About 600,000 Nvidia H200 chips are sitting in storage in Hong Kong, waiting to cross into mainland China, Reuters reported on November 5, citing two people familiar with the matter. The US licensed the sales. Beijing is the one holding them back. A year ago almost everyone assumed the chokepoint on Chinese AI chips would be American. Now it runs the other way.
How we got here
Trump said on December 8, 2025 that approved Chinese customers could buy the H200. The Commerce Department’s Bureau of Industry and Security followed on January 13 with case-by-case licensing for the H200 and AMD’s MI325X. The conditions were tight: proof the exports wouldn’t eat into supply for American customers, compliance screening by the Chinese buyer, independent testing in the US and inspection on US soil before shipment. That inspection step put each chip through a 25% US import tariff, which Nvidia has said it can’t pass on.
Licenses started allowing small shipments in February. Several large Chinese internet companies each got roughly 10,000 H200s over the following months, and smaller lots came in through intermediaries. Then Beijing stopped approving large imports.
Nvidia’s own filing shows the cost. In its 10-Q for the quarter ended July 26, it took a $0.4 billion charge for excess H200 inventory and purchase commitments, said it had shipped only a fraction of permitted volumes, under 1% of data center revenue, and said it was effectively shut out of China’s data center market without approval from both governments.
Why Beijing is holding back
The public reason is dependence. Beijing wants its tech companies building on Huawei’s Ascend line and other domestic accelerators, and every H200 that gets installed is a reason for a Chinese lab to keep writing for Nvidia’s CUDA software instead. Domestic chips still trail the H200, especially once Nvidia’s software is counted, so a free market would send buyers straight back to Nvidia.
There’s also leverage. 600,000 chips in a Hong Kong warehouse is a pile of US revenue that Beijing can release or keep frozen. Chip access has been part of US-China trade bargaining since the H200 decision, and holding back approvals costs China less than it costs Nvidia.
The market inside China shows the squeeze. By the end of October an eight-card H200 server was reportedly selling on the mainland for more than 6 million yuan, about three times the price a year earlier. Servers built on Blackwell chips, which the US still bars, went for more than 20 million yuan. Chinese companies are also buying around the problem: Reuters reported in October that ByteDance, Alibaba and Tencent had ordered more than 3 million RTX PRO 5500 cards, workstation GPUs that need no license, worth possibly more than $35 billion.
What to watch
A warehouse this size doesn’t stay secret, and it raises its own questions: who owns the stock, who pays to hold it, and how long before some of it leaks across the border through the same intermediaries that moved the first lots. For Washington, the risk has flipped. Officials spent years worrying about chips getting into China. The new worry is that Chinese restrictions protect Huawei’s home market while the chips sit a few kilometers from Shenzhen.
The signal to wait for is a Chinese import approval for one of the big platforms. If Beijing releases a large tranche to ByteDance or Alibaba right after a trade round, the chips were a bargaining chip. If approvals stay frozen while Huawei ships more Ascend clusters, Beijing has decided to accept slower AI progress as the price of not depending on Nvidia.