You're ignoring that each foundation model requires sinking enormous and finite resources (compute, time, data) into training.

At some point, even companies like Meta need to make a limited number of bets, and in cases like that it's better to have smarter than more people.