[SMM Computing Power Daily] H200 quoted at 140,000, with a same-day transaction of 128,000

SMM, September 28:

[H200: Spot listing price at 140,000 yuan with claims of rising costs; same day, 30 units of futures cleared at 128,000 yuan]

SMM has learned that a Hubei computing power service provider quoted a monthly rental of 140,000 yuan for an H200-141G SXM eight-GPU full server, equivalent to approximately 24.3 yuan per GPU per hour, with a one-year contract, one month's deposit plus one month's rent, including 100M dedicated bandwidth and one IPv4 address. One unit is available in spot, and another unit is planned to arrive in October, configured with two 8558 processors, 2T memory, 1T×2 system drives, and 3.84T×2 data drives. The provider stated that the procurement cost of H200 full servers has risen by approximately 1 million yuan over the past month, driving up rental prices for new machines. On the same day, another signal pointed in the opposite direction: a computing power service provider listed 30 H200 full servers as futures in a Jiangsu data center in early September, with an asking price of 130,000-140,000 yuan per unit per month and a plan to bring them online by month-end. All units have now been rented out at a transaction price of 128,000 yuan per unit per month, 2,000 yuan below the lower end of the quotation range, with no transactions verified at or above 130,000 yuan per unit per month. During the same period, bulk long-term contract quotes for H200 from other channels ranged from 98,000 to 128,000 yuan per unit per month, corresponding to delivery conditions of 128 units or more, five-year fixed terms, and 45 days to two months, which differ in unit count and lease term from this batch of resources. SMM believes that the two clues on that day jointly point to the fact that "listing prices and transaction prices are not the same number": resources in small batches with imminent delivery do have room for premiums, but the extent of the rise is constrained by the upper end of concurrent long-term contract quotes. It should be noted that the bandwidth terms of the two quotes are not aligned—the 140,000 yuan quote includes 100M dedicated bandwidth, while the 128,000 yuan transaction does not specify whether bandwidth is included, making the two not strictly comparable.

[Other market developments · Memory: 64G 5600 tax-inclusive median at 17,400 yuan; non-Samsung brands about 1,000 yuan lower]

SMM's tracking of DDR5 server memory channel quotes shows that the daily median for 64G 5600 on a tax-inclusive basis has dropped from 18,200 yuan on September 19 to 17,400 yuan on September 28, a decline of 4.4%. On that day, tax-inclusive quotes for this specification were 18,250 yuan for Samsung, 17,400 yuan for CXMT, and 17,100 yuan for Kingston, with the latter two approximately 1,000 yuan lower than Samsung, indicating that brand stratification has emerged in the spot market. On the same day, 64G 4800 modules from Samsung and SK Hynix were both quoted at 16,900 yuan. At the full-server level, channel listings on that day included Chilwee H200 spot units with 4T configurations, with unit counts spanning two batches and no prices provided. Pre-holiday buyback activity was notably active, with dense purchase calls appearing on the same day for 5090 fan cards, Pro6000 server editions, and server memory modules. SMM believes that buyback-side calls do not represent end-use computing power demand, and any assessment of market heat must first strip out the buyback channel.

[Policy developments: Shantou sets quantitative ceiling for token pricing; Ascend 950 supernode announced for commercial use]

SMM has learned that the Shantou Overseas Chinese Pilot Zone released the "Ten Measures for Tokens" on September 24 and launched the International Zero-Carbon Computing Power Port: computing power vouchers will be issued at 13 yuan per day per standard rack for cross-border intelligent computing racks or green intelligent computing racks with a green electricity ratio of no less than 80% and a comprehensive PUE of no higher than 1.18. Monthly rent for dedicated racks in designated data centers can be as low as 520 yuan per kW per rack (including electricity). Token vouchers are targeted at the R&D, iteration, and production of AI toys, AI micro-dramas, and digital employee agents, issued at no more than 40% of actual computing power costs, with a maximum of 1 million yuan per enterprise per year. The first phase of the International Zero-Carbon Computing Power Port launched on the same day covers approximately 413 mu, with a total IT power of 400MW and a total investment of 120 billion yuan. The area's daily average token call volume has exceeded 100 billion, and the number of enterprises approved for "processing data upon arrival" has increased to 132. On the supply side, Huawei announced at its Full Connectivity Conference on September 28 that the Ascend 950 supernode has entered large-scale commercial deployment, with services available to clients in China starting September 30 and to clients outside China starting November 30. The flagship Atlas 950 SuperPoD solution delivers 1 EFLOPS (FP8) of computing power with 1,024 GPUs per cluster, and more than 1,000 supernodes have already been deployed. The Ascend 960 roadmap has been moved up, with releases beginning in Q1 2027. Two additional signals: an article in Qiushi proposed "letting computing power follow electricity prices—compute more when wind and solar power are abundant, compute less when power is tight." Official interpretation disclosed that by the end of 2025, China had over 13.73 million standard racks of computing power facilities in use, 42 10,000-GPU-class intelligent computing clusters built, and intelligent computing scale of 1.59 million PFLOPS (FP16). SMM believes that green electricity ratio and PUE are shifting from evaluation metrics to preconditions for subsidies, and token pricing on the "eastern data" side and electricity price linkage on the "western computing" side are being institutionalized simultaneously during the same period.

SMM analysis: Four main threads today. First, H200 showed a two-way divergence of "rising quotes, falling transactions" on that day—on one hand, a listing at 140,000 yuan per month citing a procurement cost increase of approximately 1 million yuan for full servers; on the other hand, 30 futures units cleared at 128,000 yuan, below the lower end of the quote range, with no transaction support at or above 130,000 yuan. The two clues come from different channels and should not be extrapolated to an overall H200 price level judgment, but they are sufficient to show that listing prices should not be directly used as transaction prices. Second, the issue of quote comparability has resurfaced with H200: the 140,000 yuan quote explicitly includes 100M dedicated bandwidth and one IPv4 address, while the 128,000 yuan transaction does not specify bandwidth terms. Before comparing prices, four items must be aligned: whether bandwidth is included, full server or single GPU, lease term and payment terms, and attribution of IP and electricity/maintenance costs. Third, the memory channel remains weak—the tax-inclusive median for DDR5 64G 5600 fell 4.4% over nine days, non-Samsung brands are approximately 1,000 yuan lower than Samsung, and brand stratification in the spot market is taking shape. At the same time, pre-holiday buyback calls are dense, representing the buyback channel rather than end-use demand. Fourth, the policy side directly ties green metrics (green electricity ≥80%, PUE ≤1.18) to subsidy disbursement, domestic supernodes have entered large-scale commercial deployment, and supply release and demand-side pricing tools have landed simultaneously at the end of Q3.