You can write TF in Java. Most don't, for a good reason. Python is the first class citizen, everyone uses it, you can find help most easily for it, its data science ecosystem dwarves Java's, you'll have a hard time convincing scientists to port their Python notebooks into Java, etc.
As for web frontends, we've also already had GWT. In most cases you're better off going with the grain and using JS/TS, for obvious reasons.
I can continue for a while with examples where Java is inappropriate. Another one would be system daemons that need to use minimal memory. There is Graal but it has so many caveats that I can't really take it seriously yet (and I've tried, multiple times). In this same vein, serverless applications are generally more painless with languages whose ecosystems emphasize fast startup, which Java historically has been terrible at. And on top of that you'll pay for the memory the JVM hogs.
I’m pretty sure it’s temporary. There are plenty of projects working in that direction, e.g. https://github.com/i-net-software/JWebAssembly
> you wouldn't write your Tensorflow code in Java
Why?
Kernel modules and device drivers are probably the only example where you need to pick another tool.