According to a research report, the current focus of the AI industry is gradually shifting from the training era to the inference era, moving from competition over parameters to competition over tokens and agents. The report maintains a positive outlook on core AI model providers with capabilities in model development, developer ecosystems, and application deployment, including Knowledge Atlas (02513), MINIMAX-WP (00100), BABA-W (09988), and TENCENT (00700). The key points of the report are as follows.
Key Recent Developments
On the evening of June 13th, Knowledge Atlas officially launched and previewed the open-sourcing of its new-generation model GLM-5.2, concurrently introducing the ZCode3.0 intelligent coding agent product.
Supportive policies for technological innovation continue to be strengthened. On June 11th, the China Securities Regulatory Commission released measures to deepen reforms of the STAR Market and serve technological self-reliance. Simultaneously, the People's Bank of China and the Ministry of Science and Technology announced the first batch of special relending projects worth 320 billion yuan to support technology finance, reinforcing support for the tech industry.
On June 13th, following directives from the U.S. government related to export controls, Anthropic suspended access to its most advanced models, Fable 5 and Mythos 5. These directives require restricting foreign user access to relevant models.
Domestic Open-Source Models Edge Closer to the Frontier
According to disclosures from Knowledge Atlas, GLM-5.2 supports an ultra-long context of 1 million tokens and shows improved performance in long-range reasoning, code generation, and agent tasks. It is scheduled for official open-source release next week under the MIT license. Concurrently, Knowledge Atlas launched ZCode 3.0, built upon GLM-5.2, further enhancing agent programming capabilities.
From an industry perspective, the current round of model competition is gradually shifting from pure benchmark comparisons to competition in agent capabilities, long context, multimodality, and practical commercial application deployment. With the launch of GLM-5.2, domestic models are seen as continuing to close the gap with the global top tier in core areas such as code agents, enterprise applications, and developer ecosystems.
Token Demand Surges as Improved Model Capabilities Translate to Real Usage
Based on ongoing tracking of OpenRouter data, weekly token usage on the platform as of June 9th has surged over 598% compared to the beginning of the year, representing a several-fold increase over 2025 levels. The daily token consumption scale in the Chinese market has rapidly grown from hundreds of billions at the start of 2024 to the current level of hundreds of trillions.
The report posits that the industry's growth logic has progressively shifted from being training-driven to inference-driven and is now moving further into a commercialization-driven phase. The execution of agent tasks, multi-turn tasks, code generation, and enterprise-level application deployments significantly increase token consumption per user, solidifying tokens as the most critical production factor and value carrier in the AI era.
Domestic Models Transition from "Usable" to "Superior"
On one hand, domestic models represented by Knowledge Atlas, MiniMax, DeepSeek, Tongyi, Doubao, and Hunyuan continue to enhance their capabilities in reasoning, multimodality, and agent functionality.
On the other hand, domestic models hold a clear advantage in inference costs. Currently, the API costs of mainstream domestic models are generally only 1/5th to 1/20th of those of top-tier overseas proprietary models, offering stronger economic benefits during enterprise-scale deployment.
With ongoing advancements in Mixture-of-Experts (MoE) architectures, quantization techniques, adaptation to domestic computing power, and engineering optimizations, this cost advantage is expected to widen further. This is likely to encourage more enterprise clients to transition from pilot testing to formal production stages.
Policy and Capital Synergy Accelerates AI Infrastructure Development
Considering recent policies related to technology finance, technological innovation, and capital markets, the report suggests a positive feedback loop has formed in China's AI industry: "improved model capabilities — expanded application deployment — growing token demand — sustained capital investment."
This round of AI industry upgrade carries stronger strategic significance. From computing power and models to applications, the sector is receiving sustained top-down support. Recent supportive policies for technological innovation, the expansion of industry funds, and increased capital market backing for hard-tech assets are expected to further reinforce the long-term growth rationale of the domestic AI industry chain.
Potential Risks to Consider
Risks include changes in industry policy and regulation, potential shortfalls in technology R&D, intensifying market competition, earnings falling short of expectations, calculation inaccuracies, and delays in data updates.
Comments