[SMM Computing Power Flash] Jinan H20-141G Monthly Rent 52,000 Yuan, 3-Year Lease Intent Under Discussion

SMM learned that a computing power equipment supplier released two batches of 8-card H100 complete machines in the past two weeks. On August 14, the supplier listed 9 Dell 8-card H100 complete machines at 3 million yuan per unit. The equipment was used stock removed from data centers in China, had been used for more than a year, and was not covered by warranty after testing. On August 29, the supplier transferred out dozens of brand-new Supermicro complete machines with similar specifications at 3.2 million yuan per unit. They are expected to arrive in Shenzhen within one to two weeks. They are also not covered by warranty after testing; buyers need to sign a contract first and pay a deposit of 50,000 yuan per unit before customs declaration. The two batches differed in new-versus-used condition, with quoted prices differing by 200,000 yuan per unit.

Data Centre Magazine reported on Aug 27 that Prysmian is rolling out a fibre-plus-copper composite cabling programme for AI data centres, running in parallel with its $1.25 billion three-state US expansion. The programme covers pre-terminated fibre, AOC active optical cable, direct-attach copper (DAC), and rack-level power distribution copper busbars, targeting hyperscale and colocation buyers. The Molex EUR 5.5 billion long-term agreement executes alongside; Prysmian expects North America data-centre fibre:copper cable mix to shift from 60:40 toward 45:55 over 2027-2030, with AI compute nodes driving incremental demand for high-speed differential copper and 800 V busbars.

SMM has learned that a certain MaaS platform is currently selling DeepSeek-V4-Flash model tokens at 78% of DeepSeek's official website price. The sales side also has a policy: "If usage exceeds 30,000 yuan, the official website price is further discounted to 95%." The platform's developers bluntly say that the 95% price may be completely unprofitable. The platform uses a hybrid model of partly self-built deployment and partly access to large vendors' APIs, with large vendors as the upstream. Capacity is expanded whenever usage exceeds the limit, and the TPM ceiling has been raised several times to 30 million. According to its feedback, the discount spread between upstream and downstream in token resale is extremely small, margins are negligible, and profitability is far lower than that of the computing power leasing business.