We've been building this way ever since we moved to Elixir which makes distributing state across your application easy. Memory on application servers is severely underutilized
That's highly dependent on your choice of language and how you're running your app.
While we don't have the scale yet to need to cluster out and handle these kinds of issues, I've told my cofounder, that thanks to Elixir, I'm not terribly worried about the costs when does happen (not as worried as I would be if we were using something like ruby)
I think what makes actor models so nice is the explicit ownership of state. It is not possible to declare "var x = 1" in one file, and access x in another file. You always have to retrieve state explicitly, otherwise it won't be accessible within the scope of your function.
The flow is to dual write to our rdms and a local process. The process either stores in ETS or an process state map. It also broadcasts out the event to the cluster. When a node in the cluster gets a message, it executes on it. Process startup is restored from the DB which is still authority.
Generally this is really easy to do in Elixir. Maybe 50 lines of code for the system described (minus DB code), much of it definition lines.
These processes can seamlessly interoperate while running on multiple machines.
This makes building stateful application servers a pretty natural fit.