No, that's just a common choice. You can choose a covariance function (kernel) corresponding to some other function space.

http://www.gaussianprocess.org/gpml/chapters/RW4.pdf