[SMM Computing Power Midday Review] H100 in the Yangtze River Delta breaks 87,000, short-term rental budget falls through

[Market Overview] The SMM computing power rental price index rose locally today, with only one of the 19 indices—the H100 Yangtze River Delta—adjusting, while all others remained flat. The index's highest monthly rent exceeded 87,000 yuan for the first time, forming a relay upward move with long-term contract quotes. On the demand side, the first sample of a high-priced budget falling through emerged this round, and the mismatch between price and resource availability is becoming explicit.

[H100 Yangtze River Delta: Highest Price Rises Another 2.35%, Relaying Long-Term Contract Quotes] The SMM-H100-80G-full server-Yangtze River Delta-monthly highest monthly rent rose from 85,000 yuan to 87,000 yuan, up 2,000 yuan, an increase of 2.35%; the average monthly rent rose in tandem to 82,500 yuan, up 1.23%; the highest per-card-hour rate rose from 14.76 yuan to 15.1 yuan, up 2.30%. In the same region, a 4-unit H100 full-server annual long-term rental was quoted at 87,000 yuan/unit·month, with resources to be released on the 27th of this month and another 9 units of supply pending confirmation. This price is 2,000 yuan above the upper end of the mainstream range and also higher than the final batch transaction price of 85,000 yuan cleared in batches in early September. Long-term contracts and spot clearance are forming a relay upward move, with the pace of supply release remaining the main driver.

[Demand Side: Budget Upgrades and Misses Coexist, Constrained by Availability] The market saw demand for 16 H100 full servers on a 3-month short-term rental, with a budget of 90,000 yuan/unit·month. There were no requirements for data center location and no need for relocation. The lessee also accepted substitution with 10 H200 units, with the budget revised up by 19,000 yuan from similar demand on July 20, an increase of 26.8%. For the first time, the demand side adopted a substitution plan based on computing power equivalence, placing more emphasis on delivery timeliness than on the model itself. On the other hand, a 3-month H100 short-term rental in early September had a budget as high as 100,000 yuan/unit·month but failed to match resources, and the deal did not go through—the first sample of a high-priced budget falling through this round. Taken together, the key constraint on transactions has shifted from price to resource availability.

[Domestic Computing Power: Delivery Premium Persists, High-End Price Band Overlaps with Imports] In east China, 5 relocatable 910B units on a 4-year lease were quoted at 20,000 yuan for two consecutive batches, showing clear price rigidity—a premium of about 33% over the online price of 15,000 yuan, with no discount obtained even for a one-year lock-in. Two relocatable 910B3 units also rose to 20,000 yuan, catching up with the 910B4, and the model price spread disappeared again, with specification-based pricing still not firmly established. On the online channel side, the first purchase of 910B1 was at 16,000 yuan, exactly the same price as 910B2 and 910B3, with zero price difference. A single 910B3 unit online was quoted at 15,000 yuan, actually a 6.3% discount compared with the batch quote. In addition, a 910C 752T cluster of 158 units was offered for rent at 66,000 yuan/unit·month, the largest domestic cluster-level sample, with the quote already close to the H800 annual rental of 69,000 yuan. The high-end domestic and mid-end imported price bands are beginning to overlap. The H800 annual rental of 69,000 yuan is about 9% lower than the central China long-term contract in August, remaining the only H-series model with a sustained downward trend.

[SMM View] The demand characteristics of inference computing power and training computing power are diverging: inference requires low latency and rapid response, is sensitive to single-point delivery timeliness, and is therefore willing to pay a premium for immediately available resources—even high prices may fall through due to lack of inventory. Training requires large clusters and long-term stability, locking in scale and duration, but currently neither scale nor duration has entered into pricing, and both batch orders and long-term contracts struggle to obtain discounts. These two types of demand are competing for limited spot cargo in the same market, which is precisely the common source of the relay upward move in H100 prices and the rigidity of the 910B4 delivery premium. SMM believes that price signals are no longer enough to explain whether deals are concluded; resource availability and delivery form are the watershed for the next stage.