A variety of models explain why new technologies such as hybrid corn, agricultural fertilizer, and new medical treatments diffuse so slowly. A common characteristic of these models is that agents (or countries) quickest to diffuse do so optimally because of relative gains from the new technology; the debate is around why others are so slow. In this paper, we develop a nested Bayesian model of diffusion and learning with heterogeneous agents that allows for overconfidence, which can cause early innovators to exhibit below-average productivity. We apply the model to the case of implantable cardiac defibrillators (ICDs), a medical device approved in 2005 to help prevent cardiac death in patients with weakened hearts (congestive heart failure). Using a unique clinical registry of every ICD implanted during 2006-13 linked to Medicare claims data, we find remarkable variations in the speed of diffusion across hospitals and regions. The structural model matches both aggregate moments, and individual hospital-level trajectories, of mortality and utilization. We find that overconfidence raises mortality by 8% on average (and more among those most overconfident), and can explain roughly three-quarters of variation in diffusion rates and risk-adjusted mortality. In addition, the model predicts, correctly, that the most overconfident hospitals are the ones that scale back quickest. These results suggest caution in equating rapid diffusion to productivity gains, particularly in health care.