[SMM Computing Power Flash] RTX 4090 24GB Deals Dip to 7,000 Yuan, Close to Low Market Quotes

SMM has learned that a cloud service provider holds 32 spot H100 8-GPU servers in Inner Mongolia, with a monthly rental quote of 80,000 yuan per unit. In addition, a cluster of 64 H100 units in Xinjiang is expected to be launched in early August, rented as a whole without being split. Market sources disclose that current spot quotes for H100 are mainly in the range of 75,000–80,000 yuan per unit per month, and supply remains tight. However, a regional operator's H100 rental bid range is only 65,000–68,000 yuan per unit per month, creating a significant price spread from the market. Suppliers are generally reluctant to take orders and usually do not trade with operators.

Spot A100-80G supply is tight, with a smart computing service provider reporting no current stock. Two quotes have emerged in the market: 32 units near-spot in east China (2-week delivery, InfiniBand networking, 3-year contract) quoted at 38,000 per month per unit; 32 units spot in Jiangsu (3-year closed-end contract, strong counterparties, one month deposit and one month rent) quoted at 31,500 per month per unit. The price spread between the two quotes is significant, and SMM analysis suggests they may be the same batch of goods circulated through multiple intermediaries. On the client side, an intended one-year closed-end bid is only 30,000 per month per unit, showing a clear gap with supply-side quotes. The service provider indicated that its focus has shifted toward tokens and models, and it currently has no demand for hardware leasing.

A data technology company recently negotiated a computing power cooperation. Under a proposed token-sharing model, the partner would provide computing power while the company would supply clients and model tuning, with upstream connection to a major AI firm. The starting volume requirement is 200 million TPM per minute, and the partnership mainly promotes the GLM 5.2 model. The partner stated that GLM's performance on domestic computing power is currently under testing. The data technology company said it prioritizes NVIDIA cards but can also use domestic ones. It is learned that NVIDIA's high-end resources are currently tight and mostly tied to full-lease guaranteed contracts.