Even so, your numbers are still a tiny fraction of GPU units compared to concurrent users, and the limit you "imagine" is just that, imagined.

And you do need to remember that the majority of the compute at these companies is used for model training and not used for inference.