Zhipu AI, one of China’s leading contenders in the global generative AI race, has signaled a definitive shift in the industry’s commercial playbook. CEO Zhang Peng recently declared that as large language models reach a critical threshold of capability, the Application Programming Interface (API) is no longer just a delivery mechanism but the core business model of the artificial intelligence era. This strategy reflects a broader move away from traditional software licensing toward a utility-based "intelligence economy."
Financial results recently released by the firm bolster this confidence. Zhipu reported an annual revenue of 724 million RMB, representing a 131.9% year-on-year increase. Most notably, the Annual Recurring Revenue (ARR) for its Model-as-a-Service (MaaS) platform surged sixty-fold within a twelve-month period, reaching 1.7 billion RMB. These figures suggest that enterprise appetite for integrated AI capabilities is transitioning from experimental curiosity to deep structural adoption.
Zhang’s strategic logic rests on a formula he calls the "first principle" of AGI: economic value equals the "intelligence ceiling" multiplied by the scale of "token consumption." In this framework, the intelligence ceiling—the maximum capability of the model—dictates pricing power and competitive moats. Meanwhile, the volume of tokens, which are the fundamental units of data processed by these models, determines the overall market volume. This paradigm seeks to turn AI from a luxury tool into a pervasive production factor.
This shift mirrors the evolution of Western AI pioneers such as Anthropic. By delivering high-capability models via API, Zhipu aims to transform intelligence into a tradeable resource. This marks a departure from the "one-off" software sales or project-based consulting of the past, moving toward a continuous economic model where AI infrastructure powers digital transactions and industrial processes at scale.
Looking toward 2026, Zhang anticipates a transition from "Vibe Coding"—lightweight, vibes-based AI experimentation—to "Agentic Engineering." In this stage, AI evolves into autonomous digital engineers capable of self-planning, environment sensing, and iterative improvement. This evolution is expected to trigger a second exponential jump in token consumption as autonomous agents perform complex, multi-step tasks without human intervention.
Zhipu is also making a play for the foundational layer of computing with the concept of the "LLM-OS," or Large Model Operating System. Zhang argues that while traditional operating systems manage hardware resources, the next generation of platforms will manage intelligence. In this vision, the future of computing is not a stack of individual applications but a coordinated matrix of APIs and Agents where the model serving as the system kernel holds the ultimate power of definition.
Addressing the competitive landscape, Zhang noted that the tide is turning against private, on-premise deployments. As model iterations accelerate to three-to-four-month cycles, maintaining private infrastructure has become a liability for many enterprises. Consequently, customers are increasingly migrating toward cloud-based APIs to ensure they remain at the cutting edge. Furthermore, Zhang remains confident that independent AI labs can outpace tech giants due to their singular focus and lack of internal legacy constraints.