That's assuming there isn't demand for more powerful models, there's still plenty of room for improvement from the current generation. We didn't stop at GPT-3 level models when that was achieved.