Assuming you have an ordering column -- say id -- then you can do the following in SQL Server 2012:

select col,
       col - coalesce(lag(col) over (order by id), 0) as diff
from t;

In earlier versions of SQL Server, you can do almost the same thing using a correlated subquery:

select col,
       col - isnull((select top 1 col
                     from t t2
                     where t2.id < t.id
                     order by id desc
                    ), 0)
from t

This uses isnull() instead of coalesce() because of a "bug" in SQL Server that evaluates the first argument twice when using coalesce().
You can also do this with row_number():

with cte as (
      select col, row_number() over (order by id) as seqnum
      from t
     )
select t.col, t.col - coalesce(tprev.col, 0) as diff
from t left outer join
     t tprev
     on t.seqnum = tprev.seqnum + 1;

All of these assume that you have some column for specifying the ordering. It might be an id, or a creation date or something else. SQL tables are inherently unordered, so there is no such thing as a "previous row" without a column specifying the ordering