I think the GPs point is that, starting from 0, both require effort AND are reusable.

The "debate" is usually between people who have committed to either path (creating tools or learning tools) and think that just because the marginal cost of my approach, for me is 0, it must be 0 for everyone else. Which is patently false.

I feel we can safely assume that tuning the JVM GC might cause behavior that’s pathologically slow but never actually incorrect, which is inherently safer than rewriting a bunch of production code (probably more than once) hoping for a similar result.