In February 2026, international market research firm Frost&Sullivan released "China GenAI Market Insights: A Panoramic Study of Enterprise Level Large Model Call, 2025H2". The report shows that in the second half of 2025, the daily average usage of Chinese enterprise level large models will soar to 37.0 trillion tokens, an increase of 263% from the first half's 10.2 trillion. The proportion of top models has increased, among which Alibaba Cloud Qwen has the largest growth rate, rising to 32.1%, almost doubling from 17.7% in the first half of the year, expanding its leading advantage and becoming the most favored large model by Chinese enterprises.

The proportion of daily average adjustment volume of Chinese enterprise level large models in 2025. Compared to the first half of the year, the daily average usage rate of Alibaba Cloud's Qianwen Big Model jumped to 32.1% in the second half of the year, almost doubling. Sullivan Report Screenshot

2025 is a crucial year for large models to transition from laboratory to enterprise production. Based on the analysis of 870 valid questionnaires from key industries such as the Internet, finance, consumer electronics and automobile, this study found that the motivation for enterprises to apply the big model has changed substantially. In the first half of 2025, the main driving force for enterprises to use large models is to improve product performance and customer experience; By the second half of the year, improving the operational efficiency and research and development effectiveness of enterprises will become the primary driving force, and large models will also become a key tool for enterprises to cope with market competition pressure.

In this context, the enterprise level large model market is accelerating its concentration towards top manufacturers. In the second half of 2025, the top 3 daily token consumption of large models are Qianwen (32.1%), Doubao (21.3%), and DeepSeek (18.4%). It is worth noting that in the first half of 2025, the TOP3 will also be these three companies, and at that time, the total proportion of the three will be less than half; By the second half of the year, the total proportion of the three has exceeded 70%. Among them, the growth rate of Alibaba Cloud Qianwen is as high as 14.4%.

According to the report analysis, as the model gradually embeds into the main business chain, the attitude of enterprises towards "multi vendor parallel trial" has significantly decreased. In order to reduce long-term operation and maintenance costs such as interface maintenance, version management, and stability assurance, enterprises tend to streamline the number of suppliers. This also creates higher migration barriers for top vendors with advantages in computing power scheduling, stable delivery capabilities, and deep binding of cloud infrastructure.

Taking Alibaba Cloud, which has the largest market share, as an example, enterprises can directly call the Qianwen big model from Alibaba Cloud through APIs and enjoy MaaS services; The open-source Qianwen model can also be deployed and trained through Alibaba Cloud's IaaS and PaaS platforms. The number of tokens consumed on Alibaba Cloud is much higher than the reported statistics. Token consumption means real money and silver. The performance of the latest open-source Qwen3.5-Plus is comparable to Gemini 3, but the API price is only 1/18 of it, as low as 0.8 yuan per million tokens.

The report also reveals an important trend: the willingness of enterprises to embrace open source has significantly increased. The proportion of closed source users willing to add to open source has doubled, from 22.6% in the first half of 2025 to 48.5% in the second half. At the same time, Chinese manufacturers account for 90.2% of the newly added open source models worldwide in 2025H2, far exceeding overseas markets.

In the current situation where domestic open-source models frequently dominate the charts, competition is no longer limited to single point performance, but extends to a comprehensive competition of size, deployment flexibility, and cost control. Since its open source in 2023, Qianwen has opened more than 400 models, covering full size and full modality, with over 1 billion downloads, over 200000 derived models, and 201 language extensions. It is widely recognized as the world's first open source model.