China Galaxy Securities Co., Ltd. has released a research report stating that model parity is expected to persist, with the open-source ecosystem fueling growth in the application sector. In the evolution of artificial intelligence, the development of the open-source ecosystem has effectively promoted model parity and accelerated the expansion of application end-markets. The significant cost advantage of open-source inference models lowers the technical entry barrier for enterprises, fostering innovation vitality in applications and steering the AI application ecosystem into a virtuous development cycle. The firm believes that, based on the recent strategies of domestic internet giants, the trend of converging model performance is likely to continue, and the open-source ecosystem will persistently drive AI application development. Key points from China Galaxy Securities Co., Ltd. include:
Kimi K3 launched, with multiple benchmark tests aligning with top-tier closed-source models
Moonshot AI released Kimi K3 on July 17, 2026. In various benchmark tests related to coding capabilities and Agent functionality, its performance matches or even surpasses leading closed-source models. The firm believes that the narrowing performance gap between open-weight and closed-weight foundation models could enhance the competitive advantages of local internet platforms. These companies, possessing proprietary datasets and large user bases, may benefit from open-source models with capabilities close to those of top-tier closed-source models, potentially amplifying their market position.
Data resource constraints may lead to continued model performance convergence
In the coming years, publicly available human-generated data is likely to become a hard constraint on the iteration of advanced large models. Although synthetic data can partially alleviate data shortages, China Galaxy Securities Co., Ltd. judges that its lower quality may limit further model performance improvements. Based on this, the firm holds that: 1) The performance gap between open-weight and closed-source proprietary models is likely to continue narrowing; 2) As high-quality training data becomes scarcer, competition for proprietary closed-loop data, such as payments, e-commerce, and social media feeds, will intensify; 3) The industry’s long-held "bigger is better" model scaling approach may gradually diminish in effectiveness, with data supply bottlenecks replacing computing power as a key constraint on large model development.
Models trend toward commoditization, enhancing data holders' bargaining power
As foundation models increasingly become commoditized, the center of value distribution is shifting from the infrastructure layer to the application layer. With model performance converging, the core competitive differentiator for enterprises will no longer be based on proprietary model development capabilities but on exclusive data, traffic distribution channels, and user stickiness. This industry shift favors internet platforms that possess unique datasets and large existing user bases. These companies are optimally positioned to commercialize AI technical capabilities into profitable services.
Continuous token consumption growth supports long-term industry expansion
The global token usage center has been steadily rising, with weekly token consumption increasing from 3-4 trillion tokens in August 2025 to 60 trillion tokens by July 2026. The ongoing increase in token calls directly drives a surge in demand for API interface calls across downstream applications. This demand further transmits upward to model foundation manufacturers. As the core carriers for token computation and inference, model foundations handle massive industry-wide token processing tasks and API call loads. Their service revenue and economies of scale are deeply tied to token consumption volume and API call frequency. The sustained release of token demand can amplify the reuse value of model foundations, effectively dilute unit costs for model training and technical maintenance, and continuously strengthen the technical barriers and market competitive advantages of model foundation manufacturers. This serves as a core driver for the long-term growth of the model foundation industry.
Risk warnings
Risks include intensified market competition, slower-than-expected development of AIGC technology and applications, and risks related to the commercialization progress of AI applications.
Comments