According to the latest monitoring data released by OpenRouter, the global large model invocation market is undergoing a dramatic paradigm shift "towards the East." Last week (May 18 to May 24), the total global AI large model invocations surged to 28.9 trillion Tokens, up 7.4% from the previous week, showing a strong growth trend that has continued for five consecutive weeks.

In this global competition of computing power and applications, Chinese large models have not only maintained their lead over the US in scale but also shown an astonishing explosive growth rate.

Key Data Insights:

  • Global Total: The weekly global invocations reached 28.9 trillion Tokens, increasing for five consecutive weeks, indicating that AI infrastructure has entered a high-frequency and routine invocation phase.

  • China Leads: The weekly invocations of Chinese AI large models reached 9.223 trillion Tokens, surging by 19.89% from the previous week.

  • US Follows Closely: The weekly invocations of US AI large models were 4.93 trillion Tokens, up 16.27% from the previous week.

  • Leading Position: Chinese large models have exceeded the US in weekly invocations for four consecutive weeks, and the lead is further expanding.

From "Catch-up" to "Leadership": The Three Logics Behind China's AI Invocation Surge

The reason why Chinese large model invocations have achieved an "overtake" and remained the global leader for four consecutive weeks is not just about quantity, but also about a qualitative leap:

1. High-Value "Industrial-Level" API Ecosystem

Compared to the high API prices of US models, Chinese models (such as Zhipu GLM series, DeepSeek, Qwen, etc.) offer highly competitive pricing while maintaining flagship-level performance. This combination of "high intelligence + low latency + high cost-effectiveness" has quickly captured the hearts of developers worldwide, especially in emerging markets.

2. Deep Integration of Developer Ecosystems

OpenRouter data shows that Chinese large models are being more deeply embedded into the workflows of global developers. Whether it's AI Agent development, automated programming, or large-scale content generation, the compatibility of Chinese models with inference environments is improving, making overseas developers more willing to include Chinese models in the core invocation chain of model routing (Model Routing).

3. "Downgrade Attack" on the Application Side

Looking at a series of AI products launched in May, whether it's ByteDance's Lance 3B native multimodal model or Zhipu GLM-5.1's high-speed API of 400 tokens/s, the narrative logic of China's AI industry has shifted from "simply chasing rankings" to "practical cost reduction." Models that can solve real business pain points naturally show higher stickiness in terms of invocations.

Industry Forecast: The Era of "Multipolarization" in AI Infrastructure Has Arrived

Although China leads in invocations, industry analysts point out that this is not simply a "winner-takes-all" scenario, but rather an inevitable result of the multipolarization of global AI infrastructure:

  • Two-Tier System of Chinese and US Models: Global developers are increasingly accustomed to using both Chinese and US models simultaneously in their invocation chains. Choosing the most cost-effective API for different task loads has become a basic operation for developers.

  • Double Circulation of Computing Power and Data: Invocations are not just numbers; they are sources of training data and feedback loops. China's continuous leading invocations mean that Chinese AI is gaining more diverse and globalized linguistic feedback, which in turn accelerates the iteration cycle of Chinese models, forming a positive feedback loop of "increased invocations → improved performance → increased invocations again."