Intense appetite for artificial intelligence computing hardware in China has driven costs for Nvidia’s B300 servers to approximately $1 million per unit, according to industry insiders, as US export restrictions eliminate illegal supply channels.
The cost of Nvidia’s most sophisticated server technology, essential for AI operations, has climbed throughout this year but accelerated dramatically after underground markets faced increased enforcement pressure, four anonymous sources revealed.
Chinese technology firms continue driving strong demand for computing power, though many companies avoid directly owning Nvidia equipment on their financial records due to concerns about US sanctions exposure, the sources explained.
The sources requested anonymity given the sensitive nature of the topic. This marks the first reporting of the million-dollar price point.
When contacted by Reuters, Nvidia confirmed the B300 cannot be legally sold in China and emphasized that authorized partners must maintain strict regulatory compliance.
“As systems become increasingly large and complex, unlawful diversion is a recipe for failure,” the company stated.
“Nvidia does not provide any service or support for such systems, and the enforcement mechanisms are rigorous and effective.”
Within the United States, a B300 server containing eight B300 graphics processing units costs approximately $550,000, representing an increase from roughly $500,000 in late 2023, two sources indicated.
The near-doubling of Chinese prices from about 4 million yuan last year demonstrates how supply shortages created by stricter US export controls inflate costs.
This situation emerges as Chinese technology companies seek the most efficient hardware for generating tokens – the fundamental text units processed by AI systems – to profit from their models and computing infrastructure.
The supply shortage intensified after US prosecutors in March charged Yih-Shyan “Wally” Liaw, a co-founder of Nvidia partner Supermicro, sources noted.
Companies unable to afford outright purchases are exploring rental arrangements, with monthly costs reaching 190,000 yuan for one-year contracts.
Chinese AI models expanded their portion of worldwide token usage to 32% in March 2026 from 5% the previous year, boosted by improvements in coding and autonomous capabilities, according to Morgan Stanley research.
MiniMax, Zhipu and Alibaba’s Qwen each saw token usage increase six to seven times in February and March compared to December, the investment bank reported.
Nvidia’s B300, featuring 288 GB of high-bandwidth memory, provides 14 petaFLOPS of computing performance at FP4 precision, positioning it among the most capable chips for AI inference operations.
Nvidia and partners like Supermicro started delivering the chip last September.
Questions about H200 chip exports have also contributed to rising B300 prices.
Although both governments approved H200 exports, shipments to China remain stalled as officials disagree over sales conditions.
Technology giant Huawei and other Chinese AI chip manufacturers are capitalizing on this dispute while attempting to challenge Nvidia’s dominant 55% market share in China, where competitor AMD holds 4%.
