[SMM Computing Power Express] 128 Units of 5090 Futures Delivered in Batches; Lessor Explores a Five-Year Long-Term Agreement

SMM learned that 64 H100 eight-GPU servers were released in North China, with monthly rent at 75,000 yuan per unit (four-year fixed term) and 76,000 yuan per unit (five-year fixed term). Each server was configured with 8×80G H100 GPUs, an Intel 8468 CPU, 2 TB DDR5 memory, and a 400G network card. The compute cards were interconnected at high speed via an IB switch. SMM believed that the market entry of H100 high-spec clusters under a four-year fixed term reflected high-end training demand increasingly concentrating on long-duration price locking.

Since Jul 2026, Chinese computing providers are shifting from GPU leasing to token billing—some call leasing "a side business." Leasing margins shrink under public price anchors (4090 8-GPU: CNY 7,200/mo) and broker markups; DeepSeek's Aug 17 API hike (3-5x) widens reseller spreads; daily token calls hit 140T in Mar 2026, up 1,000x in 2 years. SMM: margin recovery, not substitution—high-end GPU leasing stays tight; converts: small IDCs, resellers, hyperscalers.

SMM learned that the server subleasing model continued to prevail in the market. Procurement prices for NVIDIA GPU servers fluctuated at highs and lacked a price anchor, with the procurement price of 5090 fan cards approaching 40,000 yuan per unit. Many new entrants to the computing power leasing sector shifted toward integrating existing resources rather than purchasing additional hardware. SMM believed that the lack of an anchor for procurement prices was forcing the supply side to activate existing capacity through subleasing, and that the industry had entered a phase of competition over existing resources.