By Che Pan, Liam Mo and Laurie Chen
BEIJING, April 30 (Reuters) - Strong demand for AI computing equipment in China has nearly doubled prices for Nvidia's B300 servers to about 7 million yuan ($1 million) each, industry sources said, as a crackdown on chip smuggling dries up black-market supply.
Prices of Nvidia's most advanced and powerful server, critical for artificial intelligence tasks, have climbed since early this year, but rose sharply after the grey market, a key supply channel, came under
pressure, the four sources said.
The price surge is also being driven by robust computing demand from Chinese technology companies, even as many avoid holding Nvidia hardware directly on their books for fear of exposure to U.S. sanctions, the sources said.
They spoke on condition of anonymity as the matter is a sensitive one. Reuters is the first to report the million-dollar price tag.
Responding to questions from Reuters, Nvidia said the B300 was restricted from sale in China, and its partners needed to be committed to strict compliance.
"As systems become increasingly large and complex, unlawful diversion is a recipe for failure," it warned in a statement.
"Nvidia does not provide any service or support for such systems, and the enforcement mechanisms are rigorous and effective."
SCARCITY PREMIUM DRIVEN BY TIGHTER U.S. CURBS ON EXPORTS
A B300 server, which houses eight B300 GPUs, is priced at about $550,000 in the United States, up from around $500,000 late last year, two of the sources said.
The near-doubling of prices in China, from about 4 million yuan late last year, reflects a scarcity premium driven by tighter U.S. curbs on exports.
It comes as Chinese tech giants scramble for the most cost-efficient hardware for generating tokens, the basic units of text processed by an AI model, to monetise their models and computing infrastructure.
The scarcity followed U.S. authorities' March prosecution of Yih-Shyan "Wally" Liaw, a co-founder of Nvidia partner Supermicro, the sources added.
Some companies finding the price rise has put purchases out of reach are instead exploring options for rentals, which have risen as high as 190,000 yuan a month on a one-year short-term contract.
Chinese AI models nearly tripled their share of global token usage to 32% in March 2026 from 5% a year earlier, driven by advances in coding and agentic capabilities, Morgan Stanley said in a note.
MiniMax, Zhipu and Alibaba's Qwen each reported token usage rising as much as six to seven folds in February and March compared with December, the U.S. investment bank said.
AMONG THE MOST POWERFUL FOR AI INFERENCE TASKS
Nvidia's B300, packed with 288 GB of high-bandwidth memory, delivers 14 petaFLOPS of computing power at FP4 precision, ranking it among the most powerful chips available for AI inference tasks.
Nvidia and partners, such as Supermicro, began shipping the chip last September.
Uncertainty surrounding exports of H200 chips have also fuelled the recent surge in the price of B300.
Despite receiving approvals from both governments for exports, the H200 has yet to be shipped to China as the two sides remain at odds over the conditions governing its sale.
Tech giant Huawei and other Chinese AI chipmakers have moved to exploit this disagreement, as they seek to erode Nvidia's market-leading share of 55% in China, where competitor AMD has a share of 4%.
($1=6.8302 Chinese yuan renminbi)
(Reporting by Che Pan and Liam Mo; Editing by Anne Marie Roantree and Clarence Fernandez)
