We study the evolution of R&D networks in a Cournot model where firms may lower marginal costs due to bilateral R&D collaborations.

Stochastically stable R&D networks exhibit the dominant group architecture, and, contrary to the existing literature, generically unique predictions about the size of the dominant group can be obtained. This size decreases monotonically with respect to the cost of link formation and there exists a lower bound on the size of the dominant group for non-empty networks. Stochastically stable networks are always inefficient and an increase in linking costs has a non-monotone effect onaverage industry profits.