Virtual machines don't add much overhead.
But, more generally, the problem is that tech stacks are an awful amalgamation of uncountable pieces which, taken by themselves, "don't add much overhead". But when you add them all together, you end up with a terribly laggy affair; see the other commenter's description of a web page with 30 dishes taking forever to load.
I think this is the key.
People fall in love with simple systems which solves a problem. Being totally in honey moon they want to use those systems for additional use cases. So they build stuff around it, atop of it, abstract the original system to allow growth for more complexity.
Then the system becomes too complex, so people might come up with a new idea. So they create a new simple system just to fix a part of the whole stack and allow migration from the old one. The party starts all over again.
But it's like hard drive fragmentation - at some point there are so many layers that it's basically impossible to recombine layers again. Everybody fears complexity already built.
I don't dispute that very bad javascript can cause problems, but I don't think it's the virtualization layers or the specific language that are responsible for more than a sliver of that in the vast majority of use cases.
And the pile of build and distribution tools shouldn't hurt the user at all.
If you have a dozen layers and they each add 5% slowdown, okay that makes a CPU half as fast. That amount of slowdown is nothing for UI responsiveness. A modern core clocked at 300MHz would blow the pants off the 600MHz core that's responding instantly in the video, and then it would clock 10x higher when you turn off the artificial limiter. Those slight slowdowns of layered abstractions are not the real problem.
(Edit: And note that's not for a dozen layers total but for a dozen additional layers on top of what the NT code had.)
These are things we're unlikely to get back (and by themselves already make typical PC stacks unable to deliver smooth hand-writing experience - Microsoft Research had a good demo some time ago, showing that you need to get to single-digit milliseconds on the round-trip between touch event and display update, for it to feel like manipulating a physical thing vs. having it attached to your finger on a rubber band). Win2K on a hardware from that era is going to remain snappier than modern computers for that reason alone. But that only underscores the need to make the userspace software part leaner.
--
[0] - Source: I'll need to find that blog that's regularly on HN, whose author did measurements on this.