Does this make more sense than targeting LLVM? It doesn't seem easier to maintain, and you won't get the advantages of the optimizations/work already done.
[1]: http://www.python.org/dev/peps/pep-3146/#performance-retrosp...
I'm actually proponent of JIT, over static compilation, but only for my little pet projects.
I don't know much about WEB, but for games development, static compilation is sometimes the only way to go (or script interpretation). Sometimes there is just not enough memory, and sometimes it's restricted (hell, it so much resticted that you can't change the VTABLE at runtime in C++ class even if you want to)