[SMM Computing Power Midday Review] H200 Cluster Demand Emerges, Whole-machine Warranty Premium

[Market Overview] The SMM computing power rental price index consolidated overall today, with quotes for all server models and regions remaining flat compared with the previous trading day. Two signals on the price collection side came from the demand side and the sell side respectively: a digital technology enterprise issued a procurement demand for a 128-card H200 training cluster, and a branded H20 141G server in South China was offered for sale at 3.25 million yuan per unit, with brand and warranty premiums beginning to become explicit.

[H200: 128-Card Cluster Procurement Demand Materializes, Pointing to Normalization of Large-Parameter Training] A digital technology enterprise released GPU rental and procurement demand, planning to purchase 16 H200 AI training servers, each equipped with 8 H200 141G SXM5 cards and supporting full NVLink interconnection, to form a 128-card distributed training cluster for pre-training and routine fine-tuning of a 300B-parameter large model. On the hardware side, minimum requirements were specified for CPU core count, 2,048GB DDR5 ECC memory, 15.36TB-class NVMe high-speed drives, 8 400G NDR compute NICs, and 2,800W (3+3) redundant power supplies, along with a private one-stop training platform and 7x24 operation and maintenance services. Notably, this demand centered on specifications and stability clauses without directly disclosing a price basis, indicating that the decision-making weight in large-parameter training procurement is shifting from unit price to cluster availability and long-term stable operation.

[Whole-Server Sales: Brand and Warranty Premiums Becoming Explicit] A branded H20 141G eight-card whole server in South China is pending sale at 3.25 million yuan per unit, with a three-year warranty and optional extended warranty, and only one unit is available. Compared with the 2.4 million yuan quote for a new H20 96G whole server in the same region, this unit has a higher memory specification and a price 850,000 yuan higher. Branded servers and warranty services are beginning to show explicit premiums on the whole-server sell side, consistent with earlier observations of A100 whole servers shifting from rental to sale and the parallel price spread between new and used H100 whole servers. A single bulk sample is not enough to represent the overall market level, and the whole-server sale price series still needs to be accumulated continuously.

[SMM View] Today's two signals point to a further upward shift in demand structure. First, pre-training and routine fine-tuning demand for 300B-class large models is pushing the procurement threshold toward 128-card clusters. As the most-traded model for training at this scale, H200 spot cargo and whole-server supply will remain under pressure, sharing the same root cause as the recent rise in H200 monthly rental and three consecutive monthly increases in whole-server buyout prices. Second, buyers are prioritizing specifications, platforms, and operation and maintenance terms, with price giving way to delivery certainty, indicating that the competitive dimension of high-end computing power has shifted from unit price to full-lifecycle service capability. Third, brand and warranty premiums are becoming explicit on the sell side, echoing the delivery flexibility premium on the rental side, and the pricing of computing power assets is extending from hardware performance to services and guarantees.