It is a good question.

To work around we can cast as float or bigint; but the nub of the answer is in realising that it is the internal datatype of an undeclared 2 (ie an int) that is in question.

If Sql 2005 had all numbers internally as bigint it would cost memory, if it cast based on a guessed intention it would be slated to high heaven!

So an undeclared 2 is an int and a power operation on an int returns an int. Wishing that a power operation on an int would sometimes return an int and sometimes a bigint (what about a tinyint?) seems odd.

All seems reasonable to me.