Institutions: Training reasoning resonance, rapid release of computing power demand

The General Office of the Ministry of Industry and Information Technology (MIIT) issued a notice on carrying out a special campaign to empower small and medium-sized enterprises (SMEs) with inclusive computing power. The notice states that it will innovate inclusive computing power service models. It will improve the SMEs section on the China Computing Power Platform, the computing power resource matching section on the China SME Service Network, and the national computing power internet service node section (collectively referred to as the SMEs platform sections), so as to promote precise matching between SME demand and the supply of computing power resources, and roll out flexible payment models such as charging by “card-time,” “nuclear-time,” and Token-based fees. It will explore innovative businesses such as “a computing power bank” and “a computing power marketplace,” and support SMEs in depositing idle computing power resources, enabling flexible deposit and withdrawal through cross-regional and cross-cycle scheduling.

Bohai Securities believes that, with AI agent applications represented by OpenClaw, the volume of Token consumption is growing rapidly, which directly reflects high business momentum in the inference stage and drives the release of demand for high-performance GPU cloud services, AI-dedicated computing power cards, as well as supporting storage and network resources. Against the backdrop of the rise of “tokenomics” in the AI era, global leading cloud providers have entered a pricing cycle; related companies are expected to seize opportunities for a reassessment of value. Computing power leasing and IDC providers will also fully benefit from the industrial dividends brought by the new build and expansion of intelligence-computing data centers. In terms of AI large models, as agent applications remain intensely hot, the number of model calls is set to surge significantly. Domestic large models such as MiniMaxM2.7 and KimiK2.5 are expected to become core model choices for high-Token-consumption AI agent applications such as OpenClaw, leveraging their significant Token cost advantages, and helping Tokens go global. On the AI application layer, the heat of overseas open-source agent projects like OpenClaw is rising rapidly, while domestic vendors are also launching related products in dense succession, driving a noticeably faster penetration pace for agent applications. Under the dual drivers of accelerated model iteration and large internet companies rolling out applications, the computing power industry chain has relatively clear investment opportunities.

Guojin Securities believes that, 1) domestic Token volume surpasses 140 trillion, domestic computing power providers are entering a performance inflection point, and another CPU price-hike wave is underway. 2) Training and inference resonate together, releasing computing power demand at a rapid pace. 2026 will be a key year for China’s computing power demand transformation from “cloud training” to a dual-wheel drive of “training + inference,” and the computing power gap will be released rapidly under the catalysis of more modalities and a broader range of scenarios. 3) On the supply side, there is an improvement at the external margin, along with accelerated domestic substitution. In 2026, the domestic computing power supply side will shift from a single state of tightness to structural balance; abundant computing power resources will effectively meet the surge in demand-side needs, laying a foundation for earnings realization across the computing power industry chain.