As GPU-driven supercomputing architectures become standard for HPC and AI, offering superior performance and energy efficiency, the CFD community must adapt to fully exploit these systems for high-fidelity simulations of complex industrial flows. Porting legacy codes is feasible but often inefficient, requiring compromises to adapt software optimized for homogeneous CPUs. Developing a new code from scratch, however, allows hybrid computing concepts to be integrated from the outset, yielding more efficient GPU-based solvers. Following this approach, we present SOD2D, a next-generation CFD solver purpose-built for GPU computing and high-accuracy aerospace simulations. Written in Fortran with OpenACC, SOD2D supports LES and DNS of compressible and incompressible flows, offloading all heavy computation to accelerators. It uses Continuous Galerkin Spectral Element Methods (CG-SEM) for efficient GPU computation of convective, diffusive, and iterative solver operations. Single-precision runs show 4-8× speedups over CPU nodes for large Taylor-Green vortex cases, with strong scalability as mesh and memory grow. Parallelization employs CUDA-aware MPI and NCCL, making SOD2D exascale-ready. Within the EU EXCELLERAT Centre of Excellence, SOD2D has simulated realistic aircraft aerodynamics with billions of degrees of freedom on up to 2048 GPUs on MareNostrum-5, achieving 20x faster time-to-solution and 10-12x lower energy use than CPU versions. These advances enable more accurate lift, drag, and noise predictions. SOD2D is open source under the MIT license and available on GitLab.