A factor of 10x slowdown would probably be reasonable. Is the current state of the art on ARM able to get it below 10x?
And, BTW, nobody would buy a computer that can run only Windows and Office.
To be fair, only the Office part would have to run under emulation. As soon as the code calls into core Windows, the machine would be running ARM code. I don't know where the boundary is with Office.
I do agree that only running Windows and Office is not sufficient. But its a starting point for the discussion.
Also, consider that A-15 may be competitive with current x86 designs. Who knows what will Intel and AMD bring by the time A-15/Denver machines hits the shelves.
That said, I find my modest Atom good enough for Django programming the slightly less modest Core i5 very reasonable with Java and Eclipse. It will take some iterations until ARM catches up (I assume it will move faster than x86 because of its simple architecture). I have no doubt a relatively modest ARM machine will be good enough for most of my then current usage.
Maybe we just have really fast networks.
I assumed he meant something where the network latency would impose a keystroke-level degredation in performance.
The picture may not be that grim if at least timing-critical parts of Office can be made into fat binaries. I would not bet on straight emulation of uncooperative code, however.