Hi Marc,
I regularly do stacks of hundreds of subs. Now that I'm doing short subs with CMOS cameras I'm sure the need to deal with thousands is going to arise before too much longer
A single large stack will always give the best result. It has the most information possible for the normalisation and rejection algorithms to chew on. It also offers the most consistent weighting of subs.
Of course, it may not be practical to do a single large stack and in my (limited) experience a stack of stacks (no rejection the second time) will also work well. The only caution I'd offer is that making individual stacks of data from a single night is not a good idea. Batching multiple nights will help to reduce the impact of differing sky conditions.
Cheers,
Rick.