Ehhh... well... if you're willing to put some work into the matter that may not be as true as you'd think. Yes, to speak the federated protocol to another server you're going to be speaking the XML protocol, but between a server and a client that come to accommodations with each other, you don't really have to speak XML.

You'd want to speak something isomorphic to the XML, but the XMPP XML is not that verbose on a node-by-node basis or anything (minus a couple things, see below), it really is the XML, not the semantic content. It would be fairly easy for a medium-to-high skill developer to come up with an isomorphism for use with mobile clients. (It is sufficiently tricky that there would be a few traps that could catch the inexperienced or unwary (watch those namespaces!), but it isn't that difficult.) The end result is a more bandwidth efficient protocol that is a drop-in replacement at the XMPP library layer in the client for XMPP. (Of course, if your client is programmed to fling around XML all over the place internally this won't help, but, well, that's the price of failing to abstract properly.)

Where you'd still have trouble is things like roster size, but yes, there are XEPs for that, and before long you're looking at a reasonably svelte protocol (not perfectly, but reasonably svelte) that is still largely XMPP and easy to write communication with the outside world.

What's really difficult is semantic IM translation between entirely different systems, not protocol translation. ejabberd, the server whose internals I am most familiar with, essentially works by internally speaking Erlang terms with itself, and translating to the XML protocol only at the edges. It wouldn't be very hard to drop in different serialization at the edges and get a different protocol; as proof of concept of the fact this is possible, one could easily simply serialize those Erlang terms with Erlang's term serialization, and indeed, when operating ejabberd in a clustered mode, that's actually happening across the wire as the cluster nodes communicate with each other.

However much effort that may be, it seems easier than starting from scratch.