What commenter Tadas says about deployment headaches is largely true in my experience as well.
That said, one thing to consider might be finding some way to leverage PHP as a deployment vehicle. Surely it should be possible to code up some sort of VM in PHP that executes three-address code, or to write a compiler codegen module that outputs valid PHP from its internal TAC representation.
So in the ideal world, you come up with this amazing new platform, call it JeffScript or whatever. You provide a great native tool chain and reasonable integration with the big players of web serving (Apache and IIS I guess) and then, as a "side project," you provide an alternate deployment method that compiles to PHP (having coded up API bridges to the 25-30 most popular, but not part of the base package, PHP modules.) This approach has seen some success in things like Jython, Erjang, JRuby, etc. (Hell, there's apparently even a PHP -> Java bytecode compiler -- who knew!?)
The problem I see here is that part of the attraction of PHP is that for beginners/prototypers/contract-developers, there are mountains of bad, but free and quasi-usable code out there just waiting to be (ab)used. Full CMSes! CRM applications! BBoard systems! Frankenstien-ing up hybrid web apps seems to be a major passtime of PHP contract developers, and it's a feature that I struggle to see how JeffScript-on-PHP might leverage.
Just a thought.