[SMM Computing Power Flash] Spot high-end computing power complete systems at 13 million, intermediaries pushing to 14.5 million

SMM has learned that a cloud service provider stated it can handle short-term rental needs for H100 eight-card servers for three months, with priority given to finding long-term clients for renewal after the lease expires, and short-term rental prices will be raised accordingly. Previously, a university had proposed a short-term rental demand for 10 H100 units. As a currently sought-after card model in the market, suppliers mostly offer fixed-term contracts of over one year for H100s. The three-month short-term rental model provides a flexible option for temporary demand, but the premium for short-term rental and the uncertainty of renewal upon expiration also drive up leasing costs.

SMM learned that a certain intelligent computing cloud service provider confirmed on July 23 that it had transacted an A100 40G eight-card server this week, with a monthly rental of approximately 12,000 yuan, consistent with the current mainstream quotation range in the market. The service provider, which operates its own data centers, recently reflected that the overall market was soft, with limited trading volume. As the previous generation of mainstream inference cards, the A100 40G still has certain market demand, but the pace of transactions has slowed down. The transaction price of 12,000 yuan/month matches the 12,000 yuan quotation given by a Sichuan-based IDC in mid-July, and the price of this card model has remained stable recently.

SMM learned that a domestic GPU producer revealed during the WAIC conference that enterprises currently transitioning from NVIDIA to domestic GPUs need to comprehensively consider demand volume and migration time windows. The producer stated that due to significant differences in software adaptation workload across client scenarios, it was difficult to provide a unified commitment on adaptation completion time, and that "it depends on volume and time" was a key variable in determining the feasibility of the switch. This move reflects that domestic GPUs still face uncertainties in software ecosystem compatibility, and the large-scale switch from NVIDIA to domestic chips encounters an adaptation cycle bottleneck in the short term.