China Merchants Securities: Token consumption continues to explode, and token operations ushered in an inflection point in development

The Zhitong Finance App learned that China Merchants Securities released a research report saying that in the multi-model era, token operation is expected to become an important infrastructure connecting model supply and AI applications. As the number of basic models continues to increase and different models continue to differentiate in terms of capacity, cost, speed, and stability, it is gradually becoming the norm for downstream developers to call multiple models and multiple providers at the same time. The token operation platform integrates models through a unified API and further provides services such as model selection, intelligent routing, token measurement, unified settlement, and corporate governance. Its role is evolving from simple model access tools to AI traffic distribution and enterprise call control layers. The bank believes that as the scale of token calls and usage scenarios continue to expand, the operating value of tokens is expected to increase further.

The main views of China Merchants Securities are as follows:

Industry space: Token calls are growing rapidly, and the token operation market is still penetrating in the early stages

Gartner predicts that the global token operating market expenditure will grow from about 251 million US dollars in 2026 to 429 million US dollars in 2027, and is expected to reach about 1 billion US dollars in 2028, corresponding to a three-year CAGR of close to 100%; currently, only about 15% of software engineering leaders have deployed relevant platforms, 31% are being deployed, 22% are still in the evaluation stage, and there is still a lot of room for improvement in industry penetration. The bank estimates that in 2026, models carried by the Token operating platform will use GMV of about 43-6 billion US dollars worldwide, accounting for 15% to 21% of global generative AI model spending. Domestic demand is also growing rapidly. In 2025, China's annual token call volume was about 221 trillion yuan, and in March 2026, the average daily call volume further exceeded 140 trillion yuan.

Business model: The characteristics of being asset-light are obvious, and traffic scale, route data, and enterprise stickiness build core barriers

Token operators usually do not directly bear basic model training and large-scale GPU capital expenses. Currently, they mainly form charging models such as fees, subscription systems, and “subscription+credits”. Compared to the lower technical threshold of the unified API itself, the real competitive advantage of the platform comes from a positive cycle formed after large-scale operation: the richer the models and providers, the more developers can be attracted; after the scale of token calls is increased, the continuous accumulation of call data also helps optimize model selection and routing efficiency, and further increases customer migration costs through enterprise functions such as budget, permissions, logs, security, and governance.

Overseas case: OpenRouter rapidly grew from a model aggregator to an AI traffic distribution portal

OpenRouter connects different models and providers through a unified API, and gradually establishes a two-tier system of model routing and provider routing. As of September 2026, the platform has connected 500+ active models and 80+ providers, serving more than 10 million developers and enterprises, and processing more than 10 trillion tokens per day. The platform uses Credits for unified settlement and mainly charges a 5.5% platform service fee; as more and more users leave the model and provider selection to the platform, what OpenRouter masters is no longer just a model access portal, but participates in the actual distribution of token traffic between different models and inference service providers.

Domestic layout: extending from model access to token supply and application scenarios

Tencent and silicon-based mobility focus on model aggregation, reasoning, and token supply; EasyDianxia, Maifushi, and Jihong Co., Ltd. enter from marketing, corporate services, and cross-border e-commerce scenarios to embed model call, token measurement, and agent scheduling into business processes. Rongxin culture relies on content and IP resources to explore the tokenized use of intellectual assets, while paradigm intelligence extends from enterprise AI infrastructure to model services, API calls, and agent platforms. The bank believes that domestic token operations have gradually evolved from a single model access to multiple paths of “infrastructure+enterprise application+vertical content”. As agents and enterprise AI applications deepen into business processes, the importance of token measurement, scheduling, and scenario-based monetization capabilities is expected to further increase.

Risk warning: AI applications and token calls fall short of expectations; demand for multi-model calls and intelligent routing falls short of expectations; industry competition intensifies; commercialization progress falls short of expectations; data security and compliance risks.