On the existence of universal models