This is because any 2x improvement in model capability requires 10x the compute