“Significant improvements in compute efficiency could be achieved by applying more labor to optimize existing systems, such as making kernels more efficient and improving routing between models.”