JPM released a report stating that three weeks after the launch of DeepSeek V4, incremental data has led investors in Chinas large language model sector to three conclusions: (a) Only DeepSeeks own infrastructure can operate V4 with cutting-edge economic efficiency. In cache-hit input performance, there is around a 40x gap between DeepSeeks official API and third-party cloud channels, indicating a structural first-party advantage in prefix cache reuse, traffic density, routing and compute allocation. Model weights can be freely distributed, but the cost curve cannot. (b) Token data reflects supply-constrained growth rather than zero-sum substitution. According to OpenRouter data, usage of V4 is rising, while GLM, MiniMax and older versions of DeepSeek have not seen broad-based declines in usage. (c) The promotional pricing of DeepSeek V4-Pro now defines the low-cost frontier, while GLM-5.1 anchors the preference end. MiniMax M2.7 sits within that frontier and may face positioning pressure if the current structure persists.JPM noted that for KNOWLEDGE ATLAS (02513.HK) -48.000 (-4.537%) Short selling $74.17M; Ratio 9.772% , the monetization proposition hinges on extending its model leadership rather than defending its position. GLM-5.1 ranks ahead of V4 in evaluations, anchoring KNOWLEDGE ATLAS within the domestic frontier cluster. Achieving rating leadership at a significant price multiple is now marketable. To sustain pricing premiums, the next GLM cycle needs to further widen its lead over the DS V4 model and broaden its preference advantage in workflow-related tasks, including agent-based coding, long-context reasoning, and enterprise workflows where retry costs and quality dominate token costs. If successful, KNOWLEDGE ATLAS may retain premium pricing power. If not, it may cede the price-sensitive tier to DeepSeek while retaining a smaller but higher-margin business.Related News CMSI: Alibaba Cloud In-house Chip Progress Beats Expectations; Certainty of High Growth in AI Cloud Revenue StrengthensJPM added that for MINIMAX-W (00100.HK) -24.000 (-3.493%) Short selling $50.99M; Ratio 5.871% , its infrastructure-led value proposition is facing incremental pressure from DeepSeek. In the next cycle, it may need to demonstrate measurable leadership in task-completion economics to maintain differentiation. Historically, MiniMax competed on throughput, latency and concurrency capabilities. DeepSeek now offers a low-cost, one-million-context API, and its first-party service stack appears one tier more efficient in cache-intensive workloads. Therefore, for MiniMax, the successor to M2.7 will need to show fewer cycles, fewer retries and lower overall costs than the successor to V4.JPM assigned an Overweight rating to KNOWLEDGE ATLAS (02513.HK) -48.000 (-4.537%) Short selling $74.17M; Ratio 9.772% with a TP of HKD950, and an Overweight rating to MINIMAX-W (00100.HK) -24.000 (-3.493%) Short selling $50.99M; Ratio 5.871% with a TP of HKD1,100. (da/j)(HK stocks quote is delayed for at least 15 mins.Short Selling Data as at 2026-05-21 16:25.)
AASTOCKS Financial News