Limited by traditional computing paradigms, ultra-large-scale finite element analysis has long faced challenges associated with mesh refinement, including the curse of dimensionality (the number of degrees of freedom grows rapidly with mesh refinement) and the refinement dilemma (the condition number of the global stiffness matrix deteriorates quickly as the mesh is refined). This study proposes an AI-enhanced, highly efficient mechanical analysis framework based on a substructure method, which exhibits strong generalization to boundary conditions, geometries, and loading types. Methodologically, we employ a DeepONet operator to learn an implicit mapping between the material distribution inside each substructure and the substructure's numerical shape functions, accelerating the traditionally time-consuming generation of shape functions in the workflow to the millisecond scale. To achieve extreme model-order reduction, each substructure retains only a small number of corner-point degrees of freedom (8 in 2D and 24 in 3D), complemented by specially developed oversampling techniques to ensure computational accuracy. Furthermore, by introducing the Partition of Unity (PU) concept and constructing overlapping finite elements, we effectively address potential boundary coordination issues between substructures that may arise under extreme reductions. Numerical experiments show that with roughly 95% reduction in degrees of freedom, the proposed method can maintain analysis errors within 5%; simultaneously, the trained model exhibits strong generalization capability, enabling direct application to different structural shapes/topologies, external loading conditions, and boundary conditions, thereby achieving orders-of-magnitude improvements in finite element analysis efficiency. The proposed algorithmic framework is also applicable to other AI-augmented non-classical finite element constructions.