The absence of an export license from the US Department of Commerce means NVIDIA cannot provide its latest high-end H100 and H200 GPUs to entities in China. This license is notoriously difficult to obtain, especially not for hundreds of these GPUs. However, these GPUs seem to be readily available in China, with an anonymous entrepreneur boasting of importing over two dozen Supermicro servers using H200 GPUs, housing hundreds of these chips.
This unnamed entrepreneur is believed to have acquired up to 25 Supermicro servers, each containing 8 NVIDIA H200 GPUs, totaling a massive 200 Hopper GPUs. While 200 H200 GPUs are insufficient for training large, complex language models, they are good enough for research organizations to conduct experiments and support startups in software development. Companies can initiate their projects on H200 hardware before transitioning to remote cloud data centers for operations.
Given the high demand, NVIDIA’s H200 GPUs with 141GB of HBM3 memory are hard to come by. For instance, a hard-to-find H200 server can be ordered from some companies in the US, but the estimated wait time is between 4 and 6 weeks, and prices are not publicly disclosed. In Europe, a motherboard housing eight H200 GPUs costs around $273,000, excluding VAT. A fully equipped H200 server could cost over $300,000.
Despite US restrictions preventing Chinese entities from purchasing advanced AI and HPC processors for data centers, a network of buyers, sellers, and shippers is circumventing these export controls. As of August, more than 70 distributors were openly marketing these restricted processors online, with many offering delivery of loose GPUs or entire servers within a few weeks.
NVIDIA supplies its H100 and H200 processors to companies like Dell, HPE, and Supermicro, which manufacture and sell AI servers. These companies may procure surplus processors, and these components can find their way into unofficial channels. While all the companies claim to comply with US export regulations and address illegal activities when discovered, sales to smaller retailers could slip under their radar and end up in Chinese networks.
Shipping a server weighing dozens of kilograms into China is no easy feat. Therefore, the ability to smuggle in a significant number of bulky NVIDIA H200 servers indicates the existence of a well-oiled distribution system.