[SMM Computing Power Flash] Short-Term Rental Demand Faces Scarce Resources, Consider Shifting to Private Clouds with Classified Information Processing Qualifications

Computing power rental prices moved sideways today, with monthly rental and per-card-hour quotes for all models unchanged, mainly because no new resources were released into the market. Meanwhile, DeepSeek announced that it raised V4 series API pricing and introduced differentiated peak/off-peak pricing. Cached input pricing saw significant increases, well beyond the range of cost increases, signaling a pattern of testing the market with higher prices while leaving room for future anchoring. Frontline token businesses face a single pricing source and limited bargaining power, and need to closely monitor token price trends.

SMM has learned that starting at 00:00 on August 17, 2026, DeepSeek will implement new DeepSeek-V4 API pricing. V4-Flash and V4-Pro will be charged at different rates during peak and off-peak periods, with peak periods from 9:00 to 12:00 and from 14:00 to 18:00 Beijing time. Compared with the previous flat pricing, V4-Flash cached input will rise from 0.02 yuan to 0.05 yuan per million tokens (off-peak) or 0.1 yuan (peak); uncached input will rise from 1 yuan to 1.5 yuan or 3 yuan; and output will rise from 2 yuan to 4.5 yuan or 9 yuan. For V4-Pro, cached input will rise from 0.025 yuan to 0.15 yuan or 0.3 yuan; uncached input will rise from 3 yuan to 4.5 yuan or 9 yuan; and output will rise from 6 yuan to 13.5 yuan or 27 yuan. SMM believes that Pro cached input, which will see the largest increase in this round and whose marginal cost is almost zero, will reach 500% during off-peak hours alone, clearly exceeding the range of cost-driven increases. The simultaneous introduction of peak-valley differential pricing is characterized by testing the market with high prices and leaving room for maneuver in anchoring future prices.

SMM has learned that short-term rental demand has emerged in the market for Huawei Ascend 910B series devices, with at least 64 GB of memory, a single unit, a two-month lease term, a budget based on market conditions, and a requirement that the equipment be relocated to a data center in North China for offline deployment. This demand is small in scale and short in duration, highlighting offline deployment and the compliance attributes of domestic computing power. SMM believes that short-term rental demand for Ascend models mostly involves project-based delivery, and the offline deployment requirement reflects the lessee’s preference for physical control over data and equipment.