It is straightforward to convert Python bytecode into C, however this yields little speedup. It still has to do all the same reference counting and hash-table lookups.
Replace the garbage collector with a data compressor that looks for common patterns (eg key-value stores with the same keys / types / values) and replaces them with a new more compact type. It will also look for multiple copies of the same value and convert them into a single copy. Of course it will still also throw out garbage.
Language will look and feel like a modern scripting language. However, underneath it will be a functional, lisp-like language. Perhaps point-free style to facilitate compression. Simulate all the usual mod-cons with a stack of monads -- which you won't need to know about, unless you want to do something funky.
The code will be data too, hence compressible too.