Yes, but... in order to train your next SotA model you have to do this anyway and do rejection sampling to generate good synthetic data.

So if you can do it in prod for users paying 300$/month, it's a pretty good deal.