Xinhua Finance Investigation: The Commercialization Test After China's Large Model API Calls Rank First Globally

Article body

Xinhua Finance, Beijing, November 3 — Public data shows that, according to statistics from the global AI model aggregation platform OpenRouter, during the week of October 19 to 25, 2026, the weekly call volume of Chinese AI large models on the platform reached 93.54 trillion tokens, ranking first globally for 26 consecutive weeks, while the platform's total global call volume during that period was 188 trillion tokens; according to broader statistical standards disclosed by the National Data Administration, the China Academy of Information and Communications Technology and others, in early 2024 the national average daily token call volume was only 100 billion, and by March 2026 it had soared to 140 trillion, an increase of more than a thousandfold.

However, whether the leading call volume can be effectively converted into sustainable commercial revenue remains a real question facing the industry.

The valuation logic of the large model track is shifting from "parameter competition and traffic narratives" to a commercial closed loop of "computing power autonomy, paid revenue, and replicable scenarios." The era of relying solely on call volume and leaderboard rankings to support valuations is ending, and concept-type targets lacking commercialization implementation will face valuation regression pressure.

Call volume does not equal commercial value; computing power cost is a hurdle

English text is being prepared.

Wu Hequan, an academician of the Chinese Academy of Engineering, pointed out that agents have driven token consumption, and "token consumption is not honored for being large, but for efficiency." The value of computing power should be measured by the output efficiency of tokens, rather than simply by scale. He said that the share of conversational traffic is still relatively high, and the traffic structure will undergo a fundamental shift in the future, with agent service traffic for production scenarios becoming mainstream. In his view, if traffic cannot settle into reusable workflows on the industrial side, then no matter how high the call volume is, it is only short-term heat and difficult to convert into sustainable corporate revenue.

The explosive growth of token calls on the inference side is pushing up enterprise operating costs, and computing power costs are gradually becoming a heavy burden on most model vendors. Reporter research found that small and medium-sized model teams generally face a real problem: some projects that have won API orders have thin profits after deducting computing power, labor, and operations and maintenance costs, and there is even an inversion of revenue and cost where "the more calls, the greater the loss."

In the view of Ai Zhiyuan, founder and CEO of Qujing Technology, computing power cost is the first unavoidable hurdle for large models to move toward a commercial closed loop. When agents and long texts push call volume to exponential levels, the key to industrial breakthrough lies in tackling underlying system engineering and converting computing power reserves into certain productivity that is affordable and stable to run.

Lack of pricing standards; industrial implementation is the touchstone of commercialization

At present, the industry has not yet formed a unified token pricing standard. Different vendors differ greatly in measurement standards, service boundaries, and output quality, making horizontal price comparison difficult for customers and objectively delaying large-scale procurement.

Deng Zhonghan, an academician of the Chinese Academy of Engineering, has publicly stated that the large-scale implementation of domestic high-end computing power chips still faces practical difficulties such as insufficient software-hardware coordination, high adaptation costs, and low computing power utilization efficiency. Shortcomings in the computing power foundation directly constrain the commercial profitability space of large models.

Customers in the industrial and financial fields contacted by reporters have clearly changed their attitude: from being willing to try things for free in the early stage to paying prudently, with procurement decisions focused on whether models can directly reduce costs and increase efficiency, and refusing to pay simply for technical concepts. At the same time, large model filing has become routine, and security assessment and compliance requirements continue to rise. According to data from the Cyberspace Administration of China, as of June 30, 2026, a total of 988 generative AI services had completed filing, and nearly a thousand filed models have entered an elimination race. The extensive model of relying on free subsidies to exchange for traffic is difficult to sustain.

Zhang Ping, an academician of the Chinese Academy of Engineering, proposed that token measurement and billing cannot be defined by platforms themselves, nor should there be a one-size-fits-all administrative pricing. A unified basic measurement scale should be established, specific prices should be formed by the market, and a task result verification mechanism should be provided to lay a rules foundation for the commercialization of the entire industry.

Relevant policies have already followed. In September, Beijing issued the "Action Plan for Accelerating the Development of the Token Economy in Beijing (2026–2028)", using tokens as a lever to systematically lay out new forms of the intelligent economy. Wang Jiandong, deputy director of the Big Data Development Department of the State Information Center, believes that in the future, tokens should truly assume the role of the measure of the intelligent economy. The idea of purchasing power parity can be borrowed, using standard corpora as an anchor to measure the token compression rate of each model, forming regularly published and recalculable conversion coefficients, so that the "rulers" of different vendors can be mutually recognized.

Industry divergence has already emerged; where should small and medium-sized model enterprises go?

It can be seen that AI has changed from a "concept" into a real revenue engine: Alibaba's quarterly results for the quarter ended June 30, 2026, show that revenue from AI cloud and computing power services in that quarter was RMB 48.437 billion, and both total revenue and external customer revenue accelerated to 45% year-on-year growth. This growth momentum was mainly driven by increased adoption of AI-related products. Revenue from AI-related products reached RMB 12.376 billion, achieving triple-digit year-on-year growth for the 12th consecutive quarter and becoming the core segment of revenue growth; industry divergence has already emerged, and a large number of small and medium-sized model vendors remain in the pilot polishing stage and have not yet formed stable cash flow.

Zhu Shuai, a researcher at the CCID Research Institute, reminded that evaluating AI investment cannot look only at capital expenditure scale, revenue growth rate, or chip procurement volume; it is also necessary to comprehensively examine business structure, funding sources, and operational efficiency. For the vast number of small and medium-sized enterprises, "obtaining stable, safe, and business-appropriate computing power services at a reasonable price is often more practically meaningful than directly building large computing power facilities."

Industry insiders interviewed believe that topping the call volume rankings is a phased achievement of domestic large models' supply capacity, but the industry competition logic has clearly shifted: entities that can control computing power costs, accumulate stable paying customers, and implement replicable industrial solutions are more likely to ride out cycles; while models that rely on free traffic for momentum and lack a real commercial closed loop still have commercial sustainability that remains to be tested by the market. Beneath the prosperous appearance, the commercialization capability of large model enterprises will ultimately be tested by real accounts.

(Source: Xinhua Finance) Middle of the body — embedded advertisement