[SMM Computing Power Brief] East China’s 128-unit H200 cluster to go online, monthly rent 120,000, three-year long-term agreement

SMM has learned that a certain channel has released 6 B200 eight-card servers with leases for sale. Each unit is configured with 8 B200 NVLink GPUs, equipped with dual Intel Xeon 6960P processors, 3TB of DDR5 memory, and 8 ConnectX-7 single-port 400G IB cards, belonging to the high-bandwidth training model. The lease term for this batch runs until February 2027, with the quoted monthly rent per unit at 100,000 yuan. The sale-with-lease model is adopted, where after the buyer takes over, the rent, after deducting cabinet and electricity costs, belongs to the buyer. Such second-hand circulating resources with confirmed leases provide buyers with computing assets that can immediately generate rental returns. The sale-with-lease model transfers stable rental cash flow to the buyer upfront, reflecting the increased activity in the B200 second-hand circulation market.

Benefiting from the explosion of demand from AI intelligent computing centres, the "East Data, West Computing" initiative, and transoceanic submarine cables, the upstream optical fibre preform expansion cycle takes 18-24 months, leaving industry capacity supply tight, and fibre optic prices have been continuously raised significantly in 2026. Hollow-core fibre and ultra-low-loss G.654E high-end fibre are rigid demand for computing power backbones, and top-tier players have locked in orders through long-term agreements until 2028.

The NDRC held a press conference for July. A spokesperson of the NDRC stated that in H1, China’s pace of independent AI innovation accelerated, achieving rapid breakthroughs across multiple links of the entire industry chain. First, the computing power foundation was consolidated; the first fully homegrown 100,000-card AI supercluster was officially put into operation, and as of month-end June, the national intelligent computing power scale reached 2.8 times that of the same period last year. Second, model capabilities leaped forward; local companies such as DeepSeek and Moonshot AI successively released multiple open-source large models with trillion-level parameters, total global downloads of domestic large models surpassed 10 billion times, and domestic models were rapidly adapted to domestic computing power chips. Third, factor mobility was vibrant; over 120,000 high-quality datasets were built, and a virtuous cycle of mutual reinforcement among computing power, models, and data took shape at a faster pace. Driven by coordinated efforts across all areas, AI became a surging source of economic growth momentum, and related sectors maintained high growth of over 30%.