[SMM Computing Power] Huawei 910B2 Exit Dilemma: Monthly Rent 13,500, "Equipment Relocation" Bottlenecks

SMM July 3 News:

Pricing side, a certain operator in the Yangtze River Delta is currently quoting a Huawei 910B server (4×Kunpeng 920 + 8×Ascend 910-B2 64G, 1TB DDR4, NVMe + RoCE v2 networking) at 13,500 yuan/unit/month, equivalent to 2.34 yuan per card per hour. Another operator's 910-B2 specification is quoted at about 15,000 yuan/month, equivalent to 2.61 yuan per card per hour, showing strong price competitiveness in the domestic computing power leasing market. Online public pricing side, the standard 910B specification has a monthly rent of around 22,000 yuan, equivalent to 3.82 yuan per card per hour; however, factoring in annual rental discounts and other considerations, actual transaction prices could have room to decline.

Demand and clearance side, SMM has learned that overall sales of the 910-B2 are struggling. Only enterprises with "high vacancy rates and needing server occupancy for digestion" show purchasing interest. A previous flash report indicated that structural obstacles exist in clearing the 910-B2: interested parties' existing equipment is deployed in their own IDCs, and new leased resources need to be integrated into a unified cluster for coordinated scheduling; however, the operator's resources are located in off-site intelligent computing centers, and constrained by equipment attributes, off-site management (equipment relocation) is difficult to implement, directly stalling transaction progress. Regarding the 910C, a certain operator previously had a batch that has been cleared. According to online information, Huawei plans to double the production of the 910C in 2026, with a clear trend of resources tilting toward the C-series.

SMM analysis shows that the 910 series presents a differentiated landscape of "tight supply of 910C and sluggish sales of 910B2." The core pain point of the 910-B2 is not price—13,500–15,000 yuan/month is not considered high in domestic computing power—but rather the lack of reasonable absorption scenarios on the demand side. If demand for domestic large-model inference continues to grow in the future, the 910-B2's advantage of being positioned for both training and inference could be repriced; however, in the short term, it remains a buyer's market.