It's not widely appreciated how efficient OSM is: the whole project runs on roughly $170k a year. Wikimedia spends around $65m (there are, obviously, many differences in what the organisations do, but it's an interesting comparison).
It's astonishing that a project like this runs on a budget similar to what it costs to get one (!) software engineer in the Valley. Well done to everyone involved!
That's very impressive. I'd have thought bandwidth costs alone would be high, but thinking about it a bit more I suspect map tiles cache incredibly well even without any special tricks, since most tile requests will be from locals hitting the same few ISP caches.
Except, naturally, that openstreetmap.org is mainly aimed at OSM contributors, which typically are actually changing the data and by that invalidating the tiles created for our standard style (which update in minutes).
The flipside to this is how limited OSM is. There are many more things it could be doing, IMHO, but the organization has made the choice to stay small and lean. It does what it is doing quite well though and the folks who run it are impressively efficient.
Rather than the OSMF doing everything, and having one official geocoder, or one official routing engine, instead there is an ecosystem of software and 3rd party companies around OSM who provide that