This paper studies the effectiveness of collusion in the DRAM cartel. Like other high technology products, DRAM is characterized by learning-by-doing and multiproduct competition. I hypothesize that collusion is more difficult to sustain on a new generation, where learning is high, than an old generation, where learning is low. A higher learning rate makes defection from a collusive equilibrium more attractive by reducing future cost. Empirical analysis exploits variation between cartelization and competition to estimate the change in firms' output decisions on each generation. Consistent with the hypothesis, cartel participants are estimated to cut output more on the oldest generation than newer generations. Output decisions on the newest generation also show evidence consistent with defection from collusive equilibria. Lastly, the paper presents a theoretical framework to analyze collusive equilibria with learning-by-doing and multiproduct competition. The model motivates various pieces of evidence that competition authorities can compile to guide antitrust investigations in high technology markets.