In my experiance, any operation on that many points is likley to run into a wall.
try splitting off a chunk and running Near on that.
You can create a set of Feature Layers based on the $recno, excuse me, the ObjectID and run each in turn.
The datum issue I think, will hinge on how close together the polygons lie.
The difference in lat/longs will likely be, at most, in hundreds of meters.
If your polygons are tens of kilometers apart or more, I doubt a datum mismatch will be significant.