Tacx trainers calculate power using a set of curves giving power as a function of resistance setting and roller speed. Most of these trainers allow for some calibration of the power calculation. Some people have done some tests with PowerTap and SRM, and found that in steady speed conditions, the calculation is pretty good (i.e. within 10% of that measured on PowerTap or SRM). See the Tacx forum, there are a few threads on the subject. There is a calibration procedure involving timing the spin-down from a specific speed; I don't know how much that helps accuracy.
Any transient measurements (i.e. accelerating or slowing down) will be wrong, since the calculation assumes steady conditions. So accelerations will be under-estimated, and decelerations over-estimated. The typical example of this is hitting a slope on a Virtual Realilty trainer: the initial reading is very high power, since you are going fast, and the resistance is high, regardless of the fact that your speed is dropping like a rock.
No matter how precise or imprecise it is, it is repeatable (as long as you maintain the same roller pressure, and the same tire pressure), so that you can compare your progress, e.g. measure how much HR you need for a given power output, and see the evolution with your fitness level.