[SMM Computing Power Express] Two buyers each seek 128 H100 units; suppliers question authenticity

SMM learned that a GPU cloud service provider disclosed on July 10 that the monthly rental quote for bare-metal H100 servers was within 80,000 yuan, with actual on-cloud transaction prices ranging from 75,000 to 83,000 yuan per month. Due to the premium for cloud services, these prices were higher than those of closed long-term agreements. Regarding the market, the provider noted that the market price for the 4090 was around 8,000 yuan per month, the 5090 was quoted at 12,000 yuan per month, and the highest quotes for the 5090 reached 13,000-14,000 yuan. Meanwhile, the source explicitly denied any off-season in the market, stating, "There has been consistent demand, and supply is relatively tight." On the high-end computing power side, the service provider stated they did not engage in this segment and described relevant market quotes as "essentially fake." Domestic Huawei Ascend computing power has not yet been launched but is planned for future introduction. The source also pointed out that domestically produced graphics cards have no hardware issues, but optimization still requires the work of engineers. SMM also learned that DeepSeek spent all of last year on adapting domestically produced graphics cards.

SMM learned that on July 8, Tencent WorkBuddy’s inference computing resource consumption peaked due to the launch of the Hy3 large model, with the afternoon queuing rate once exceeding 50%, and the team completed emergency capacity expansion that day. This computing resource crunch directly stemmed from the exponential growth in user scale — WorkBuddy’s monthly active users had reached 20 million and daily active users had exceeded 13 million. Unlike the “one-time heavy computation consumption” on the training side, inference computing consumption is linearly linked to the number of users: each interaction by every user generates continuous token consumption, and the daily office scenarios (document processing, code generation, agent tasks) of 13 million users create high-frequency, massive inference computing demands. Data from the OpenRouter platform shows that in May 2026, Chinese models accounted for six of the top ten spots in token calls, and domestic model inference traffic is rapidly catching up with mainstream North American models.

July 8: 32 units of spot H100 in Inner Mongolia quoted at 80,000 yuan per month, 64 units in Xinjiang to go online in August; two large orders of 128 H100 units surfaced, suppliers said "no one in the market can supply"; operators bid 65,000–68,000 yuan per month for H100, a price spread of over 10,000 yuan from the market; RTX 4090 24G traded at 7,000 yuan, near the low end