Musk Praises NVIDIA's AI Architecture, Projects SpaceX's Compute Capacity to Reach 10 GW by 2024

In the same computing power leasing market, a clear dividing line in contract duration is emerging: the Huawei Ascend 910B/910C series is generally leased starting from one year in mainland China. NVIDIA H100/H200 leasing resources typically requires a minimum three-year lock-in contract, with concessions for a five-year contract still limited. Under export controls, computing power has shifted from a "commodity" to a "scarce resource“, contract duration is precisely the price of that scarcity.

SMM learned that a certain intelligent computing service provider reported that the current three-year lease quotation for H100 with cabinets and electricity was already above 75,000 yuan/month. For RTX4090 8-GPU servers, they still had dozens of units in stock, with the 1TB memory version’s monthly rent of 7,200 yuan as their floor price. The market price for similar 512GB memory versions had risen to 8,000 yuan/month. SMM also learned that the quotation range for H200 was 120,000-135,000 yuan/month and showed an upward trend this week. Rental prices for high-end cards remained in an upward channel. The service provider said that there were many intermediaries marking up prices in the market, and first-hand resources were scarce.

SMM has learned that a computing power service provider in Southern China has released 20 RTX 5090 8-GPU servers for lease, with a monthly rental quote of 13,000 yuan per unit. This batch of machines is deployed as standalone units without networking interconnection, suitable for lightweight scenarios such as distributed inference or rendering. SMM believes that the non-networked configuration offers high flexibility and a low barrier to entry, but the lack of cluster interconnection means limited adaptability to large-scale training and distributed tasks; lessees should evaluate based on their specific scenarios.