Interesting. The parallels between ML and Scheme, two languages invented in the '70s and in an academic setting, are striking: the core language is great, but the community is divided into ever-so-slightly incompatible implementations, so, in practice, there's very little sharing. Therefore, it's impractical to do most "large" projects in the language, due to lack of critical mass of useful libraries. ML/Scheme end up being learning languages, for the most part, and then everyone moves on to {OCaml, Haskell}/Common Lisp to get real work done.

I would say these similar outcomes are a repudiation of the idea that multiple implementations of the same language is a good thing, but that hasn't hurt C's popularity. More likely, it has something to do with their roots in programming language research. Maybe ML and Scheme are too good for their own good, and other academics can't resist using them as a basis for their own pet ideas. Perhaps a truly successful (read, "widely-adopted") programming language has to be flawed just enough to discourage anyone from trying to improve on it. Worse is better strikes again?