To me, the real deciding factor that tells me that LuaJIT will win in the long run is that LuaJIT traces the high-level semantics of Lua, whereas PyPy has to work with Python bytecode, which may have a considerable amount of lost information. There are some good examples in the lambda-the-ultimate thread (for example, optimizing based on the knowledge that HLOAD and ASTORE will never alias).
PyPy can't feasibly optimize the high-level Python AST, because the language is very large compared to Lua, and more or less behaviorally defined by Python's bytecode compiler.
This is one of the interesting things about LuaJIT: it demonstrates the value of a small, well-designed language in a shockingly awesome way. The care that the Lua designers put into making Lua simple and regular is one of the reasons why a single programmer (a demigod, yes, but still — one demigod) is able to make so impressive an implementation.
PyPy can't feasibly optimize the high-level Python AST, because the language is very large compared to Lua, and more or less behaviorally defined by Python's bytecode compiler.