That's nothing like why VRML failed
VRML was a declarative scene description. So it failed for the same reasons that all of the scene graphs around then — Direct3d retained mode, RAVE, etc — also failed. Developers don't want that. It makes it incredibly difficult to produce an efficient rendering because you can't apply domain-specific algorithms. So you can't do Doom because there's no way to provide a BSP tree or write special-case linear-along-spans polygon fillers. You can't do Descent because there's no way to walk a portal tree. You can't do most racing games because you can't reduce visibility and clipping of extra-track detail to a one-dimensional problem. Etc, etc.
Scene graphs are in better shape now just as you wouldn't think twice about using a compiler: they're better at management (thanks in part to dynamic tools like the occlusion query reducing some of the need for ahead-of-time calculation) and, anyway, the processing budget is so huge that optimising for project creation time is smarter. This was definitely not the case in 1994.
Arguing that VRMLs issues were overwhelmingly on the business side is disingenuous. Think of almost any significant 3d program from the mid-'90s. VRML is a bad fit for the program you're thinking of.