Abstract: The AI-native steel plant architecture is designed to address prevalent industry pain points in the process of empowering intelligent manufacturing with large language models, such as fragmented architecture, bolted-on AI, and insufficient production safety constraints. The reference architecture comprises three layers from bottom to top: the Digital Asset Interface Layer (DAS), the Intelligent Core Layer, and the Application Agent Layer. Taking an industry-specific large model as the domain-wide intelligent engine, the architecture relies on the DLA digital twin ontology to accomplish structured modeling of production lines and activities, delineate the inference boundaries of the large model, verify generated outputs, and constrain model behavior, thereby achieving a dialectical unity between unleashing large model capabilities and enforcing industrial production safety constraints. This paper analyzes the positioning, core functions, and innovative value of each of the three layers, providing theoretical guidance and an implementation reference for the engineering deployment of AI-native steel plants.