...is vastly exaggerated. The performance overhead of FP languages currently typically vastly exceeds the available parallelism in general purpose hardware, and for special purpose hardware alternative languages are available.
a process algebra supporting compositionality, interleaving, and non-determinisim like Hoare's CSP handles the sort of proof required to ensure your concurrency model (if not your actual code) is correct. Oxford University (under Bill Roscoe) has produced tools like FDR2 (FDR3 coming out shortly) which provide tooling to help coders verify their concurrency models via trace/failures refinements.
Or pesky things like I/O. Or efficient data structures.