Orient Securities: Optimistic about model intelligence to accelerate transformation and catalyze commercialization to the next level

The Zhitong Finance App learned that Orient Securities released a research report saying that it is still recommended to focus first on model manufacturers' technical iterations and ARR data post-testing. In particular, the intergenerational improvement of model performance can bring opportunities for ARR to grow nonlinear. In addition, the commercial growth of AI coding and AI data labeling continues to benefit from the improvement of the base model. The bank believes that attention should be paid to the scarce value of existing model manufacturers, as well as companies leading in multi-modal technology and closed commercialization.

Orient Securities's main views are as follows:

The total ARR of the two major overseas model manufacturers has exceeded 100 billion US dollars. The recent release of new models is expected to continue to drive upward commercialization

According to SemiAnalysis data, the ARR of Anthropic and OpenAI in June '26 is estimated to be 62 billion and 43.5 billion US dollars respectively, and the total of the two companies has already exceeded 100 billion US dollars. Compared to OpenAI's focus on the consumer market (about 950 million active users), Anthropic's revenue comes from API calls, and its net monthly ARR increase since March '26 has basically stabilized at over $10 billion.

In July, the two model makers also released new models one after another. OpenAI launched a three-level version of GPT-5.6. The hierarchical strategy targets specialization, balance, and daily tasks, and the core focus is to achieve better unit task costs

In terms of product architecture, OpenAI also incorporates the three modes of Chat, Work, and Codex into the same app, integrating capabilities such as conversation, coding, and long-range task execution. Anthropic released Claude Opus 5, with a level of intelligence close to Fable 5 and only half the price of the latter. The bank believes that judging from the manufacturer's strategy, OpenAI is more clearly working for developers and enterprise-level customers. After the new model and product architecture are launched, it is expected to further accelerate commercialization.

The commercialization slope of domestic model manufacturers accelerated upward, and Smart Spectrum achieved the ARR target by the end of the year ahead of schedule

Domestic model makers' ARR slope performance is still quite impressive. Smart Spectrum ARR increased from US$250 million in March to US$1 billion in July, achieving the previous target of US$10-15 billion by the end of the year ahead of schedule. Dark Side of the Moon ARR went from 100 million US dollars in March to surpassing 200 million US dollars in May to reaching 300 million US dollars in mid-June (of which API revenue already accounts for more than 70%), and the release of the K3 model brought about several times the rapid growth of corporate ARR. MiniMax ARR doubled from US$150 million in February to US$300 million in May, and pre-training of a new generation of 2.7 trillion parametric models is progressing smoothly. The bank believes that domestic models have more R&D accumulation and engineering implementation experience in architectural efficiency, which is transmitted to the training stage with higher efficiency in the use of computing power, while the inference stage is more cost-effective inference costs. Breaking through intelligent boundaries can bring commercialization to the next level at the same time.

Mercor ARR surpassed $2 billion, and high-quality inference trajectory data in the field of data labeling is still scarce

Mercor is one of the world's largest data platforms for AI experts. Unlike traditional data labeling, Mercor provides guidance for professionals such as doctors, lawyers, financial analysts, etc. to evaluate, correct, score, and reason model outputs. The bank believes that with the gradual exhaustion of Internet text data, the model's ability to perform certain professional tasks has improved, depending on the reasoning trajectory and professional evaluation of these tasks during the RL stage. Mercor ARR doubled its growth during the year. As long as model makers continue to iterate, demand for professional data labeling will be the driving force behind its commercial growth.

Risk Alerts

Large model technology iteration falls short of expectations, big model commercialization progress falls short of expectations, competition increases risk