The thing is, like 20% of your program ends up being non-idiomatic, ghc-specific Haskell. I suppose for certain applications it could end up being the whole thing, but in general I'm thinking optimized Haskell is in that ballpark tops.
And idiomatic Haskell's performance is plenty acceptable for a variety of common use-cases (so not "nearly always is")
I've deployed plenty of web services & data processing jobs in Haskell and the RTS has never been a bottleneck or issue. So I'm guessing you weren't doing an IO-bound activity (where Haskell is quite good.)
And idiomatic Haskell's performance is plenty acceptable for a variety of common use-cases (so not "nearly always is")