[SMM Computing Power News] Surge in Office AI Agents Boosts Inference Computing Power; Domestic Cards Fill the Gap

SMM previously learned that the rental price of Ascend 910B2 bare-metal servers exhibits significant regional differences. Market feedback shows that quotations in the Yangtze River Delta region range from about 12,000 (lowest) to 21,000 (highest) yuan per unit per month, while in the western region, the price drops to as low as 10,000 yuan per unit per month, resulting in a price spread of nearly double for the same model across regions. A channel vendor reported that they dare not relay their quotation to clients, while another channel is currently offering 910B2 at about 13,000 yuan per unit per month, which the client is still confirming. The rental pricing of domestic computing power cards is significantly affected by region, with price dispersion notably higher than that of the NVIDIA series.

SMM learned that a digital company under a certain operator issued a framework contract for intelligent computing services for a batch of H100 64-card clusters, with a tax-inclusive unit price of 74,000 yuan per unit per month, a service period of 36 months, 64 units totaling 4.736 million yuan per month, and payment terms of one-month deposit plus monthly prepayment. The direct contract price was 1,000-2,000 yuan per unit per month lower than the channel price. The contract stipulated that goods inspection would be conducted within seven working days after order signing, network setup would be completed within 30 days, the goods owner needed to conduct a background check on the lessee, and verbally required a deposit of one month plus three months’ prepayment and a one-year letter of guarantee. In large long-term agreements for H100 clusters, the price spread between the direct price and the channel quoted price serves as a key reference for market pricing.

SMM has learned that a cloud provider’s AI infrastructure spending has surged significantly, using prepayments combined with long-term agreements to lock in GPU resources and expand its own computing resource pool. SMM believes that cloud providers securing high-end computing power through prepayments and long-term agreements reflects that, amid tight supply, top-tier producers are using upfront capital to exchange for resource certainty.