I really have mixed emotions regarding the two approaches of building on the JVM vs. building from scratch.
On the one hand, building on top of the JVM gives quite a head start in many ways, and I think it's a bit ridiculous how poor the interop situation is for native programs despite the availability of FFIs (although some/much of this may be fundamental to language differences).
On the other hand, it seems the least common denominator problem always creeps up when building on a VM plus you have an extra dependency in the chain, etc.
If the Y axis is features and the X axis is time, it seems clear that building on a VM gives you a higher Y intercept. The question is whether the slope is unduly impeded. I suspect it might be, but that's totally subjective on my part.