I study the effectiveness of collusion during an international cartel in DRAM, a type of semiconductor characterized by learning‐by‐doing and multiproduct competition. First, by exploiting time and firm‐level variation in cartel activity, I estimate that cartel participants cut output more on the oldest product generation than on newer generations. This is consistent with a hypothesis that higher learning rates in newer generations make defection from collusive equilibria more attractive. Second, I formulate a test for defection from collusive equilibria in learning‐by‐doing industries. Third, I formalize these results in a theoretical framework and discuss implications for antitrust policy in high technology markets.