Recent data reveals that Chinese AI models have achieved a milestone by exceeding the United States in token usage volume for the first time. According to statistics from OpenRouter, the world's largest API aggregation platform, during the week of February 9–15, Chinese models processed 4.12 trillion tokens compared to 2.94 trillion tokens for U.S. models, marking a historic reversal. The following week, from February 16–22, Chinese model usage further increased to 5.16 trillion tokens—a 127% growth over three weeks—while U.S. model usage declined to 2.7 trillion tokens.
Even more striking is that four of the top five most-used AI models globally are from China: MiniMax’s M2.5, Moonshot AI’s Kimi K2.5, Zhipu AI’s GLM-5, and DeepSeek’s V3.2. Together, these four models accounted for 85.7% of the total usage among the top five.
Why is this data significant? OpenRouter, with over 5 million developer users, serves as a key barometer for global AI adoption trends. Crucially, the platform's user base is predominantly overseas developers—47.17% from the U.S., while Chinese developers represent only 6.01%. This indicates that the surge in Chinese model usage reflects genuine global developer preference, not internal hype.
What factors contributed to this shift? First is extreme cost efficiency. For example, processing 1 million tokens costs only $0.30 using models from MiniMax or Zhipu AI, whereas comparable overseas models like Claude Opus charge up to $5—approximately 16.7 times more. For startups and developers, this difference is not just about savings but survival.
Second, technical architecture improvements have driven down costs. Many Chinese models employ Mixture of Experts (MoE) architecture, which significantly reduces expenses. Unlike traditional models that activate all parameters for every task—wasting computational power—MoE selectively engages specialized "expert" networks. This on-demand activation cuts GPU memory usage by 60% during inference and increases throughput by 19 times.
Third, the underlying logic of token consumption has shifted from conversational to procedural use. Previously, AI served as a Q&A tool, consuming hundreds to thousands of tokens per dialogue. Now, AI acts as a digital employee, with single tasks potentially consuming hundreds of thousands of tokens.
It is important to note that leading in token usage does not equate to overall technological superiority. Chinese AI development still faces challenges, including a shortage of original foundational theories. This milestone is not an endpoint but a starting point.
Comments