[SMM Computing Power Midday Review] Index consolidates, with delivery premiums for liquid-cooled data centers and servers standing out

[Market Overview]

Today, the overall SMM Computing Power Leasing Price Index consolidated, with all model-by-region quotes unchanged from the previous day. Signals from price collection were concentrated on the supply side and delivery terms: high-power cabinets were deployed in a concentrated manner in South China, operator data center resources in east China tightened, and H20 and 5090 showed a clear delivery-capability premium.

[Supply Side: Cabinets Evolve Toward Higher Power, 100,000-Card Builds Force Earlier Planning]

A data center in South China offered cabinets for lease externally. Two buildings with a total of 8,000 cabinets had been delivered, and the total was expected to increase to 12,000 within seven months, supporting liquid-cooling retrofits, with single-cabinet power upgradable to 20 kW. Compared with the previous quotes for high-power data centers in North China—7 kW at 6,300 yuan/cabinet/month and 10 kW at 9,000 yuan/cabinet/month—the power cap for this batch was clearly higher, and the supply structure of data centers was shifting from general colocation to AI-computing adaptation. Over the same period, an operator’s data center resources in east China became increasingly tight. Frontline feedback indicated that 100,000-card-scale concentrated builds should be planned in advance and started concurrently, mutually corroborating earlier signals that cluster delivery in the region had been delayed to Q4; the lead times for cabinets and power supporting facilities were difficult to compress in sync.

[Delivery Premium: Reflected Separately in H20, 5090, and 910B2]

In North China, H20 141G bare metal was quoted at 54,500 yuan/unit/month, to be released after the existing lease expired on September 30. Bandwidth was provided by the base carrier, IP was not charged separately, and 100M shared bandwidth was included. Compared with east China, where H20 96G cloud instances were presold at 46,000 yuan and existing annual contracts were at 39,000 yuan, this was more than 18% higher (different specifications, recorded on a different basis). The 5090 eight-GPU complete unit was quoted at 16,000 yuan/unit/month; the lessor acknowledged it was 2,000–2,500 yuan above the market, about 15%–18% higher than the mainstream 13,500–15,000 yuan, with multi-site data center footprint and test response speed priced separately. Demand side, the budget for leasing one 910B2 on a three-year term reached 20,000 yuan/unit/month and required migration to a data center in east China; spot orders with extended contracts still offered a premium.

[SMM View]

Today’s signals pointed to the same conclusion: beyond price movements, the source of premiums was shifting from the model itself to delivery format and supporting capabilities. First, the concentrated release of high-power liquid-cooling cabinets corresponded to upgrades in AI-computing power density, but the mismatch between 100,000-card-scale demand and construction cycles meant the supply-demand gap will become the main support for leasing prices in the next stage. Second, as bandwidth and IP shifted from chargeable items to basic services, and multi-site data centers and test response were priced separately, delivery capability has become an independent premium item. Third, demand for domestic computing power migration was robust; limited supply of classified private-cloud deployments, together with rising data-isolation needs, further strengthened the premium for deployment certainty. SMM believed that subsequent price collection should record delivery terms and supporting-caliber definitions in parallel in order to restore the true transaction price.