The rate limits are not because of compute performance or the lack of. It's to stop people from training their own models on the very cutting edge.