Erlang is not about parallelism, it's about fault tolerance. It's about "making reliable distributed systems in the presence of software errors" (as Armstrong's thesis is titled).
How will your STM example behave if the computer managing the transactions suddenly has a massive hardware failure?