This contribution presents recent advances enabled by the CERISSE [1] high-order compressible CFD framework for aerospace and energy applications, demonstrating extreme-scale large-eddy simulations (LES) of reactive flows exceeding 10 billion grid points. CERISSE is designed for modern high-performance computing, combining adaptive mesh refinement (AMR) with GPU-native parallelization to efficiently target heterogeneous exascale systems. The simulations showcased here scale to O(10³) GPUs, enabling good resolution of coupled shock–turbulence–combustion phenomena while maintaining strict control of numerical accuracy and stability. CERISSE employs high-order shock-capturing schemes together with state-of-the-art LES closures, including LES-PDF [2] approaches for sub-grid turbulence-chemistry interaction and non-equilibrium wall-models, as well as embedded/immersed boundary methods to represent complex geometries. This allows detailed representation of finite-rate chemistry, sub-grid scale variability, and non-equilibrium effects in regimes where classical models fail. AMR is used dynamically to concentrate resolution on shocks, shear layers, reaction zones, and plume interfaces, reducing the overall computational cost while preserving fidelity in critical regions. Two applications relevant to security and defense are discussed. The first concerns scramjet combustor simulations, where LES captures unsteady shock trains, ignition and extinction dynamics, flame anchoring mechanisms, and strong coupling between turbulence and chemistry under realistic operating conditions, all of which are critical for defense systems operating at Mach >5. The second focuses on supersonic retro-propulsion for entry–descent–landing, a problem of direct relevance to space access and rocket recovery. Here, large-scale LES resolves the interaction between rocket plumes, bow shocks, and turbulent boundary layers, revealing new transient flow regimes, plume-induced instabilities, and mode-switching behavior that are not accessible in lower-fidelity simulations. Beyond the individual applications, the results demonstrate how large-scale LES enabled by AMR and GPU acceleration fundamentally can change our understanding of the dynamics of safety-critical flows.