Western Securities has released a research report stating that the commercialization of large AI models has achieved a qualitative leap from "1 to 10." Leading domestic model firms, represented by KNOWLEDGE ATLAS (02513) and MINIMAX-WP (00100), are experiencing exponential growth in paid token consumption and Annual Recurring Revenue (ARR), with particularly rapid user growth in core scenarios such as AI coding. Large model applications are gradually forming a virtuous cycle of "volume and price rising together," validating their commercial value, while model developers are gaining pricing power in key areas like programming. In terms of AI engineering, the emergence of the new concept Harness Engineering, coupled with the inadvertent open-sourcing of Claude Code, is expected to accelerate the large-scale deployment of Agent applications. Key points from Western Securities are as follows:
KNOWLEDGE ATLAS has achieved an exponential increase in paid token consumption, with its open platform API ARR reaching 1.7 billion yuan in March. KNOWLEDGE ATLAS released its 2025 performance announcement: revenue reached 724 million yuan, a year-on-year increase of 131.9%; adjusted net loss was 3.182 billion yuan, widening by 29.1% year-on-year. The company's revenue structure has shifted significantly, with MaaS API revenue占比 rising substantially. In 2025, by business segment, open platform and API business revenue reached 190 million yuan, up 292.6% year-on-year; enterprise-grade intelligent agent revenue grew 248.8% to 166 million yuan; enterprise-grade general large model revenue increased 70.5% to 366 million yuan. The number of paid developers for KNOWLEDGE ATLAS's GLM CodingPlan surpassed 242,000. In February 2026, the company proactively raised prices by 30% and eliminated first-purchase discounts, marking a new phase of "volume and price rising together" for tokens. Additionally, the company launched the one-click installable AutoClaw and introduced the Claw Plan in March 2026, which gained over 100,000 subscribers within just two days and exceeded 400,000 subscribers within 20 days. Following the earnings report, the company disclosed a real-time figure: as of March 31, 2026, the ARR for its open platform API had surged to approximately 1.7 billion yuan (about $250 million), a more than 2.4-fold increase from the roughly 500 million yuan at the end of 2025, and a staggering 60-fold increase compared to 12 months prior.
Previously, MINIMAX-WP also announced that its ARR had exceeded $150 million by February. In 2025, MINIMAX-WP's revenue was $79.04 million, a 159% year-on-year increase. Growth accelerated further at the start of 2026—ARR surpassed $150 million in February, with a revenue run rate nearly double that of the full year 2025. The direct driver of this jump was a surge in usage following the release of the M2.5 model: daily average token consumption for the M2 series text models in February grew over 6 times compared to December of the previous year, with usage in programming scenarios increasing more than 10-fold.
The establishment of the new Harness Engineering paradigm is becoming key infrastructure for the large-scale engineering implementation of AI Agents. In February 2026, OpenAI published a technical blog titled "Harness Engineering: Leveraging Codex in an Agent-First World." The article detailed an experiment where a team of just 3 engineers (later expanded to 7) used CodexAgent to generate over 1 million lines of production-grade code within 5 months, merging approximately 1,500 pull requests, without a single line of code being handwritten by humans. However, what truly ignited industry discussion was not the figure "AI wrote 1 million lines of code" itself, but the proposal of a全新的 engineering paradigm: Harness Engineering. Harness Engineering is a unification of the operating system and software engineering methodology for the AI era, encompassing elements like memory, system prompts, knowledge bases, and orchestration under the Agent paradigm, as well as text flows under the OpenClaw paradigm, such as Agent.md, Soul.md, and User.md, all designed to facilitate better interaction with models. Anthropic subsequently published a blog post titled "Harness design for long-running application development," which described Harness as an external framework, control structure, and orchestration system that supports the operation of complex AI Agents. It is not a single algorithm but a comprehensive set of engineered "scaffolding" used to manage and amplify AI capabilities. While prompts determine the quality of a single interaction, Harness determines the execution flow and reliability of multi-turn, multi-agent, and long-duration tasks. The core function of Harness is to address the "going off the rails" problem when AI performs complex, time-consuming tasks, compensating for inherent model weaknesses (such as context anxiety, self-embellishment) through external control mechanisms. Both OpenAI and Anthropic explicitly identify Harness as critical for the successful implementation of Coding Agents.
The leak of over 510,000 lines of Claude Code source code is expected to further promote the adoption and development of Agents. On March 31, due to an npm packaging error by Anthropic, approximately 512,000 lines of Claude Code source code were inadvertently "open-sourced" to global developers. The leak included 4,756 source files, over 40 tool modules, and several unreleased features. Exposed by a researcher, the code did not involve model weights or user data but revealed the architecture, prompts, tool invocation mechanisms, and even undisclosed features like Kairos persistent processes and undercover mode, providing the most complete external view of the Claude Code architecture to date. This was the company's second similar mistake, raising supply chain security concerns. Although the relevant files have been officially removed, the code has spread within the community, significantly lowering the barrier to entry for AI Agent development and potentially accelerating industry competition and technological innovation.
The large AI model industry is accelerating its transition from a phase of technological exploration to one of large-scale commercial implementation. Firstly, the commercialization process has achieved a qualitative leap from "1 to 10." Leading model firms like KNOWLEDGE ATLAS and MINIMAX-WP are seeing exponential growth in paid token consumption and ARR, with particularly strong user growth in core scenarios like AI coding. More importantly, large model applications are forming a virtuous cycle of "volume and price rising together"—companies are proactively raising prices and users are accepting them, signifying that large models are evolving from "usable" tools into "essential" productivity infrastructure. Their commercial value is being market-validated, and model developers are gaining pricing power in key areas like programming. Regarding AI engineering, the emergence of the Harness Engineering concept and the inadvertent open-sourcing of Claude Code are expected to accelerate the large-scale deployment process of Agent applications.
Regarding investment targets, the following are suggested for attention: KNOWLEDGE ATLAS (02513), MINIMAX-WP (00100), Wangsu Science & Technology (300017.SZ),卓易信息 (688258.SH),寒武纪 (688256.SH),海光信息 (688041.SH),中科曙光 (603019.SH),华丰科技 (688629.SH).
Risks include changes in industrial policy, slower-than-expected technological progress, slower-than-expected AI development, intensifying industry competition, and changes in the international environment.
Comments