The research report from CITIC Construction Investment Securities sheds light on two primary investment logics that are shaping the global AI industry. From the perspective of hardware and infrastructure, the 'two-phase framework' encompassing large-model inference and Agent orchestration workflows has led to a decline in the marginal returns of general-purpose GPUs. The competition for computing resources has evolved from a mere technological contest to one that encompasses energy and power grid considerations. This transformation is driving structural changes in server architectures, with a preference emerging for ASICs boasting high data reuse rates (such as Google's TPU) and high-performance command-layer CPUs (where the CPU-to-GPU ratio is moving towards 1:1). Concurrently, driven by expansion needs, prices for DDR5 DRAM have experienced a significant uptick.

Turning to models and cloud ecosystems, the release of OpenAI's GPT-5.5 has seen the gap in download and installation volumes between its Codex ecosystem and Anthropic Claude Code gradually close. Given the typical 3- to 6-month lag between a model's technological leadership and corresponding shifts in market share, computing power and cloud partners aligned with OpenAI, such as Microsoft, Oracle, and CoreWeave, are well-positioned to capitalize on investment opportunities. These opportunities arise from the need for narrative correction and strategic market share expansion, as they navigate the dynamic and evolving AI landscape.