So it would be faster and more efficient when sticking to the new subset and Nx slower then using the emulation path.

So this is kind of a useless question, because in such a timespan anything can happen. 20 years ago computers had somewhere around 512MB of RAM and a single core and had a CRT on desk.

Obliterating x86 in that time would take quite a lot more than what the ARM trajectory is now. It's had 40 years to try by now and the technical advantage window (power efficieny advantage) has closed.

Now apply that to weapons systems in conflict against an enemy that DOES have modern production that you (no longer) have... it's a recipe for disaster/enslavement/death.

China, though largely hamstrung, is already well ahead of your hypothetical 2005 tech breakpoint.

Beyond all this, it's not even a matter of just slower, it's a matter of even practical... You couldn't viably create a lot of websites that actually exist on 2005 era technology. The performance memory overhead just weren't there yet. Not that a lot of things weren't possible... I remember Windows 2000 pretty fondly, and you could do a LOT if you had 4-8x what most people were buying in RAM.