Industrials: Earnings vacuum period; continue to focus on domestic computing power, physical AI, and AI4S

Industrials released a research report stating: continue to accumulate leading domestic computing power names on dips. Affected by increased volatility in overseas computing power leaders, the domestic computing power sector has been fluctuating at the bottom. Recently, the sector has corrected fairly fully, large supernode orders keep landing, and the country's first Ascend 950dt all-liquid-cooled supernode cluster has been put into operation in Horinger, Inner Mongolia; meanwhile, localization of core chips is accelerating and being delivered. We recommend trading time for space and continuing to accumulate core leaders in the sector on dips. Domestic open-source flagship models are being released intensively. This week, Kimi open-sourced the weights of K3.1 Pro, with total parameters of 2.8 trillion and a context length of 1 million; Tencent released the official version of Hy4, introducing visual understanding and computer operation capabilities for the first time, and plans to open-source the weights, with WorkAgent capabilities expected to continue improving. According to OpenRouter, global large-model calls totaled 194 trillion tokens from 1026-1101, up 3.19% week-on-week; among leading models, domestic models accounted for 99.55 trillion tokens in calls, with DS V4.1 Flash retaining the top spot. Continue to focus on the physical AI and AI4S industries. We are currently in an earnings vacuum period, and model iteration and new application scenarios are expected to see rotational market activity. On the one hand, the generalization capability of physical AI models has recently been improving at an accelerating pace, with incremental demand across full-stack GPUs, physical AI models, embodied data, and on-device AI. On the other hand, pay attention to application sub-sectors at the bottom: AI4S, FDE, AI marketing, etc.