Recent work focuses on how homophily (tastes for similarity) can lead to network segregation (the lack of intergroup friendships). Yet studies seldom consider whether and how levels of network segregation could also lead to the trend towards homophily over time. Instead, existing cross-sectional studies argue that intergroup exposure exacerbates homophily. By neglecting longitudinal data on changes to friendships and focusing on intergroup exposure rather than initial intergroup friendships, existing studies are likely to present an overly pessimistic view on the benefits of intergroup contact. Using longitudinal data and stochastic actor-oriented models, I study how levels of initial ethnic network segregation between students in Swedish Classrooms with "native" backgrounds and immigrant-origin students are related to subsequent levels of ethnic homophily. Results show that more initial network segregation in classroom friendship networks is linked to more ethnic homophily in network evolution, which suggests that beyond mere exposure, optimal conditions for contact and actual intergroup friendships are crucial for positive intergroup dynamics, and that their benefits can appear longitudinally.