Oh, and one other thing.
Might I remind you of the procedure of going through scientific experiements? The idea is to vary just one variable in order to test that single variable's influence on the outcome.
When you compare the GeForce4's anisotropic to the Radeon 9700's, you are changing far more than one variable, and therefore it is impossible to accurately tell which of the numerous changes is the one that is the cause of the performance discrepancy.
In other words, just because the GeForce4 has more of a performance hit than the Radeon 9700 with anisotropic enabled doesn't mean that you can't have an anisotropic implementation that is invariant upon rotation about the z-axis with a performance hit similar to the Radeon 9700's. You don't have enough information to make that statement, and looking at the technical aspects of the situation contradict that statement.