Automotive surface pressure prediction studies the problem of accurately predicting aerodynamic flow patterns around vehicle geometries to accelerate design optimization. While traditional computational fluid dynamics methods achieve high fidelity, they require prohibitive computational resources that limit rapid design iterations. Recent neural operator approaches have shown promise in learning Partial Differential Equations (PDEs) solution mappings, yet existing methods struggle with multi-scale geometric features and memory bottlenecks when processing large-scale automotive meshes. In this paper, we propose PyramidGTO, a memory-efficient graph-transformer operator that effectively captures multi-scale aerodynamic phenomena through hierarchical processing. We first introduce a three-tier sampling strategy with learnable positional encoding to capture geometric features from global vehicle contours to fine surface details. Then, we develop Streaming Projection Attention that achieves linear memory complexity through block-wise processing and kernel fusion, enabling efficient computation on million-node meshes. Additionally, we design an Adaptive Geometry Network with progressive reconstruction that employs geometry-aware branch selection for specialized processing across diverse surface characteristics. Extensive experiments on Ahmed-Body and DrivAerNet benchmarks demonstrate that PyramidGTO achieves superior performance with 4.5 and 3.7% improvements in surface pressure prediction while maintaining the lowest computational cost among competitive methods.