AI应用正当时：国产大模型调用量首超美国，看好大模型出海与商业化落地进展加速

Investment Rating - The report maintains a "Positive" investment rating for the software and services industry [8]. Core Insights - Recent explosive growth in AI model invocation in China, with OpenRouter platform's invocation surpassing that of the US for the first time [2][5]. - The rapid increase in invocation volume indicates a strong collective rise of Chinese AI manufacturers, not reliant on a single blockbuster product [10]. - The report anticipates accelerated commercialization and international expansion of domestic AI models, driven by significant price competitiveness [10]. Summary by Sections Industry Overview - China's AI model invocation reached 41.2 trillion tokens, exceeding the US's 29.4 trillion tokens for the week of February 9-15, 2026 [10]. - The following week saw further growth to 51.6 trillion tokens, marking a 127% increase over three weeks, while US invocation dropped to 27 trillion tokens [10]. Model Performance - In the week of February 16-22, 2026, four out of the top five models by invocation volume were from Chinese manufacturers, contributing to 85.7% of the total [10]. - Notable models include MiniMax's M2.5, which contributed 14.4 trillion tokens, and KimiK2.5, which saw its revenue exceed its total for the previous year shortly after launch [10]. Competitive Landscape - Price competitiveness remains significant, with MiniMax-M2.5 and GLM-5 priced at $0.3 per million tokens, compared to Claude Opus4.6 at $5, highlighting a 16.7 times difference [10]. - The report suggests that the value of tokens is shifting from internet traffic to essential resources in the AI era, indicating a potential new business model [10]. Future Outlook - The year 2026 is projected to be a pivotal year for the commercialization of large models, with key changes anticipated in business models and market dynamics [10]. - The report emphasizes three main themes: new entry points and commercialization of large models, domestic chip development, and the restructuring of software through agents [10].