sequence of interpolation of spatial attributes

Say I have two related observations that can be combined to derive a third:

c = a + b

and I want interpolations of all three.

I could do something equivalent to

(form 1) c = krige(a) + krige(b)

or

(form 2) c = krige(a + b)

The issue become important when a is a subset of a very large set.  In 
my case, a is a set of observations at one of five depths on one of 385 
dates.  b is static throughout.

For each date, I want a and c, so I have to perform krige(a) in any 
case.  I want krige(b) so that has to be performed once.

After the static krige(b) is complete and a krige(a) has been performed, 
the compute time of form 1 is less than a second.  If I used form 2, the 
compute time of each depth and date doubles.  If I want other derived 
attributes, I add another computationally expensive step.  In real 
terms, I have four derived attributes.  Form 2 will take 240 hours 
versus 60 cpu hours for form 1.

So my question is, are there reasons one form might be better than 
another?  Intuitively, form 2 is better, but if the original 
interpolations are good, I can't think of any reason the second form 
would be invalid.  I can overlay (randomly selected) observations onto 
interpolations and qualitatively compare the two.  Sometimes 1 seems 
better and sometimes 2 seems better.  I am willing to randomly compare 
the two methods if that makes sense and if there is a quantitative 
comparison I could make.

Anyone have any suggestions?

Thanks!