The column of ones incorporates the intercept term. One simple example is the null model where there is no predictor and just a mean y=!+":

We can assume that E"=0 since if this were not so, we could simply absorb the nonzero expectation for the error into the mean ! to get a zero expectation.