This is one thought that I have >

Assuming there is a realiable PK field in the table:

Bring in the 70 million rows to a staging area locally on the destination server, index that by the PK field, finally perform a 'Merge' operation to the destination table by way of that PK field. If there are really not that many additions and updates ( I am ignoring deletes, but maybe that factors in too) the statistics you have are likely still usable.

You could do all of this in SSIS. Of course I would call a stored Proc or have the merge statement directly in a t-sql task. I dont promise this will be fast, but it might be done in the desired time frame and eliminate the need to re-build indexes each time.