NVIDIA B300 Server Prices Soar, Nearly Doubling Since Late Last Year
The transaction price of NVIDIA B300 AI servers in mainland China has surged to approximately 7 million RMB (about $1 million) per unit, nearly doubling from around 4 million RMB last year. For reference, the same server with 8 B300 GPUs costs about $550,000 in the US market, a slight increase from $500,000 last year, representing a price difference nearly double that of the Chinese market.
The price surge has already prompted some capital-constrained AI companies to shift from purchasing to leasing. It is reported that the price of a one-year lease contract for a B300 server has risen to 190,000 RMB per month, and there is still a queue to wait for resource availability.
The market pricing is supported by the hardware specifications of the B300, which is equipped with 288GB of HBM3e high-bandwidth memory and can deliver 14 petaFLOPS of computing power under FP4 precision, making it one of the most powerful chips currently available from NVIDIA for AI inference tasks.
The uncertainty surrounding H200 exports to China is also driving up the price of the B300. Although the US government has approved export licenses for the H200 to China, the two sides still disagree on specific terms and have not yet made actual shipments. This stalemate has made the B300 even scarcer in the market.
Chinese domestic AI chip manufacturers are gradually increasing their competitiveness. Huawei has established a clear product iteration roadmap, planning to launch four Ascend series chips in phases between 2026 and 2028, including the Ascend 950PR in the first quarter of 2026.
Currently, Hygon, Cambrian, Alibaba’s T-Head, and other manufacturers are also simultaneously advancing the iteration of AI training and inference chips.