A research report from GLMS SEC suggests that investment strategies should be structured around the theme of "rising Agent adoption - Token inflation - repricing of cloud and models." In the short term, cloud platforms are prioritized due to their clear business models, rapid implementation, and potential for quicker profitability improvements. For the medium term, model developers are a key focus, as they are positioned to handle higher-frequency calls at lower costs, thereby amplifying revenue and subscription growth. Long-term prospects are favorable for domestic computing chips, which are expected to demonstrate stronger earnings resilience amid expanding inference demand and the trend toward technological self-sufficiency. Overall, the report recommends focusing on core assets across three primary lines: one "chip," two "models," and three "clouds." The main viewpoints from GLMS SEC are outlined below.
The evolution from chatbots to proactive "Agent" intelligences. The breakthrough of OpenClaw signifies Agents' transition from merely answering questions to actively executing tasks. Their core value lies in enabling AI to operate 24/7, autonomously break down tasks, utilize tools, read and write files, perform cross-platform scheduling, and achieve closed-loop delivery. AI is thus evolving from a conversational assistant into a practical "digital employee." This shift indicates that the billing logic for internet and enterprise IT is moving away from competition based on traffic, duration, and access points, toward metrics such as Token consumption, task completion rates, security governance, audit trails, and resource provisioning.
The popularity of "Lobster" highlights the beneficiaries of rising Token demand. The viral success of "Lobster" suggests that the Agent era will bring nonlinear growth in Token demand, potentially leading to unexpectedly high AI computing power requirements and increased usage of large models. Meanwhile, cloud providers are no longer just selling computing power but are beginning to compete for Agent deployment entry points, model distribution rights, and subsequent expansion privileges. It is anticipated that the continuous operation of Agents will significantly increase demand for Token consumption, web searches, state storage, persistent connections, and multi-step inference, thereby driving a systemic reassessment across three layers: chips, models, and cloud platforms.
One "Chip": Domestic AI chip manufacturers such as Cambricon, Hygon, and CloudMinds. Computing needs in the Agent era extend beyond GPU inference to include tool invocation, environment setup, task scheduling, long-term memory, and high-concurrency CPU coordination, elevating the importance of the computing foundation. Domestic chip makers benefit both from the expansion in inference demand driven by Token inflation and from the trends toward domestic substitution and supply chain autonomy.
Two "Models": MiniMax and Zhipu. The Agent era is driving nonlinear growth in demand for large models' coding capabilities. At the same time, comprehensive requirements for low unit costs, long-text processing, multi-step reasoning, and programming proficiency give domestic models a clear advantage. MiniMax stands out for its cost-effectiveness, long-context capability, and programming/visual execution strengths, making it suitable for high-frequency workflow integration. Zhipu possesses globally leading coding technology, and its recent price adjustments reflect a shift in the model layer from price competition toward tiered subscriptions and enhanced pricing power.
Three "Clouds": Kingsoft Cloud, Wangsu Technology, and UCloud. The cloud sector represents a high-certainty "picks and shovels" opportunity in this cycle. Kingsoft Cloud and UCloud benefit from demand for one-click deployment, model hosting, knowledge bases, Agent platform subscriptions, and enterprise-grade governance. Wangsu Technology gains from the increased value of CDN, API gateways, edge security, and traffic management as machine traffic rises. Overall, the cloud layer offers high certainty, the model layer presents greater elasticity, and the chip layer serves as the long-term foundation.
Risks include slower-than-expected Agent adoption, lower-than-anticipated monetization rates for model developers, price wars among model providers, disruptions in computing supply, and miscalculations in CPU demand elasticity.