[SMM Computing Power Flash] An intelligent computing cloud service provider completed a transaction for an A100 40G eight-card server, with a monthly rent of about 12,000 yuan.

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 at the WAIC conference that the spot high-end computing power complete machine is currently quoted at about 13 million yuan on the market, while intermediary parties are quoting between 13.5 million and 14.5 million yuan. First-hand supply is difficult to access directly, and transactions must be concluded through tail cargo owners, with intermediary levels commonly existing in the transaction chain. It is also learned that another high-end computing power GPU complete machine was quoted at about 3.9 million yuan two months ago, and its spot price has now risen to between 4.9 million and 5 million yuan, a significant increase.

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.