It forces the variance of $Y$ to be constant for all $\mu$.

It makes the link function linear in the parameters, which simplifies the interpretation of coefficients.

It ensures that the sufficient statistic for $\theta$ is equal to $Y$ in the exponential family representation.

It allows the model to ignore the distribution of the errors entirely.