After being pressured by the "AI consumes SaaS" narrative, the software sector is undergoing a revaluation. The explosion in Token usage driven by AI Agents is shifting the core value of software from "functionality presentation" to "process orchestration, data, and permission governance." Infrastructure and platform software are likely to be the first to capture this incremental value. A recent catalyst from the industry is the breakout popularity of OpenClaw. According to HSBC Qianhai, AI Agent Token calls can reach up to 15 times that of chat interactions. Weekly Token calls on the OpenRouter platform rose to 14.8 trillion in the week of March 2, representing growth of approximately 160% within two months. This has simultaneously elevated the monetization curve for AI applications and the consumption curve for computing power. For the A-share market, HSBC Qianhai Securities believes that strong growth in Token call volume will boost demand for AI software infrastructure and IDC services. Chinese models, leveraging their higher cost-effectiveness, are continuously increasing their global share of Token usage.
AI Agents Evolve from "Chatting" to "Executing": Amplifying Token Consumption
HSBC Qianhai Securities suggests that as foundational models improve their capabilities for productivity tasks, AI Agents are becoming the "first breakout AI application." OpenClaw is a representative case, capable of automatically performing tasks such as software installation, script execution, file organization, and web monitoring. Popularity and usage are rising simultaneously. HSBC Qianhai Securities disclosed that since its release in November 2025, OpenClaw's GitHub stars have surpassed 299,000, exceeding those of well-known open-source projects like Linux and React. More critically, the consumption structure is changing: According to Anthropic data, AI Agent Token consumption can be up to 15 times higher than that of chat interactions, making Token usage a "hard metric" for measuring AI application penetration. Data from OpenRouter reinforces this trend. HSBC Qianhai Securities pointed out that driven by active deployment of multi-agent systems like OpenClaw by individual users and enterprises, weekly Token calls on the OpenRouter platform have grown rapidly since early 2026, reaching 14.8 trillion in the week of March 2.
Chinese Models Gain Market Share: Cost Advantages Amplify Infrastructure and IDC Resilience
HSBC Qianhai Securities notes that since February 2026, Chinese models such as Zhipu's GLM5, Minimax's M2.5, and Moonshot AI's Kimi K2.5 have been steadily capturing Token market share from US large models, leveraging advantages in performance and cost-effectiveness. According to OpenRouter data, in the week of February 9, the weekly Token call volume share of Chinese models among the top 9 models on OpenRouter exceeded that of US models for the first time. Pricing structures are also driving penetration. An analysis of mainstream AI programming service subscription plans by HSBC Qianhai Securities shows that, under similar usage limits, the monthly subscription pricing of Chinese providers is about 20% to 30% of that of overseas providers, making cost efficiency a key variable driving Token migration. HSBC Qianhai Securities judges that robust growth in Token call volume will increase demand for AI software infrastructure, potentially creating structural opportunities for Hyper-Converged Infrastructure (HCI) and Software-Defined Storage (SDS). Simultaneously, the growth in computing demand driven by the proliferation of AI Agents is expected to boost orders for domestic IDC suppliers.
OpenClaw's Breakout: Agents Don't Replace Software, But Form an Overlying "Calling Layer"
Shenwan Hongyuan Research points out that OpenClaw is positioned as a personal AI assistant, capable of integrating with various communication channels like WhatsApp, Telegram, and Feishu. It manages sessions, tools, events, and skills through a local Gateway, demonstrating that Agents are evolving towards an "executable" product form. Architecturally, OpenClaw uses a local WebSocket Gateway for authentication, orchestration, configuration management, and log processing. Applications and devices connect as nodes, exposing command execution, message sending/receiving, and device capabilities to the Agent. Shenwan Hongyuan concludes that Agents are not a new species detached from existing software, but rather an "intelligent calling layer" and "task orchestration layer" built upon existing software, data, and permission structures. Their core value lies in enhancing automation and process management efficiency. According to Shenwan Hongyuan Research, the previous pressure on the software sector stemmed mainly from AI rewriting valuation frameworks, not from a deterioration in fundamentals. Earlier this year, investors were concerned that general-purpose large models and Agents could bypass traditional SaaS interfaces, directly eroding functional value and user stickiness, even leading to an extreme "SaaS apocalypse" narrative. The marginal change is that deploying enterprise-level Agents is not as simple as "plugging in a model." It requires deep integration with enterprise data and permission controls, setting up security safeguards, and ensuring observability, evaluability, and auditability. Software companies are not passively accepting this challenge; instead, they are leveraging their existing process entry points and data barriers to rapidly productize Agents. ServiceNow explicitly emphasizes that its AI Agents are built-in, integrating AI agents, data, and workflow on a single platform to unify third-party Agents. The software sector currently appears to be in the early stages of a recovery following a sentiment bottom: the pessimistic narrative hasn't completely disappeared, but indiscriminate selling has ceased. The market is beginning to assess which companies face the least impact, or may even benefit from the wave of AI adoption.
Comments