600,000 Nvidia H200s Sit in Hong Kong as Beijing Withholds Import Approval

On 5 November 2026, Reuters reported that about 600,000 Nvidia H200 GPUs are sitting in Hong Kong, licensed by Washington but waiting on an import approval from Beijing that has not come. The report, summarised by TNW and Investing.com via Yahoo Finance, cites two people familiar with the matter and says the total number of foreign chips China eventually approves could end up below 200,000. It is the clearest sign yet that the bottleneck on Nvidia's China business has moved from Washington to Beijing.

What happened

The H200 is a Hopper-generation part, one step behind the Blackwell line that the US still does not allow into China. President Trump allowed limited H200 sales in December 2025, and in January 2026 the US began reviewing H200 licences for China case by case, according to TNW's account of the Reuters story. Beijing approved a first batch of imports for ByteDance, Alibaba and Tencent in January, then largely stopped.

According to the Reuters sources, several large Chinese internet companies have each received about 10,000 H200s in recent months. The rest of the inventory waits in Hong Kong. Nvidia, China's industry ministry and its main economic planning agency did not respond to Reuters' requests for comment.

Beijing's motive is not hidden. TNW's summary says Beijing "wants firms such as Alibaba, ByteDance, Tencent and DeepSeek to use Chinese-made chips instead", especially for inference. In September Huawei had pulled its Ascend 960 roadmap forward, which gives the policy a domestic product to point at.

The diplomatic track has not moved chips either. Before the 24 September Trump-Xi meeting, CNBC reported that US Trade Representative Jamieson Greer said chip controls were not on the agenda in the preliminary talks; the dialogue was framed around AI safety and an incident channel instead.

What Nvidia's own filing says

The best primary source is Nvidia's 10-Q for the quarter ended 26 July 2026, filed on 26 August. Three passages matter:

• The US "granted licenses that would allow us to ship small amounts of H200 products to specific China-based customers, but such sales were restricted by the PRC government."
• In the first half of fiscal 2027 Nvidia took a $0.4 billion charge on H200 excess inventory and purchase obligations "as the demand for H200 products diminished".
• The licences require the H200s to pass an inspection process in the United States before shipment, so they are subject to a 25% tariff on importation into the US. Nvidia says it has been unable to pass any of that tariff on to customers.

The same filing puts the stakes in proportion. Data Center revenue was $89.0 billion for the quarter, up 117% year on year, driven by Blackwell Ultra, and Hopper shipments to China were less than 1% of it. Vera Rubin began production shipments in the current quarter. China is a footnote in that income statement, not a pillar.

The numbers

Why it matters

Export controls were designed on the assumption that China would buy everything the US let it buy. That assumption has broken. Washington now permits a Hopper-class part, collects a 25% tariff on the round trip, and Beijing declines most of it to protect its own chip industry. Both governments are rationing the same product from opposite ends.

The prices tell you the rationing is real. A server that costs roughly three times what it did a year ago, and Blackwell machines on offer at over 20 million yuan despite a US ban, mean demand inside China is not satisfied by domestic silicon. The reported RTX Pro orders point the same way: Chinese buyers are taking whatever compliant Nvidia parts they can get while the H200 question stays open.

For Nvidia, the direct financial exposure is small. The filing says China Hopper shipments were under 1% of Data Center revenue, and the $0.4 billion charge is a rounding error against $89.0 billion a quarter. The bigger long-run risk, in my view, is a large market learning to build models on someone else's software stack.

What sceptics say

Treat the 600,000 figure carefully. It comes from two unnamed sources, Nvidia has not confirmed it, and it may count chips at different stages of logistics. The 10-Q describes shipments as "a fraction of the allowed shipments", which fits a stockpile story but does not quantify it.

There is also a reasonable argument that the stockpile is not Nvidia's problem at all. The filing says demand for H200 "diminished" and the charge has already been taken. If China never approves the rest, the cost is mostly sunk, and Blackwell Ultra and Rubin demand elsewhere absorbs the capacity. Meanwhile the gray-market prices suggest that some volume moves regardless of either government, which is its own policy failure.

Finally, Beijing's position may be a negotiating stance rather than a permanent ban. Approvals came in January; they could come again after the next round of trade talks.

What it means for engineering and DevOps teams

My take, as someone who rents and operates GPU capacity rather than trades it: this story barely touches the supply most Western teams buy from. Nvidia's filing says it is "currently experiencing certain supply constraints" while it ships both Blackwell and Rubin, and H200s licensed for Chinese customers are not going to show up in your cloud quota.

Where it does touch you is compliance. If you run a GPU cloud, an inference API or a reseller business, export rules now come with inspection, certification and know-your-customer obligations, and CNBC notes that Chinese firms have reportedly accessed compute remotely through Southeast Asian data centres. Remote access is exactly the route regulators are watching. Write down who your largest GPU tenants are and where their parent companies are headquartered before someone asks you to.

Second, vendor concentration risk now includes geopolitics on both sides. A Chinese team that bet on H200 availability got stuck; a Western team that bets everything on one accelerator family carries a quieter version of the same risk. Portable serving stacks, as covered in my piece on LLM inference engines, are cheap insurance. And if you are pricing capacity, the neocloud economics matter more to your bill than anything happening in Hong Kong.

FAQ

Can Nvidia legally sell H200 chips to China in 2026?

Yes, in small amounts. Nvidia's 10-Q says the US began granting H200 licences for specific China-based customers in February 2026, with US inspection and a 25% tariff on importation into the US. Blackwell remains barred. The binding constraint is now Beijing's import approval.

Why is China blocking Nvidia H200 imports?

Reporting attributes it to Beijing's push for domestic chips, particularly for inference, and a reluctance to deepen reliance on a US supplier that Washington can regulate. China has not published a formal rule explaining the hold.

How much does China matter to Nvidia's revenue now?

Very little on current numbers. Nvidia said Hopper shipments to China were less than 1% of Data Center revenue in the quarter to 26 July 2026, when Data Center revenue was $89.0 billion.

Related: GPU Clouds in 2026: H100s at $2.59 an Hour, $104B of CoreWeave Backlog on $35B of Debt, and a Six-Year Bet on Silicon, LLM Inference Engines 2026: vLLM vs SGLang vs TensorRT-LLM vs Dynamo and GPUs on Kubernetes Explained (2026): Device Plugin, MIG, Time-Slicing, DRA.