That's becoming less true. The cost of inference has been rising with bigger models, and even more so with "reasoning models".

Regardless, at the scale of 100M users, big one-off costs start looking small.