Institutions: Domestic computing power full industry chain sentiment accelerating

On March 14, the Securities Times reporter learned from the 2026 Shanghai Global Investment Promotion Conference that, to address the current hottest artificial intelligence sector and solve the problems of large enterprises struggling to find computing power and small and medium-sized enterprises facing high costs, Shanghai has established the largest domestic computing power scheduling platform. It allocates 1 billion yuan in computing power vouchers annually, offering a pay-as-you-go and no-application-needed model to help enterprises quickly and affordably access the city’s 140,000 PetaFLOPS of heterogeneous computing resources.

Guojin Securities believes that: 1) Oracle intelligent computing ROI is being realized, and Tencent Cloud’s price increase confirms demand. 2) Training and inference resonance is rapidly releasing computing power demand. 2026 will be a critical year for China’s shift from “cloud training” to a dual-driven model of “training + inference,” with the computing power gap accelerating under the catalysis of more modalities and broader scenarios. 3) External marginal improvements on the supply side and increased domestic production will shift domestic computing power supply from scarcity to structural balance. Abundant resources will effectively meet the explosive demand, laying a foundation for performance realization across the computing industry chain. 4) The entire domestic computing power industry chain is accelerating in prosperity, with expectations of both volume and price increases. Under strong supply and demand logic, it is predicted that 2026 will see the industry enter a “full-chain inflation” cycle, with prosperity spreading from core chips to AIDC, cloud and computing services, supporting power equipment, servers, and other segments.

Orient Securities believes that the engine driving computing power demand is shifting from intermittent “dialogue” needs to sustained “execution” needs, leading to an accelerated increase in computing consumption. Unlike traditional conversational AI, OpenClaw is an agent framework with high permissions and autonomous scheduling capabilities. When handling complex tasks, it frequently uses underlying models for thinking, tool invocation, and environment interaction. This operational mode significantly increases token consumption compared to traditional single-task models. For infrastructure, this not only means explosive growth in inference traffic but also higher requirements for real-time processing and stability. This structural change in demand is expected to further boost overall computing power needs, benefiting cloud providers and domestic chip manufacturers.