Jefferies: China's AI enters a period of token consumption expansion, with low-cost models reshaping the competitive landscape.

According to Mars Finance, in its AI series report released on June 22nd, Jefferies stated that as AI usage accelerates its adoption in code, agents, office work, and content generation scenarios, its use is shifting from "chatting" to "workflows," driving rapid growth in token consumption. OpenRouter data shows that in the week ending June 22nd, platform token consumption increased by 4.7% week-on-week, reaching 46.7 trillion. Within this trend, the share of Chinese models is expanding. The report states that, according to OpenRouter's data, Chinese models consumed 18.8 trillion tokens that week, higher than the 5.8 trillion of US models. DeepSeek V4 Flash ranked first with 4.94 trillion tokens, followed by Xiaomi MiMo-V2.5, MiniMax M3, and Tencent Hy3 preview. Behind this change is the more competitive balance that Chinese models are achieving between performance and cost. Jefferies states that the intelligence gap between Chinese and American models is narrowing, while API call costs are only a fraction of those of American models. The report attributes this cost advantage to MoE (Moment of Energy), attention mechanism optimization, and higher model computing power utilization. However, increased token usage does not necessarily translate to a corresponding increase in spending. Jefferies' Silicon Data LLM Token Expenditure Index remained between 1.64 and 1.68 from June 14th to 19th, down from 2.04 on May 31st, indicating that while enterprises and developers are increasing call volume, they are also shifting towards cheaper and more efficient models. Jefferies believes that this combination of "more calls, lower unit price" will benefit companies such as Baidu, Alibaba, Tencent, Kingsoft Cloud, AI Labs, and Kuaishou Keling. Cloud service providers will benefit from the growth in model calls and MaaS demand, while AI applications are expected to find clearer commercialization paths in code, agents, office software, and medium-to-long-form content generation. The report specifically mentions that Volcano Engine is about to hold its FORCE conference, where the market will focus on the daily token consumption of the Doubao large-scale model, subscription package pricing, the code and agent product progress of Seedance 2.0, TRAE, Arkclaw, Coze, etc., as well as MaaS revenue targets. Regarding Tencent, WeChat's AI "Xiaowei" is currently undergoing small-scale testing, and Jefferies predicts it may be officially released in the fourth quarter of 2026. The user base on the application side is also expanding. In the May global ranking of monthly active users for AI mobile applications, ChatGPT still ranked first with approximately 958 million, but Chinese applications have already occupied several top positions, including Doubao, Qwen, Quark, DeepSeek, and Yuanbao. The report believes that the key to the next stage of AI applications in China will shift from user growth to penetration and commercialization in high-frequency scenarios. In the long term, Jefferies, citing data from F&S and CAICT, states that the Chinese large-scale model market is projected to grow at a CAGR of approximately 64% from 2024 to 2030, exceeding RMB 100 billion by 2030. Enterprise-level and on-premises deployments are expected to be the primary revenue drivers.