Cutting AWS costs through inference infrastructure improvements