China Tops the List for 15 Consecutive Weeks! Why Does Its AI Model Call Volume Outpace the US? OpenRouter's latest weekly token call rankings show that from August 3rd to August 9th, the total global AI model call volume reached 69 trillion tokens. Among them, China's AI model token call volume reached 34.25 trillion, a 21.76% increase week-on-week; during the same period, the US AI model token call volume was 9.17 trillion, a 109.36% increase. This means that on the OpenRouter platform, China's AI model call volume has surpassed that of the US for 15 consecutive weeks, firmly holding the top spot globally. Last week, the top four AI models in terms of global call volume were all from China. The top-ranked model was DeepSeek-V4-Flash-0731 (the official version of DeepSeek-V4-Flash), with a weekly call volume of 8.83 trillion tokens, a surge of 570% compared to the previous week. Tencent's Hy3 model ranked second with 8.05 trillion tokens, a 67% increase week-on-week. OpenAI's GPT-5.6 Luna model ranked fifth. This data should not be interpreted simplistically as "Chinese AI has surpassed the US." However, this ranking reveals a trend: Chinese AI models are demonstrating increasing competitiveness in commercial applications due to their open ecosystem, performance, and cost. There are many reasons for this trend. Currently, the focus of AI industry computing power expenditure is shifting from model training to larger-scale inference capabilities. AI is gradually becoming a fundamental tool in daily enterprise operations, making cost increasingly important. And cost is precisely where Chinese AI models excel. First, the training cost of some leading Chinese models is significantly lower than that of their American competitors. According to publicly available data and industry estimates, the training cost of some Chinese models may be only one-tenth of that of their American counterparts. Second, Chinese models generally adopt a MoE (Hybrid Expert) architecture, which can reduce inference costs by lowering the parameter activation ratio. The activation ratio of mainstream Chinese models in a single inference is generally in the single digits to around 10%, while this ratio for American models reaches 15% to 30%. Third, China also has advantages in power supply, data center construction, and maintenance, further reducing the price of AI services. According to OpenRouter data, the cost of Chinese models is as low as 18 cents per million tokens, while that of American models is $4, creating a cost difference of more than 20 times. Previously, American companies such as OpenAI, Anthropic, and Google held an advantage in the AI ​​competition due to the powerful performance of their closed-source models. However, as AI enters the commercialization stage, companies no longer focus solely on performance when choosing models, but rather consider both performance and cost on the same level, providing new competitive opportunities for Chinese AI. More importantly, Chinese AI models are not only cheaper, but their performance is also continuously improving. While the cost of current Chinese AI models is only a fraction of that of similar American products, their performance is already close to that of top-tier American models, with a gap of only a few months. The possibility of surpassing them is very high.