That's not TCO, that's just an optimization for individual recursion. You might be able to use it for multiple forms under the same recur (like w/ "labels"), but it's very different when it applies to everything.

With TCO, you can use a set of mutually-recursive functions - you can set up a FSM via tail calls, use CPS for backtracking (and many other things), a re-entrant virtual machine where each opcode is a tail call (rather than a giant switch/case or lots of gotos), etc. Being able to use function calls like "gotos with arguments" (but with local scopes, etc.) can make some constructs much easier to read, reason about, etc.

I'm not that familiar with the implementation of the JVM, but AFAICT it's due to difficulties setting up tail calls efficiently in JVM bytecode.