In theory, it's also possible that you could get any 1,000 rows that are already present in buffers and thus don't require physical I/O. Without an ORDER BY, any 1,000 rows are valid.

Also, since SQL now "shares" I/O among multiple query requests if possible, so you could even get non-sequential, i.e. random, rows that just happened to be in the buffers at that moment. Other dbms's have done this for quite some time, SQL Server just started doing it.