Deep learning has emerged as a powerful paradigm for surrogate modeling of partial differential equations (PDEs), giving rise to a new class of methods often referred to as neural PDE solvers. Despite their promise, scaling these models to industrial simulations with highly diverse geometries and billion-cell meshes remains a fundamental challenge. In this talk, I will first present the Transolver family—Transolver, Transolver++, and Transolver-3—which together form a scalable backbone for high-fidelity physics simulation on complex industrial geometries. Transolver tackles the geometric irregularity of industrial meshes by projecting them into intrinsic physical state representations, enabling more effective and generalizable learning for engineering design problems. Along with the efficiency optimization in Transolver++ and Transolver-3, Transolver makes neural PDE solving feasible on industry-scale geometries with up to billions of cells. By scaling this backbone, we develop GeoPT, a unified pre-training model for general physics simulation. GeoPT introduces a lifted-space pre-training paradigm that breaks the data bottleneck and can generate millions of self-supervised samples within days, enabling the million-sample pre-training regime in the physics domain. Together, these works outline a possible path toward scalable, learning-based PDE solvers for real-world industrial applications.