China's large models have amassed a total of nearly 80 trillion token calls, outperforming the United States for 25 consecutive weeks.

Last week, the total call volume of global AI large models on OpenRouter, the API aggregation platform, reached 164 trillion Tokens. Among them, Chinese models accounted for 79.18 trillion — roughly four and a half times the 17.36 trillion of US models.

According to the latest data from OpenRouter cited by National Business Daily, this figure marks that the weekly call volume of Chinese large models on this platform has exceeded that of the United States for 25 consecutive weeks, firmly ranking first in the world. More notably, all top five positions in the global call volume ranking last week were occupied by Chinese AI models.

At first glance, this is irrefutable evidence that China's artificial intelligence application layer has taken a comprehensive lead. However, an in-depth look at the data structure and business model shows that this "report card" not only reveals the real industrial demand, but also exposes a problem that cannot be ignored: out of such a huge call volume, how much can be converted into real revenue for leading manufacturers?

The Regional Scissor Gap Revealed by the OpenRouter Ranking

OpenRouter is not an ordinary app store, but one of the most mainstream AI API aggregation gateways in the world at present. It acts as an "intermediary" — uniformly encapsulating the inference capabilities of various mainstream large language models in the market, and providing standardized access interfaces for developers.

Continuous data from the platform shows that for more than half a year, Chinese models have maintained a significant advantage over US models in weekly call volume. Behind this gap lies the structural differences in the development paths of the AI industry between China and the United States.

The West still maintains a leading position in R&D of underlying basic models. US models such as the GPT series and Claude are still leading the industry in technological cutting-edge performance. However, in actual application scenarios, the Chinese market, with its huge manufacturing base and active internet ecosystem, has spawned massive API call demands.

Most of these demands come from the daily operations of enterprises. Whether it is e-commerce customer service, intelligent translation, code assistance or content generation, the penetration rate of AI tools among Chinese enterprises is rising rapidly. As an aggregation platform that takes cost performance as an important consideration, OpenRouter has precisely become the main outlet for such demands.

It is worth noting that the platform implements a "zero markup" strategy — directly passing on the pricing of the underlying Model Provider to users, and only making profits by charging a fixed handling fee. This means that the call volume distribution on the ranking largely reflects the real preference of developers, rather than the result of platform algorithm recommendation.

In addition, OpenRouter also supports the BYOK (Bring Your Own Key) mode and flexible routing suffix control, such as options of ":nitro" for priority throughput and ":floor" for priority low price, which further lowers the access threshold and trial-and-error cost for developers. This flexibility enables more small and medium-sized teams to use Chinese models at low cost for product verification and commercialization attempts.

The Cost Account of 79 Trillion Tokens: Who is Paying the Bill?

What is the concept of 164 trillion Tokens? Roughly converted, estimated on the basis that each Token is approximately equal to 0.75 English words, it is equivalent to processing more than 270 trillion tokens per week.

In the 79.18 trillion composition of Chinese models, a considerable part comes from batch calls in low-cost or even free modes. OpenRouter's statistical caliber is based on Token count, not the number of independent requests. The platform supports the BYOK mode, and users have a free processing quota of up to 25,000 US dollars per month, and the handling fee is withheld at a 5% proportion only for the part exceeding the quota.

From the perspective of business logic, this is a typical stage of "subsidy for drainage". Domestic public cloud infrastructure suppliers — Alibaba Cloud, Tencent Cloud, Huawei Cloud, etc. — are locking enterprise customers by reducing the price of inference services. For downstream ISVs (Independent Software Vendors), the call cost has been reduced from several dollars to a few cents, making embedding AI functions into products an economically viable investment.

However, it also brings new contradictions: while low-price competition promotes the popularization speed of AI applications, it also squeezes the profit space of upstream large model manufacturers. When the call volume approaches the level of hundreds of trillions, the unit price difference of every 0.1 cent means a revenue gap of billions. Although this strategy of "burning money for scale" has expanded the market coverage in the short term, in the long run, if a stable payment conversion mechanism cannot be formed, it may eventually lead to the value shrinkage of the entire industrial chain.

High Call Volume ≠ High Revenue: Monetization Model Remains an Unresolved Issue

Large call volume does not mean high revenue, which has become a consensus in the current AI track.

From the perspective of international experience, OpenRouter's own business model is of great reference value. The platform does not earn the price difference of APIs, but realizes monetization through paths such as fixed handling fees when purchasing Credits, premium function subscription services, and exchanging privacy for discounts.

For Chinese large model manufacturers, the real test is how to convert massive calls into sustainable revenue. At present, there are two main payment models in the domestic market: pay-as-you-go for enterprise B-end API calls, and C-end subscription system for individual users.

The advantage of the B-end lies in high customer unit price and strong willingness to renew, but the decision-making cycle is long, and small and medium-sized enterprises tend to choose low-cost solutions rather than high-priced customized models. The C-end has a low threshold and wide coverage, but user stickiness and payment willingness have long been in doubt.

The fact that all top five positions are taken by Chinese models cannot automatically solve this problem. More importantly: out of these calls, how many come from paying enterprise customers? How many come from free trials? How many come from automated tests?

Without a transparent bill disclosure mechanism, it is difficult to judge which party has realized the commercial closed loop solely by the call volume ranking. This is also the biggest information asymmetry that investors are currently facing.

It is worth special attention that with the development of more open-source models, the barriers of the self-built ecosystem of large manufacturers are being weakened. Third-party open-source communities are continuously launching high-quality alternative solutions, enabling enterprises to choose cheaper self-deployed models, thus further compressing the profit space of aggregation platforms.

What to Observe Next

For investors and practitioners who focus on the AI industry, the following four indicators are worthy of continuous tracking:

First, the growth rate of inference revenue of cloud service providers. When the revenue growth rate of inference services exceeds that of training services, it indicates that the market has shifted from "competing for computing power scale" to "competing for application implementation". It is necessary to closely track the changes in the quarterly financial reports of the three major operators and leading technology enterprises.

Second, the enterprise-side payment conversion rate. It is necessary to observe the change in the proportion of API revenue in the financial reports of leading large model manufacturers. If the call volume rises while the revenue stagnates, the price war is still squeezing profits. It is recommended to focus on the changing trend of the number of monthly active paid accounts and the ARPU value.

Third, the AI penetration rate in core industries. If the actual application coverage data of key industries such as manufacturing, finance, and medical care can be obtained, it can be verified whether the "usage dividend" can form a long-term competitive barrier. This requires cross-departmental data and research cooperation, which is difficult but of extremely high value.

Fourth, the commercialization progress of open-source models. With the development of domestic and foreign open-source communities, more high-quality models may no longer rely on a single platform for distribution, thus changing the weight distribution pattern of aggregation platforms. Especially when the self-deployment cost is lower than the API call cost, aggregation platforms may lose their pricing power.

This article is from the WeChat Official Account "BT Business Technology", written by Zhi Yu, and authorized for release by 36Kr.