It can do things (as the author of the article you linked to pointed out) which would be harder or impossible in real world hardware.

Never impossible. Even the software VM eventually goes down to hardware CPU(s). But it is true that the designer of instruction sets for a software VM usually won't bother to ask how easy a given instruction could be implemented in hardware. Often, they may not even have the know-how to figure out how to design a hardware-based architecture (which is fine, since their aim is software).

Some might be interested in an old architecture, the Burroughs B5000 (poke around Google for information), which was supposedly optimized to run higher-level languages. In the recent ACM interview with Alan Kay, he mentioned a benchmark on that machine (made in 1979) and a more modern one. The modern machine ran it 50 times faster. However, Moore's "Law" should have given it roughly 50,000 times speed up. Somewhere along the way, we lost around three orders of magnitude in speed due to bad architectures.