International market research firm Frost & Sullivan recently released the report "China GenAI Market Insight: Enterprise-Level AI Large Model Call Panorama Study, H2 2025". Data shows that in the second half of 2025, the average daily call volume of enterprise-level large models in China surged to 37.0 trillion tokens, a 263% increase compared to 10.2 trillion in the first half of the year, marking a phased leap from initial experimentation to large-scale application.
Top Player Landscape: Top Three Account for Over 70%, Alibaba Cloud Qwen Doubles Its Growth
In terms of market share, the top three enterprise-level AI large models by average daily token consumption in the second half of 2025 were Alibaba Cloud Qwen, Doubao, and DeepSeek, accounting for 32.1%, 21.3%, and 18.4% respectively, with the combined share of the top three exceeding 70%. In the first half of 2025, although the top three were the same companies, their combined share was still less than half.
Among them, Alibaba Cloud Qwen showed the most significant growth, with its market share jumping from 17.7% in the first half to 32.1%, almost doubling its share and further expanding its leading advantage. This data reflects the continuous evolution of enterprises' comprehensive consideration of performance, stability, and cost during model selection.
Qualitative Shift in Application Drivers: From Enhancing Product Experience to Driving Operational Efficiency
The report points out that 2025 is a crucial year for large models moving from the laboratory to enterprise production. Based on an analysis of 870 valid questionnaires from key industries such as internet, finance, consumer electronics, and automotive, the study found a fundamental change in the drivers for enterprises adopting large models.
In the first half of 2025, the main driver for enterprises using large models was to improve product performance and customer experience; by the second half of 2025, improving enterprise operational efficiency and R&D effectiveness had become the primary driver, with large models becoming a core tool for enterprises to cope with market competition pressure. Currently, large models are widely covering core business areas such as content production, knowledge acquisition, data analysis, and R&D support. The top five application scenarios are text content creation (19.4%), Q&A assistants (13.7%), data processing and analysis (12.9%), internal knowledge Q&A (12.6%), and code generation and assistance (12.4%).
Growth Drivers: Dual Engines of Agent Implementation and Engineering Capability
The report analysis suggests that the explosive growth in enterprise-level large model call volume in China mainly comes from two driving forces: first, the accelerated implementation of Agent-related architectures on the enterprise side, driving large models from one-time Q&A to task-oriented continuous calls and tool execution; second, the continuous improvement of model capabilities and supporting engineering, driving more teams and business processes to stably use large models, simultaneously increasing user scale and call intensity per user.
From a business scenario perspective, new enterprise calls mainly come from two types of expansion needs: one is capacity expansion for core systems and external services. These scenarios emphasize stability, continuity, and accountability, so new calls often still prioritize closed-source models. The other is the addition or rapid expansion of internal efficiency tools and business support applications. These scenarios are more cost-sensitive and have relatively controllable stability requirements, making them more suitable for open-source models.
Open Source vs. Closed Source: Migration Willingness Continues to Increase, Closed-to-Open Source is the Mainstream Direction
The report shows that the willingness of closed-source users to add open-source configurations has increased from 22.6% to 48.5%, while the willingness of open-source users to add closed-source configurations has increased from 5.0% to 7.5%. The overall migration direction remains primarily from closed-source to open-source. Token consumption means real financial investment. Taking Alibaba Cloud, the market leader, as an example, its latest open-source Qwen3.5-Plus performs comparably to Gemini 3, but its API price is only 1/18th of models with equivalent performance, as low as 0.8 yuan per million tokens, demonstrating a significant price advantage.
Global Comparison: Chinese Vendors Dominate Open-Source Pace
From a global perspective, the release strategies of global AI vendors showed significant divergence in the second half of 2025. The report indicates that Chinese vendors, relying on large-scale investment, dominate the pace of open-source releases. In the second half of 2025, Chinese vendors accounted for 90.2% of the newly added open-source large models globally, far exceeding overseas markets. Meanwhile, overseas vendors continue to focus on the closed-source mainline. This pattern reflects the differentiated choices in AI technology routes and business models between China and the US, and also provides abundant underlying model resources for the further expansion of enterprise-level AI applications in China.
