Roughly 600,000 Nvidia H200 artificial intelligence chips are sitting in Hong Kong, waiting for permission to enter mainland China, according to people familiar with the matter cited by Reuters this week. The holdup is not in Washington this time but in Beijing, which has not approved large-scale imports despite strong demand from Chinese technology companies.
Neither Nvidia nor the Chinese regulators involved commented, and the stockpile figure has not been confirmed by the company. Sources said Beijing has let some chips through in recent months, with several large Chinese internet companies each receiving about 10,000 units, while smaller quantities have arrived through intermediaries and indirect routes.
A deal approved on one side only
The H200 was once off-limits to China under US export controls. In January, the US Commerce Department moved to case-by-case reviews for exports of the H200 and AMD’s MI325X, provided buyers and sellers meet a list of conditions. These include independent testing in the United States, customer screening by the Chinese buyer, and a 25 percent levy collected by the US side.
Licences issued from February allowed small shipments, but Nvidia has told investors that Chinese restrictions kept it from selling the full amount it was cleared for. In its filing for the quarter ended in July, the company said H200 shipments to China were under one percent of its data centre revenue, and it took a $0.4 billion charge for excess H200 inventory and purchase commitments. Its conclusion was blunt: without approval from both governments, it is effectively shut out of China’s data centre market.
Beijing’s caution is linked to concern about depending on American technology, and to its push to build up domestic chipmakers. Yet Chinese alternatives have not matched the H200 for training large AI models, especially when paired with Nvidia’s software tools. That gap shows up in prices. By the end of October, an eight-chip H200 server was being offered in mainland China for more than 6 million yuan, around $900,000 and roughly three times its price a year earlier. Servers with Nvidia’s newer Blackwell chips, which remain under US controls, were listed at more than 20 million yuan.
Chinese firms are also turning to less powerful products that fall outside the controls. ByteDance, Alibaba and Tencent were reported last month to have ordered more than 3 million of Nvidia’s RTX Pro 5500 workstation cards, worth over $35 billion at list prices.
What it means here
The episode is a reminder that AI computing power is now rationed by politics as much as by price. Chips that are physically available, even sitting in a warehouse a short flight away, can still be out of reach for years.
Bangladesh is not caught in the US-China dispute directly, but it buys into the same tight market. The government has opened a shared GPU cloud facility for researchers and has said it wants to launch a “sovereign AI” system within a year. Both depend on hardware whose cost and availability are shaped by decisions in Washington, Beijing and Nvidia’s order book. Prices of ordinary computers have also risen this year as AI demand swallows memory supply.
For planners in Dhaka, the practical lesson is to budget for volatility, avoid relying on a single supplier or route, and weigh renting cloud capacity against buying hardware outright. For universities and startups, sharing national GPU resources efficiently may matter more than owning the newest chips.
