[SMM Computing Power News] 16 H20 units in Southwest China quoted at 53,000, including cabinet, electricity, and tax

SMM learned that a public AI cloud service provider issued an announcement stating that, due to business strategy adjustments, its public cloud platform will officially cease services at 23:59 on September 21, 2026, covering all public cloud businesses including GPU, CPU, shared storage, and model APIs. The platform requires users to complete data backup and migration before shutdown, after which resources will be uniformly released and data cleared, with no liability assumed for data loss, and unused vouchers will be voided simultaneously. SMM believes that the exit of small and medium-sized cloud platforms will bring about the reallocation of existing GPU resources, potentially creating regional resource release and leasing windows in the short term.

SMM learned that the L20 48G series has entered the consideration of investors, who intend to relocate the machines to a carrier's self-owned data center, primarily to lock in resources, with a five-year lease contract to be signed. Currently, spot quotes for a single card of this model in North China are over 32,000 yuan, while quotes for newly dismantled cards are over 31,000 yuan, a difference of about 1,000 yuan, indicating limited depreciation. Compared with quotes of over 42,000 yuan for a single 5090 32G card in the same period, the L20 offers higher memory specifications but a unit price about one-quarter lower. SMM believes that the shift by investors from leasing to owning and their proactive resource lock-in reflects rising market willingness to allocate long-term computing power assets.

The Index Consolidated Today. Supply Side: In South China, the number of cabinets was expected to increase from 8,000 to 12,000 within seven months, supporting 20 kW liquid-cooling retrofits; operator data centers in east China were tight, and 100,000-card buildouts were required to be planned in advance. Delivery Premium: In North China, H20 141G bare metal was priced at 54,500 with free bandwidth and IP included; the 5090 eight-GPU server carried a premium quote of 16,000 with a premium range of 15–18; the three-year rental-seeking budget for 910B2 was 20,000 and required relocation to east China. Demand for domestic relocations was robust; classified private clouds and data isolation pushed up the certainty premium for deployments, and delivery capability became a standalone pricing item.