Beijing is restricting the import of Nvidia H200 AI chips to those used for the training of AI models, barring their use for deployment in inference tasks. The shift in policy is intended to increase the reliance on domestic hardware for the operational phase of AI deployment and reduce dependence on U.S. technology.
The move has left approximately 600,000 H200 processors stalled in Hong Kong. This bottleneck has pushed gray-market prices for an eight-card H200 server to 6 million yuan, or about $900,000, as of late October, which is roughly three times the cost from a year earlier. While large shipments are blocked, several of China’s largest internet companies have received smaller approved allotments of about 10,000 chips each.