Computer ratings are not the same as chess.com ratings that are not the same as FIDE ratings that are not the same as USCF ratings ...
As far as I know, the computer ratings are established by C vs C play, and not by letting a statistically significant number of humans play rated games against them.
One other problem with Computer-ELOs is that you rarely know what hardware was used and that hardware's influence on an engines performance cannot be neglected.
I would add that the "computer" moves are "psychologically" very different too, and that can screw up the ratings.
To take an example of bad chess software, there is a program on windows 7 called "chess titans". On the max difficulty level, I usually get mopped in the opening / early middlegame, losing at least a pawn. But then as the endgame comes the computer just plays weak moves, not tactical mistakes, but moves without a plan, and basically I win 80% of the time, when I did not lose too much material in the first phase.
Compared to a human player, I would say it plays the opening / early middlegame as a 2000, and the endgame like a 1500. So it's hard to put an average rating on that.
I suspect serious softwares have the same problem.
Recently I got Rybka 4 and have been playing rated games with it where you can adjust its ELO rating to whatever you want. After playing about 50 games with it, it seems to me that it is significantly harder to beat Rybka at any given ELO level than it is online opponents at that same level. Is this just my imagination? And if not, does this mean that the online ELO system is inflated somehow?