Open Source Securities: Agents are expected to drive a new round of tokenization, with domestic models transitioning to an S-curve and increasing market size.

Open Source Securities released a research report stating that according to Gartner data, the global AI industry is expected to reach $1.8 trillion by 2025, with AI models accounting for 1% of the market share. The future compound annual growth rate (CAGR) for the next 2 years is projected to be 74%. Additionally, according to IDC data, the number of active Agents is expected to grow rapidly from approximately 28.6 million in 2025 to 2.216 billion in 2030, indicating significant potential. The top 3 players hold over 50% market share, with domestic large models emerging as new disruptors. In mid-February, MiniMaxM2.5 tokens consumption ranked first in the OpenRouter list. Mainstream large models still have room for improvement in practical scenario solutions and long-context learning. The evaluation of AI in the second half may exceed training, signaling a shift from the paradigm of "bigger but comprehensive" to "precise and accurate". OpenAI vs Anthropic: Growth momentum of B-end+Coding mode, more efficient utilization of computing power From 2023 to 2025, Anthropic's Annual Recurring Revenue (ARR) is expected to grow at a rate of 10 times per year, while OpenAI is around 3 times. In February 2026, OpenAI achieved an ARR of $25 billion, with Anthropic reaching $20 billion in early March, continuing to catch up with OpenAI. Both companies have a long-term gross profit margin outlook of over 70%, with the unexpected increase in reasoning computing power leading to a downward revision in gross profit margin outlook. By 2024, the computing power costs associated with OpenAI's training (R&D expenses) and operations (inference costs/operational costs) are projected to account for 125% of its revenue, with OpenAI investing significantly more in computing power than Anthropic. By 2028, Anthropic's revenue volume is expected to approach that of OpenAI, while requiring only about one-third of OpenAI's computing power investment. B/C-end: Accelerated penetration of B-end domestic models in the top tier C-end: ChatGPT's monthly visits have consistently exceeded 5.5 billion, significantly leading other competitors, while Gemini continues to catch up with its multimodal capabilities. The Chinese large AI model market is growing rapidly, with enterprise-level market leaders emerging, capturing a 32% market share in the second half of 2025. The integration of Agent with the Spring Festival bonus has driven an increase in monthly activity, with Alibaba, ByteDance, and Tencent leveraging their respective resource endowments with different strategies. B-end: The market concentration of API calls is high, with Anthropic dominating a significant portion. Domestic large model enterprises are gradually moving into the top tier, and the current landscape is expected to shift. Leveraging the advantages of "open source + cost," domestic models are poised to maintain long-term competitiveness. Valuation: With each valuation update, Anthropic's valuation typically doubles or triples from the previous assessment, driven by the rapid growth of ARR. In the early stages (May 2023), when the business was essentially at a very low level, the valuation reached 46 times that year. Since 2025, Anthropic's valuation has peaked at over 20 times, typically dropping below 10 times the following year. In the context of large models in China, using KNOWLEDGE ATLAS as an example, if Anthropic's valuation is taken as a benchmark, the CAGR of KNOWLEDGE ATLAS's revenue from 2025 to 2027 would need to reach 176%. Considering that the growth of models driven by agents is not a simple linear progression, the ARR growth of domestic large models is expected to fully absorb the current valuation levels. Investment recommendation: Bullish on investment opportunities in models, AI infrastructure, and computing power in the context of Agent index upgrades OpenClaw accelerates the penetration capability of Agents, with the ability of domestic large models gradually reaching the top tier globally. Beneficiaries in this sector include BABA-W (09988), BIDU-SW (09888), MiniMax-W (00100), KNOWLEDGE ATLAS (02513), fourth paradigm (06682), BAIRONG-W (06608) in the AI infrastructure layer, and KINGSOFT CLOUD (03896), GDS-SW (09698) in the computing power sector. Risk warning: Model iteration falling behind expectations, intensified industry competition, and Agents failing to land as expected.