Orient Securities has released a research report advising investors to first focus on model vendors' technological iterations and ARR data validation. In particular, generational improvements in model performance could create non-linear growth opportunities in ARR. Additionally, the commercial growth of AI coding and AI data annotation continues to benefit from foundational model enhancements. The firm believes that the scarcity value of existing model vendors should be emphasized, along with companies that have a technological edge in multimodality and a closed commercial loop.

Overseas model vendors' combined ARR surpasses $100 billion, with new model launches expected to further drive commercialization

According to SemiAnalysis data, as of June 2026, the ARR for Anthropic and OpenAI is projected at $62 billion and $43.5 billion respectively, with their combined total exceeding $100 billion. While OpenAI primarily targets the consumer market (with approximately 950 million weekly active users), Anthropic generates about 75% to 85% of its revenue from API calls. Since March 2026, Anthropic's monthly net new ARR has consistently remained above $10 billion. In July, both model vendors released new models. OpenAI launched three tiered versions of GPT-5.6, with a segmentation strategy targeting specialized, balanced, and routine tasks, aiming to achieve better unit task costs. In terms of product architecture, OpenAI integrated three modes—Chat, Work, and Codex—into a single app, consolidating capabilities for conversation, coding, and long-term task execution. Anthropic released Claude Opus 5, which offers intelligence levels close to Fable 5 at half the price. From the vendors' strategies, OpenAI is clearly pivoting toward developers and enterprise clients. With the new model and product architecture integration, further acceleration in commercialization is expected.

Domestic model vendors' ARR slope accelerates upward, with Zhipu hitting its year-end ARR target ahead of schedule

The ARR slope for domestic model vendors remains impressive. Zhipu's ARR grew from $250 million in March to $1 billion in July, achieving its previously set year-end target of $1 billion to $1.5 billion ahead of schedule. Moon's Dark Side's ARR rose from $100 million in March to over $200 million in May, reaching $300 million by mid-June (with API revenue accounting for over 70%). The release of the K3 model led to several times rapid growth in enterprise ARR. MiniMax's ARR doubled from $150 million in February to $300 million in May, and pre-training for its new 2.7 trillion-parameter model is progressing smoothly. Domestic models have accumulated more research and development experience and engineering implementation expertise in architectural efficiency. This translates to higher utilization of computing power during the training phase and more cost-effective inference costs during the inference phase. As each company breaks through the intelligence ceiling, commercialization can simultaneously reach a new level.

Mercor's ARR surpasses $2 billion, with high-quality reasoning trajectory data in the data annotation field remaining scarce

Mercor is one of the world's largest AI expert data platforms. Unlike traditional data annotation, Mercor provides professionals such as doctors, lawyers, and financial analysts to evaluate, correct, score, and guide model outputs through reasoning. As internet text data is gradually exhausted, the improvement of models in specific professional tasks relies on reasoning trajectories and professional evaluations during the RL stage. Mercor's ARR has doubled within the year. As long as model vendors continue to iterate, the demand for professional data annotation will remain a driving force for its commercial growth.

Risk warnings

Technological iteration of large models may fall short of expectations; commercialization progress of large models may be slower than anticipated; and risks of intensifying competition.