OS-level threads are a crude big hammer way of doing parallelism - want to run a task? Just switch out the whole 8kb of stack, then switch back to a stack that's 99% identical. Cooperative multitasking was always known to be much more efficient if you could ever get it to work safely - but one misbehaving task killed everything

Ironic that with deployment we're moving in the opposite direction - we've decided we really do need the complete isolation of separate VMs/containers and are willing to pay the performance overhead because process separation isn't enough.