It's true that that is no longer the attitude. But if we want to understand why one language is more popular than another, we usually need to look at things that happened in the past, not just things that are happening right now. Even C's meteoric rise took 15 years from the first C compilers until it was the undisputed queen of programming languages, partly as a result of missteps made by the communities of ALGOL-68, ALGOL-W, Lisp, PL/I, SNOBOL, TRAC, FORTRAN, BCPL, and, perhaps most interestingly, MULTICS, 20 years earlier.
(It's also possible that my memory that the OCaml team treated the native-code compiler as a second-class citizen is incorrect, either because I misremember what they were saying or because I misinterpreted it.)
Figuring out why OCaml didn't get popular in the past is not a super useful exercise. It's more useful to understand what is needed to popularize it today.
Projects like KDE, the GIMP, GNOME, Lucene, Jython, LLVM, Asterisk, Audacity, CMake, and Danger that got started in C, C++, or Java in the 01997–02002 period are still based on those languages today. OCaml would have been a reasonable language for all of these, though a new interpreter for Python in OCaml would have lacked Jython's key feature, the ability to easily script other things written in Java.
(Danger? Well, Danger as such ended in 02011, but Andy Rubin founded Danger in 01999 and left to found Android in 02003, which ended up making a Java-based smartphone much like Danger's Java-based smartphone, but less locked down.)
The fact that Java focused heavily on performance starting in 01997, when Sun bought Animorphic, meant that a lot of things were able to be written in Java in the 02000s, things that were previously just unthinkable. Lucene predates HotSpot by a bit, but the whole Hadoop ecosystem that grew out of it only makes sense in a HotSpot world. Minecraft obviously pushes performance heavily. SPARK is written in Scala, but OCaml would have been a good fit too—if Hadoop had gone that direction.