China stalls 600,000 Nvidia H200 chips to prioritize domestic hardware

The Hong Kong bottleneck

A massive stockpile of approximately 600,000 Nvidia H200 AI chips is currently sitting in Hong Kong. While the U.S. government has permitted the export of these high-performance semiconductors, the Chinese government has not approved their large-scale entry into the mainland. This represents the first time the specific quantity of these chips held in the region has been publicly reported.

Major Chinese internet companies are currently navigating these hurdles to secure the hardware necessary for their AI ambitions. Reports indicate that several large firms have managed to secure about 10,000 H200 units each through various channels. However, these quantities remain small compared to the total volume of chips currently waiting in the port.

Small amounts of the H200 chips continue to enter the country through indirect channels and various brokers. Despite these small trickles, authorities are maintaining a firm stance against approving large-scale imports. This creates a significant backlog for companies that require high-performance hardware to maintain their competitive edge in the global AI market.

A policy of self-reliance

The current situation is the result of a collision between U.S. export controls and China's internal goals for chip self-sufficiency. Since 2022, the United States has limited China's access to advanced AI semiconductors and the equipment needed to make them. These restrictions were designed to slow down China's progress in high-end technology development.

China's own policies are now creating a secondary layer of restriction on top of the U.S. bans. If the Chinese government allows a flood of U.S. chips into the country, domestic companies might continue to favor Nvidia products because they outperform local alternatives. This would undermine the government's goal of expanding the market share and technical capabilities of Chinese chipmakers.

To manage this, Chinese authorities are attempting to distinguish between different types of AI tasks. They are allowing limited U.S. chip access for AI model training, which is the process of processing data to improve a model's performance. Conversely, they want domestic chips to be used for inference, which is the task of using a trained model to generate actual outputs or answers.

The performance gap

The core issue for Chinese tech firms is that domestic chips currently do not match the performance of Nvidia's offerings. The H200 is not Nvidia's newest chip, but it is still considered superior to what Chinese substitutes can currently provide. It is particularly valuable when paired with Nvidia's software ecosystem for training large-scale models.

High demand from major players like Alibaba, ByteDance Ltd., and Tencent is driving the scramble for these high-performance units. Because the H200 is so effective, these companies are looking for any way to secure them despite the import limits. The lack of a viable domestic alternative makes it difficult for these firms to pivot away from U.S. technology quickly.

The government's strategy aims to balance these needs by securing enough performance to keep AI development moving while forcing the growth of the local industry. However, this creates a burden for companies that need to scale their services. They must find a way to balance the need for high-end performance with the regulatory requirement to support domestic hardware.

Surging gray market costs

The intersection of U.S. export controls and Chinese import limits has caused a spike in prices within the unofficial chip market. In mainland China, a server equipped with eight H200 chips was offered for sale at more than 6 million yuan in late October. This price is roughly triple what the same server cost just one year ago.

The situation is even more extreme for next-generation hardware that is subject to even stricter U.S. export controls. Servers equipped with Nvidia's Blackwell chips were reportedly offered for more than 20 million yuan per unit. These high costs reflect the scarcity and the difficulty of moving these chips across borders legally and efficiently.

Analysts suggest that China's push for chip self-reliance could significantly raise development costs for domestic AI companies in the short term. If the government continues to restrict imports while domestic chips remain less capable, securing the necessary hardware will become increasingly difficult. The future of Chinese AI development may depend heavily on whether the government eventually approves larger import volumes.