[SMM Computing Power Daily] High-End Computing Power Server Procurement Peaks at 14 Million, Payback Period Extends to Over Five Years

SMM, July 15:

High-End Computing Power Procurement Hits 14 Million Yuan Peak, Payback Window Significantly Lengthened

SMM learned that the procurement price of a complete system for high-end computing power equipment has touched a peak of approximately 14 million yuan, nearly doubling from around 4 million yuan at the end of 2025. According to an earlier survey, on July 13, a second-hand GPU dealer reported that complete high-end computing power systems rose from about 9 million yuan to 13 million yuan, "just like trading stocks," with incoming goods absorbed within a week of arrival. At the current high of 14 million yuan, based on a monthly rent of 75,000-80,000 yuan for an 8-GPU H100 server and annualized revenue of approximately 900,000-960,000 yuan, excluding cabinet power, operations and maintenance, and depreciation, it would take at least 14-16 years to recoup the purchase cost. Even when assessed using quotes of around 190,000 yuan per month for similar equipment from some platforms, the payback period still exceeds six years. Coupled with rising expectations for next-generation computing power architecture iterations, existing high-priced assets face accelerated depreciation risks—GPU server annual depreciation rates can reach 20%-35%, meaning the book value shrinks by about 30% after just one year.

A100-40G Monthly Rent Rises to 12,000 Yuan, Significant Price Divergence Across Card Types

On July 15, the latest quotes from an IDC service provider showed: the 4090 8-GPU monthly rent was 8,000 yuan/month, positioning it in the upper-middle range of the 6,800-8,800 yuan market price range, as understood by SMM; the 910B2 price remained unchanged from the previous 20,000 yuan/month; the A100-80G was quoted at 36,500 yuan/month, sustaining the upper end of the early July price range; the A100-40G was quoted at 12,000 yuan/month, a price difference of about 36% compared to early July cloud machine quotes. Market participants noted that the A100-40G, favored for its cost-effectiveness in inference scenarios, has seen stable demand, with recent supply reductions driving its price rise, though downstream acceptance post-increase still needs monitoring. According to SMM, in the previous month’s quotes, the H20-141GB 8-GPU server commanded annual contract monthly rents of 42,000 yuan and monthly contract rents of 46,000 yuan, earning recognition from an IDC service provider, who also expressed clear demand interest for H20 resources. As NVIDIA’s main compliant supply model for China, the 141GB large-memory H20 version continues to be favored in inference and fine-tuning scenarios.

High-End Computing Power Supply-Demand Gap Persists, Entry Barriers for Small and Medium Buyers Sharply Soar

SMM learned that on July 14, a contact from a major internet company referred two client demands: one for the buyout of 1 unit of H20 plus 2 units of L20, and the other for the rental of 12 complete high-end computing power systems. The latter was considerably more difficult to match. Additionally, an IDC service provider reported on July 10 that high-end computing power like H100 "gets snapped up instantly." Delivery cycles for high-end computing power have been extended to Q1-Q2 2027. Meanwhile, online data indicated that the world’s four major tech companies’ combined AI capital expenditure in 2026 totaled approximately $725 billion, putting small and medium computing power participants at a natural disadvantage in supply chain negotiations. A dealer further noted that in Shenzhen’s spot market, "you need a hundred million yuan in your account just to buy high-end computing power—they ask if you agree on the price, and only if you do will they let you see it; if not, you won’t even get through the door"—the real threshold for high-end computing power is no longer merely price, but channel trust relationships.

SMM Analysis: The current market has shifted from the earlier "supply-demand tight balance" to deeper structural issues—the mismatch between hardware investment and leasing returns is intensifying. Complete high-end computing power systems surged from about 4 million yuan at the end of last year to the current peak of 14 million yuan, an increase exceeding 250%, while the leasing side saw increases of only 15%-30% over the same period. The widening price scissors directly extend the payback period to over five years. Against the backdrop of accelerating technological iteration (NVIDIA’s Vera Rubin platform is expected to ship in H2), the depreciation risk for high-priced inventory assets cannot be overlooked. Some participants have adopted a more cautious pace of market entry. The pricing order in the computing power leasing market still requires further optimization, necessitating a transition from a simplistic "highest bidder wins" to a more sustainable pricing mechanism.