Ran into this today. Code as follows, with tablesample commented out. Run it, get the expected results of 5000050000.

-- tested in SQL Server 2014

use tempdb;

IF OBJECT_ID('dbo.ints', 'U') IS NOT NULL  
    DROP TABLE ints;

CREATE TABLE ints(x bigint NOT NULL);

WITH MakeInts AS
(
    SELECT CAST(1 AS bigint) AS x
    UNION ALL
    SELECT x + 1 AS x
    FROM MakeInts
    WHERE x < 100000
)
INSERT INTO ints 
    SELECT *
    FROM MakeInts
    OPTION (MAXRECURSION 0);

DECLARE crsr CURSOR READ_ONLY
FOR select x 
from ints
--tablesample (5 percent)
;

DECLARE @x bigint;
DECLARE @sum bigint = 0;
OPEN crsr;

FETCH NEXT FROM crsr INTO @x;
WHILE (@@fetch_status <> -1)
BEGIN
    IF (@@fetch_status <> -2)
    BEGIN
        set @sum += @x;
    END
    FETCH NEXT FROM crsr INTO @x;
END

CLOSE crsr;
DEALLOCATE crsr;

print @sum;

Now comment tablesample back in and run again. Same result of 5000050000.

Am I overlooking something completely obvious?

(Martin Smith's comment pointed out that tablesample samples at the page level and my table was small enough to fit on one page. Good catch, and amended code above so it makes a much larger table. Doesn't fix though)