to Reuters, which first reported the stockpile on 5 November 2026 citing multiple sources, roughly 600,000 H200s are piled up in Hong Kong warehouses awaiting approval to enter mainland China. The shipment is worth $20 billion to $30 billion at list prices — the first time a specific figure for the H200 inventory has become public. Hong Kong's government, China's Ministry of Industry and Information Technology, the National Development and Reform Commission, and the companies involved — Nvidia, Alibaba, ByteDance, DeepSeek and Tencent — did not respond to requests for comment.

Hong Kong matters here because it is the natural staging point. It is a separate customs territory from mainland China, with its own import rules, and it has long been the place where goods bound for the mainland are held, sorted and forwarded. The chips are not lost, damaged or hidden. They are sitting in commercial warehouses in one of the world's busiest logistics hubs, waiting on a piece of paper: large-scale customs clearance from Beijing.

The scale is hard to overstate. Six hundred thousand H200s is not a shipment — it is a significant fraction of Nvidia's H200 production. For comparison, the largest Chinese buyers have each received about 10,000. At that rate, clearing the whole stockpile through the current trickle would take years. That is why the number moved markets when Reuters published it: it turned an abstract policy dispute into a warehouse full of unsold inventory worth tens of billions of dollars.

The reversal: Washington opened the door, Beijing shut it

To understand how strange this situation is, rewind to the start of the year. The United States has controlled exports of advanced AI chips to China since 2022, and the H200 was on the restricted list. After US-China trade negotiations, the Trump administration changed course: in December 2025 it allowed H200 sales to approved Chinese customers, and a formal licensing framework followed in January 2026, moving the chip to case-by-case licensing review. The deal was presented as a controlled concession — American chips could flow to vetted Chinese buyers, under conditions, rather than not at all.

The conditions were onerous. Chips must be inspected in the United States before shipment, and a 25% tariff applies when they are brought back into the country; Nvidia is absorbing the full cost rather than passing it on to customers. Even so, the door was open. The US government had decided that selling a previous-generation chip to approved customers was acceptable, and the licensing machinery was running.

Beijing then introduced its own restrictions. Chinese authorities have allowed only limited volumes for some companies in recent months, while large-scale imports remain unapproved. The policy has a deliberate shape: according to a July 2026 report by The Information, Chinese authorities granted Alibaba, ByteDance, DeepSeek and others limited permission to buy H200s for AI model training, while requiring that domestically made processors be given priority for inference — the stage where trained models are put into actual service. The calculation is to allow some American chips at the training stage, where substitution is difficult in the near term, while nurturing a domestic chip ecosystem — including Huawei's Ascend 950 — at the inference stage, where the market is larger.

Some observers expect total approved volumes to come in below 200,000 units. That would leave roughly two-thirds of the Hong Kong stockpile with no approved route into the mainland at all, under current policy.

The timing adds a further political dimension. Key parts of the US-China trade agreement — including the suspension of Washington's Affiliates Rule and Beijing's rare-earth export measures — were due to expire around 9–10 November 2026. Both sides have since announced extensions of the relevant measures until 10 January 2027. The chip standoff is therefore unfolding alongside wider negotiations over trade and strategic resources, and any change in the H200 position is likely to be part of a larger bargain rather than a standalone decision.

Who got chips anyway

The block is not absolute. In recent months, Chinese authorities have allowed limited volumes for some companies. Several major internet companies, including Alibaba and ByteDance, have each secured about 10,000 H200s, and small quantities have entered through brokers and roundabout channels. But large-scale imports remain unapproved, and the gap between 10,000 and 600,000 is the story.

There are workarounds, and Chinese companies are using them. One source cited in the Reuters reporting said the authorities are steering companies to tap the computing power of chips installed in Hong Kong data centres over remote networks, while controlling physical shipments into the mainland. The chip stays in Hong Kong; the compute crosses the border as a service. That arrangement suits Beijing's purposes: Chinese companies get access to Nvidia performance without a physical import that would undercut the domestic-first policy.

Other workarounds reach further afield. Chinese AI developer Moonshot AI has explored renting European computing capacity to access advanced Nvidia hardware — reported discussions about clusters containing 8,000 or more Nvidia Blackwell GPUs, primarily for inference, with some interest in training capacity. Those were reported discussions rather than evidence of a completed lease, but they show the lengths to which leading Chinese developers are going. The irony is hard to miss: a country trying to build an independent AI industry has some of its leading developers looking outside the country for the computing power they need.

Meanwhile, the contrast with US allies is sharp. Sakura Internet began operating about 1,100 of Nvidia's Blackwell GPUs in February 2026 in containerised data centres at its Ishikari site in Hokkaido, and SoftBank added GB200 NVL72 systems to its enterprise cloud service in March 2026. While Chinese companies cannot import the previous-generation H200 in large volumes, Japanese companies are receiving the latest generation as normal. The customs barrier applies to China specifically, not to the region.

The grey market: triple the price, and a $35 billion workaround

With official supply blocked, prices are soaring on the informal market. As of late October 2026, servers equipped with eight H200s were being offered in mainland China for more than 6 million yuan — approximately $900,000 — roughly triple the level of a year earlier. Servers based on the next-generation Blackwell platform, which is subject to US export controls, were reportedly offered at more than 20 million yuan (approximately $3.0 million) per unit.

The six-million-yuan server is the most revealing number in this story. If Beijing's restrictions were successfully moving the industry toward domestic independence, why are companies willing to pay three times the previous price for foreign processors? The answer is simple: computing power has immediate economic value. AI companies cannot postpone their development plans indefinitely while waiting for local suppliers to catch up. They need processors now, because access to computing resources affects how quickly they can train models, develop applications and attract customers.

The informal market is therefore more than an inconvenience. It shows that the government's industrial policy is working against the purchasing decisions of its own largest technology companies. Instead of allowing demand to move naturally toward better domestic products, Beijing is creating artificial scarcity and encouraging expensive, less transparent supply channels.

Chinese Big Tech firms are also pursuing a different workaround at industrial scale. ByteDance, Alibaba and Tencent recently ordered more than 3 million of Nvidia's RTX Pro 5500 graphics cards — a lower-specification product that complies with US export rules. At list prices, the orders are worth more than $35 billion. Because order volumes exceed Nvidia's planned output, the company has reportedly notified customers and partners that it cannot guarantee delivery dates. The workaround demand is so large that it is straining supply of a product that was never meant to carry the AI market on its own.

The two embargoes, side by side

It helps to see the two restrictions next to each other, because they are often confused — and because the confusion is exactly what the viral graphic trades on.

	US export controls	China's import block
What is restricted	Exports of advanced AI chips (Blackwell barred; H200 allowed under licence)	Large-scale imports of H200s into mainland China
Since when	Controls since 2022; H200 licensing framework from January 2026	Large-scale clearance withheld through 2026; training-only permissions from July 2026
Who decides	US Commerce Department (BIS licensing)	Chinese customs and industrial authorities (MIIT, NDRC)
Who it hits	Chinese buyers of the newest chips; Nvidia's Blackwell sales	Chinese buyers of the previous-generation H200; Nvidia's H200 sales
Current status	Door open for H200s to approved customers; Blackwell still barred	Door shut for large volumes; ~10,000 per major buyer; total approvals expected below 200,000

The table makes the reversal visible. The US restriction and the Chinese restriction now point in opposite directions on the same product: Washington will let H200s out, Beijing will not let them in at scale. Both governments are using market access as a political instrument. American controls restrict what can be supplied, while Chinese import approvals determine what can actually enter the domestic market. The idea that Washington alone determines the availability of advanced AI chips in China is increasingly outdated.

What it costs Nvidia

The policy risk is already showing up in Nvidia's financial statements. According to the company's Form 10-Q for the second quarter of fiscal 2027, ended 26 July 2026, Nvidia recognised $400 million in charges during the first half of the fiscal year for excess H200 inventory and purchase obligations. The charges stem from the company being unable to sell even its licensed volumes in full because of Chinese government restrictions. During the same period, China-bound shipments under the new licences accounted for less than 1% of data centre revenue. In the filing, the company assessed that it cannot compete in China's data centre computing market without approvals from both governments.

The longer trend is a shrinking China business. Nvidia's Chinese revenue contribution has fallen from approximately 21.5% in fiscal 2023 to around 12.5% in early fiscal 2026. Beijing may read that decline as a sign of reduced technological dependence. But it also represents business lost from a market containing some of the world's largest potential AI customers — and the decline is driven by policy on both sides, not by Chinese customers choosing to buy less.

The next test comes on 17 November 2026, when Nvidia reports results for the third quarter of fiscal 2027. The revenue forecast sits at $108 billion at the midpoint, with a range of plus or minus 2%, and gross margin is guided at 74%, plus or minus 0.5 percentage points. Crucially, the revenue outlook excludes China data-centre computing revenue — a caveat that keeps the guidance from being read as an unrestricted statement about every regional opportunity. If Beijing were to approve large-scale imports, the Hong Kong stockpile would become sellable inventory almost overnight; if the block hardens, the $400 million in charges may be only the first instalment.

For investors, the distinction that matters is between demand, approval and delivery. The stockpile proves demand exists and that supply is ready. It proves nothing about whether the two will be allowed to meet. Anyone treating the entire 600,000-chip inventory as future revenue is skipping a step in the chain.

The ghost-company approvals

A second story broke two days after the Reuters report, and it complicates the picture further. A New York Times investigation published 7 November 2026 found that the US Commerce Department approved two Chinese companies to assemble or handle Nvidia H200 chips despite apparent ghost offices and ties to US-sanctioned entities.

The first is New Matrix, a Beijing firm founded approximately 17 months ago and approved to assemble H200 chips into AI servers. Reporters who visited its listed address in an industrial park on the northwest edge of Beijing found an empty building — a grey shell with wires dangling where lights should be — and property managers said they had no record of the company. Its owners are connected to Wise Road Capital, a Chinese private equity firm the Commerce Department's Bureau of Industry and Security placed on the Entity List in 2024. Two analysts cited by the Times said New Matrix may be a front for companies tied to that fund — their assessment, not a government finding, and no charges have been reported.

The second is Maginfra, approved in September 2026 as an H200 assembler. Reporters who visited its listed office found it deserted, and analysts cited by the Times link Maginfra to Inspur, a server maker sanctioned for supplying technology to the Chinese military, citing shared staff and patents as the basis.

Three distinctions matter here, and the reporting is careful about all of them. First, both companies were approved as assemblers or handlers, not necessarily as end-users of the chips — assembling H200s into servers is a different role from buying them to run AI workloads, and neither company should be described as an approved end-user buyer on the basis of this reporting. Second, whether any H200 processors physically reached either company is unconfirmed; the Times does not establish that either firm received a single chip. Third, an empty listed address is not proof of wrongdoing, and a corporate link to a blacklisted fund is not a finding by any agency that a licence was improper.

What the investigation does establish is that the vetting process cleared firms whose ownership traces to sanctioned entities. The approved list runs to roughly two dozen Chinese companies, according to people familiar with the approvals cited by the Times, and includes Baidu, BYD and Tesla Shanghai alongside the two assemblers. Shanghai Jiatang, an entity traced to state-owned Datang Telecom, which supplies technology to the Chinese military, was also identified as an approved H200 handler. Commerce has issued no public statement on the approvals, and has not publicly revoked them or confirmed it is reviewing them. Nvidia spokesman John Rizzo said it "makes no sense for the Chinese military to depend on American technology," pointing to Beijing's own restrictions on H200 imports as a mitigating factor.

The timing sharpens the irony. A Commerce rule filed on 5 November 2026 keeps the 50% Affiliates Rule — which extends Entity List limits to majority-owned subsidiaries — suspended until 10 January 2027. The suspension was extended two days before the Times investigation was published. The rule that exists to catch ownership links to blacklisted parties was switched off while licences were being issued to companies with exactly those links.

Why Beijing would block chips its own companies want

The hardest question in this story is the one Beijing's own policy raises: why would a government deliberately limit access to technology its largest companies are paying triple the price to obtain? There are two honest answers, and both deserve a hearing.

The case for the block starts with dependence. China has spent years trying to reduce its reliance on American technology, and Huawei's Ascend processors, along with other domestic semiconductor projects, represent a major part of that effort. The government's calculation is straightforward: as long as Chinese companies can easily purchase Nvidia's best available products, they have little financial incentive to invest heavily in domestic alternatives. Why spend money developing a less efficient processor when a more powerful Nvidia chip can be purchased instead? Restricting imports changes that calculation. ByteDance, Alibaba and Tencent must look more seriously at domestic suppliers, allocate money to local development and help identify weaknesses in Chinese hardware.

There is a genuine argument here. Building a competitive semiconductor industry requires more than government subsidies and impressive laboratory demonstrations. Domestic manufacturers need customers, production volume, software developers and real-world testing. A guaranteed domestic market could give Huawei the financial support and practical experience needed to improve its Ascend range. Over several years, this could produce a stronger industry with greater control over its own supply chain.

China also has a legitimate national security concern. A country building its future economy around technology designed and controlled by a geopolitical rival remains exposed to sudden policy changes. Washington has already demonstrated its willingness to use semiconductor exports as an instrument of foreign policy. Beijing has no reason to assume that American export rules will remain stable — the H200 itself was restricted, then allowed, within a single year.

The case against the block starts with readiness. The biggest problem with Beijing's approach is the assumption that Chinese companies can replace Nvidia quickly enough to avoid losing ground. Developing a competitive AI processor is extremely difficult. The physical chip is only one part of the problem. Companies also need reliable manufacturing, high-bandwidth memory, efficient connections between thousands of processors and software that allows developers to use the hardware effectively. Nvidia has spent years building an ecosystem around its CUDA software platform. Developers know how to work with it, data centres are designed around it, and many AI applications have been optimised for its hardware. Huawei may produce increasingly competitive processors, but convincing companies to move their entire computing infrastructure to another platform takes considerable time and money.

Forcing that transition before domestic alternatives are ready creates immediate costs. Engineers must spend time adapting software, fixing compatibility problems and testing performance instead of concentrating on improving AI models. Every month spent solving hardware problems is a month in which competitors elsewhere can develop and release new products. And the grey market is the market's verdict on the policy: companies are paying roughly $900,000 for an eight-GPU server on the informal market rather than waiting for domestic chips to catch up. If the restrictions were successfully moving the industry toward domestic independence, buyers would not be paying triple for foreign processors.

There is also a cost that revenue figures cannot measure. China could reduce its purchases of Nvidia products while remaining dependent on foreign semiconductor equipment, advanced manufacturing processes and other parts of the global chip supply chain. A falling Nvidia sales figure does not automatically mean Huawei has become competitive. Strategic independence may be worth pursuing. Delaying the development of your own AI industry to achieve it is a much harder decision to defend.

What happens next

Three dates and one decision will shape what happens next. The first date is 17 November 2026, when Nvidia reports third-quarter results. The earnings call will be the first time the company has to explain the Hong Kong stockpile, the $400 million in charges and the exclusion of China data-centre revenue from guidance in its own words, to its own shareholders. Watch for any change in the language about China — and for whether analysts press on the size of the warehoused inventory.

The second date is 10 January 2027, when the current extensions of the trade measures expire — including the suspension of the Affiliates Rule and Beijing's rare-earth export measures. The H200 question is unlikely to be settled in isolation; it will be part of whatever bargain the two sides strike over trade, rare earths and technology controls. A green light for large-scale H200 imports would most plausibly arrive as one element of a wider deal, not as a standalone customs decision.

The third thing to watch is not a date but a signal: whether any verified intermediary publishes a test, a delivery or a customs document. A GPU-Z screenshot, a delivery record or a customs filing would change this story from a listing into a measurable flow overnight. Until something like that appears, the honest description is that roughly 600,000 chips are warehoused, a small number have entered, and the rest are waiting on a political decision.

What a green light would unlock is straightforward to describe and hard to overstate. The stockpile would become sellable inventory against reported mainland demand, without Nvidia having to find a single new customer. The grey-market premium would collapse. Alibaba, ByteDance and Tencent could stop paying triple on the informal market and stop ordering millions of lower-specification cards as a substitute. And the $20–30 billion of warehoused value would start converting into revenue — the largest single swing factor in Nvidia's China business. What a continued block would mean is equally clear: more charges, more workaround orders, more capital spent on expensive foreign compute, and a domestic chip industry forced to prove itself under artificial scarcity rather than open competition.

Verdict

The chips are real, the warehouses are real, and the blocker is Beijing. Washington spent four years building a wall around advanced AI chips, then opened a gate in it for the H200 — and China has chosen not to walk through. That is a remarkable reversal, and it is the story the viral graphic gets backwards: the country that wants the chips is the one keeping them out, and the country that restricted them is the one offering to sell.

Whether Beijing's gamble pays off depends on a race it cannot fully control — between Huawei's improving processors and the months its own AI companies lose waiting for them. The 600,000 chips in Hong Kong are the scoreboard for that race: every month they sit there is a month Chinese companies pay more for less compute, and a month Nvidia carries inventory it cannot sell. Watch the customs stamp, not the headline number. The stamp is the whole story.

Frequently Asked Questions

Are the 600,000 Nvidia chips seized?

No. Nothing has been seized or confiscated. The chips are sitting in commercial warehouses in Hong Kong, awaiting customs clearance to enter mainland China. They belong to the normal supply chain — Nvidia and its distributors — and no seizure, raid or confiscation has been reported by any outlet.

Who is blocking the chips — the US or China?

China. The United States cleared H200 exports to approved Chinese customers after trade negotiations, with a case-by-case licensing framework in place since January 2026. Beijing is withholding large-scale customs clearance as part of a push for semiconductor self-reliance. The US still bars the newer Blackwell chips, but the H200s in Hong Kong are held up by China's import approvals, not by American export controls.

How many chips are there, and what are they worth?

Roughly 600,000 H200s, according to Reuters, citing multiple sources — the figure is an estimate, not an exact count. At list prices the stockpile is worth an estimated $20 billion to $30 billion. Alibaba and ByteDance have each secured about 10,000 chips, and total approved volumes are expected to come in below 200,000.

Can Chinese companies buy H200s at all?

Yes, in limited amounts. Alibaba, ByteDance, DeepSeek and others have limited permission to buy H200s for AI model training, while domestic processors get priority for inference. Small quantities have also entered through brokers and roundabout channels, and companies can access Hong Kong-based compute remotely. What is blocked is large-scale physical import into the mainland.

What is an H200?

The H200 is Nvidia's previous-generation flagship AI accelerator, built on the Hopper architecture with 141GB of high-bandwidth memory. It is designed for training and running large AI models. It has been succeeded by the Blackwell generation, which the US bars from China — which is why the H200, older but still highly capable, is the chip Chinese companies want most.

What happens next?

Nvidia reports third-quarter results on 17 November 2026, with guidance that excludes China data-centre revenue. The current trade measures, including the suspension of the Affiliates Rule, run to 10 January 2027. A green light from Beijing would unlock the stockpile against existing demand; a continued block would mean more inventory charges for Nvidia and more grey-market premiums for Chinese buyers.

Sources

• Reuters — "Nvidia chips pile up in Hong Kong as China withholds import approval" (Exclusive, 5 November 2026, multiple sources). Direct reuters.com URL could not be established from this machine — reuters.com returns 401 to this VM and the search index never surfaced the article URL — so the reporting is cited via the 200-verified coverage below. No URL has been guessed.
• BigGo Finance — "Nvidia's 600,000 H200 Chips Stuck in Hong Kong as China Blocks Mass Customs Clearance in Push for Semiconductor Self-Reliance" (7 November 2026, reporting the Reuters exclusive): https://finance.biggo.com/news/9cfaec47-9327-4e6a-9896-1cc956d084cc
• Adalytica — "Nvidia H200 chips await China import approval" (9 November 2026, citing Reuters): https://adalytica.com/news/nvidia-h200-china-import-approval-2e4529f0
• Kozminski Techblog — "The Reverse Embargo: Why Beijing Is Blocking Nvidia Chips While China Falls Behind in the AI Race" (9 November 2026): https://techblog.kozminski.edu.pl/2026/11/09/the-reverse-embargo-why-beijing-is-blocking-nvidia-chips-while-china-falls-behind-in-the-ai-race/
• ad-hoc-news — "Nvidia's China Revenue Is Sitting in a Warehouse, and Wall Street Is Still Looking for Buyers" (5 November 2026): https://www.ad-hoc-news.de/boerse/news/unternehmensnachrichten/nvidia-s-china-revenue-is-sitting-in-a-warehouse-and-wall-street-is-still-looking-for-buyers/70464490
• New York Times — "U.S. Cleared Nvidia Chips to Chinese Firms With Empty Offices and Troubling Ties" (7 November 2026, investigation). Direct nytimes.com URL could not be established (paywalled); cited via the 200-verified summaries below.
• TFTC — "Commerce Cleared Ghost-Company H200 Chip Handlers" (8 November 2026, summary of the New York Times investigation): https://www.tftc.io/commerce-ghost-companies-nvidia-h200-chips-export-controls
• SantAge AI — "Two Chinese Nvidia H200 Licensees Had Empty Offices, NYT Finds" (8 November 2026): https://santageai.com/news/2026/11/08/nvidia-h200-china-licensees-empty-offices
• China Factor — "H200 licensee had empty office: NYT" (9 November 2026): https://chinafactor.news/2026/11/08/u-s-cleared-chinese-firm-with-empty-office-to-assemble-nvidia-h200-servers-nyt-reports/
• Nvidia — H200 product page (141GB HBM3e): https://www.nvidia.com/en-us/data-center/h200/

Related reading: Modded RTX 5090 with 96GB of VRAM Listed in China · Jensen Huang resigns from Nvidia · Nvidia SPAN: home mini data centres, explained