It's not just about context switching costs. One advantage to non-preemtable concurrency is that data races are much easier to reason about, because any operation that you know won't block becomes atomic by default (this doesn't apply to all models, eg. ones that have m:n scheduling).

There's also other resource overheads. For example, I would think nothing of spinning up 10000 greenlets/green threads in python. They cost maybe 1KB of memory each. But spinning up 10000 OS threads? That's a little more dicey.