I imagine it's because until about 5 years ago, there weren't many choices that were significantly safer than Java (no null references, failing to account for all cases in a match statement is a compile error, etc.) and generally within a factor of 2 of C's speed. Haskell's laziness by default can mean less predictable performance, or at least fatter tails on the performance distribution.
It's a reasonable language choice, and the OCaml's type safety was obviously an important factor. However, the key reason why it was such a brilliant choice was for attracting top talent. People who show the initiative to learn a functional language tends be very self-motivated, intelligent, and passionate about their craft. As a result, Jane Street has long been known as one of the (if not the) top places to go if you want to do functional programming full-time.
I believe for many of the same reasons that Lisp was used with great success in the past. OCaml is descended from the ML family of languages and grounded in solid mathematics, like Haskell. In the hands of the right person it's a formidable tool and arguably provides barriers to entry for competitors.
