For us its the simplified production environment. We use BEAM's clustering capability to forgo things like redis and/or a message queue. It probably doesn't matter for every company, but it helps for us where we have zero full time ops people.
In a current project I'm also able to use OTP to create a single microservice for my monolith that runs on its own node and the two can communicate with run-of-the-mill message passing without the need for an HTTP API.
Dist messaging and mnesia scale pretty far. Although it must be noted that they were original built around rock solid networking (two nodes in one chasis), so if your networking is flakey, you have to handle that; my recommendation is to make your network not flakey, but that's easier to say than it is to do.
That said, redis is in memory key-value, and Mnesia does in memory key-value too. If you're using Postgres's relational stuff or data bigger than memory, that's a different thing. Mnesia has some relational facilities, but I dunno if anybody uses them much; mnesia can use disc_only_copies, but that's not a good time. Of course, memory only gets you pretty big tables these days; I ran some mnesia nodes with 768GB, but a couple TB isn't too hard to put together these days.
The reality is that asynchronous, parallel, distributed computing is hard, Erlang makes it easier. Not easy, but easier.