Western: The industrialization process of AI is accelerating, and domestic leading model manufacturer ARR has achieved a leap.

Western has issued a research report stating that the commercialization process of AI large models has achieved a "quantum leap" from 1 to 10, with Chinese leading model manufacturers such as KNOWLEDGE ATLAS (02513) and MiniMax-W (00100) experiencing exponential increases in paid token consumption and ARR. Particularly in core scenarios such as AI Coding, user growth has been rapid. The application of large models is gradually forming a virtuous cycle of "quantity and price rising," with their commercial value being validated by the market, and the manufacturers of large models gaining pricing power in core scenarios such as programming. In terms of AI engineering, the emergence of Harness Engineering as a new concept, as well as the passive open-source of Claude Code, is expected to accelerate the deployment process of Agent applications on a large scale. Western's main points include: KNOWLEDGE ATLAS's paid token consumption has experienced an exponential increase, with the ARR of the open platform API reaching 1.7 billion RMB in March. KNOWLEDGE ATLAS released its 2025 performance announcement: In 2025, the company achieved a revenue of 724 million RMB, a year-on-year increase of 131.9%; adjusted net loss was 3.182 billion RMB, a 29.1% increase year-on-year. The company's revenue structure has undergone a transformation, with MaaS API revenue accounting for a significantly higher proportion. In 2025, in terms of business segments, the revenue from the open platform and API business reached 190 million RMB, a year-on-year increase of 292.6%; enterprise-level intelligent body revenue increased by 248.8% to 166 million RMB; enterprise-level general large model revenue increased by 70.5% to 366 million RMB. KNOWLEDGE ATLAS's GLM CodingPlan has exceeded 242,000 paid developers, and in February 2026, the company proactively raised prices by 30% and canceled first purchase discounts, entering a new phase of "quantity and price rising." Additionally, the company released AutoClaw for one-click installation, and in March 2026 launched Claw Plan, with over 100,000 subscribers within two days of launch and over 400,000 subscribers within 20 days. After the financial report, the company announced real-time data: as of March 31, 2026, the ARR of the open platform API had soared to approximately 1.7 billion RMB (about 250 million USD), more than 2.4 times the amount at the end of 2025, and an increase of about 60 times compared to 12 months ago. Previously, Minimax also announced that the company's February ARR had exceeded 150 million USD. In 2025, MiniMax's revenue was 79.04 million USD, a year-on-year increase of 159%. In the beginning of 2026, the growth rate further accelerated - with February ARR exceeding 150 million USD, the revenue run rate approaching twice that of the previous year. The direct driver of this jump is the high increase in usage after the launch of the M2.5 model: the daily token consumption of the M2 series text models in February increased by over 6 times compared to December of the previous year, and by over 10 times in the programming scenario. The establishment of the new paradigm of Harness Engineering is becoming a key infrastructure for the large-scale engineering deployment of AI Agents In February 2026, OpenAI released a technical blog titled "Harness Engineering: Leveraging Codex in an Agent-First World." The article revealed an experiment: a team consisting of only three engineers (later expanded to seven) generated over 1 million lines of production-level code using Codex Agent within 5 months, merging about 1500 pull requests, with no code written by humans. However, what really ignited industry discussions in this article was not the number of "1 million lines of code written by AI" itself, but a new engineering paradigm it proposed: Harness Engineering. Harness Engineering is a unified body of the operating system and software engineering methodology of the AI era, including aspects such as memory under the Agent paradigm, system prompts, knowledge bases, scripting, etc., as well as text flows under the OpenClaw paradigm, such as Agent.md, Soul.md, User.md, etc., all designed to facilitate better communication with models. Anthropic subsequently published a blog post titled "Harness design for long-running application development," which mentioned that Harness refers to an external framework, control structure, and orchestration system that supports the operation of complex AI agents. It is not a single algorithm but a set of engineered "scaffolding" used to manage and amplify the capability of AI. Prompts determine the quality of a single conversation, while Harness determines the execution flow and reliability of multi-round, multi-agent, long-duration tasks. The core function of Harness is to address the problem of "going off the rails" when AI is completing complex and time-consuming tasks, compensating for the inherent flaws of the model (such as contextual anxiety, self-glorification) through external control mechanisms. Both OpenAI and Anthropic explicitly recognize that Harness is the key to the landing of Coding Agents. Claude Code's leaked source code of over 512,000 lines is expected to further drive the dissemination and development of Agents On March 31, due to a packaging error in an npm package, Anthropic had approximately 512,000 lines of Claude Code source code leaked, including 4756 source files, over 40 tool modules, and several unreleased features, which were then passively "open-sourced" to global developers. The code, exposed by a researcher, did not involve model weights or user data, but it revealed architecture, prompts, and tool invocation mechanisms, as well as previously undisclosed features such as Kairos persistent processes and undercover modes, providing the most complete view of the Claude Code architecture to date. This was the company's second similar mistake, raising questions about supply chain security. While the official had removed related files, the leakage significantly lowered the development threshold for AI Agents, potentially accelerating industry competition and technological innovation. The AI large model industry is accelerating from a period of technical exploration towards scaled commercial implementation Firstly, the commercialization process has achieved a "quantum leap" from 1 to 10. Leading model manufacturers such as KNOWLEDGE ATLAS and MiniMax, have seen exponential growth in paid token consumption and ARR, especially in core scenarios such as AI Coding, with a rapid increase in users. What is more important is that the application of large models is gradually forming a virtuous cycle of "quantity and price rising" - enterprises proactively raising prices and users accepting them, marking the transition of large models from "usable" tools to "necessary" productivity infrastructure. Their commercial value is being validated by the market, and the manufacturers of large models have gained pricing power in core scenarios such as programming. As for AI engineering, the emergence of the new concept of Harness Engineering and the passive open-source of Claude Code are expected to accelerate the deployment process of Agent applications on a large scale. Targets to watch: Recommended to focus on: KNOWLEDGE ATLAS (02513), MiniMax-W (00100), Wangsu Science & Technology (300017.SZ), Jiangsu Eazytec Co., Ltd. (688258.SH), Cambricon (688256.SH), Hygon Information Technology (688041.SH), Dawning Information Industry (603019.SH), Sichuan Huafeng Technology (688629.SH). Risk Warning: Changes in industry policies, unexpected technological advancements, AI development falling short of expectations, intensified industry competition, and changes in the international environment.