In global virtual teams (GVTs) individuals from different cultures collaborate for a specific work purpose. Cultural differences have been studied intensively as a predictor of team performance, yet the findings have been inconsistent, sometimes showing positive, negative, or non-significant effects. These inconsistent findings point towards previously overlooked mediating and moderating effects, particularly in GVT environments. Drawing on social categorization theory, this study develops a model examining the negative influence of perceived cultural differences within GVTs on team performance via the mediating effect of team dynamics (i.e., collective efficacy). As a moderating effect, we argue that a higher level of shared working language weakens the negative effect of cultural differences on collective efficacy by reducing subgroup formation. To test our hypotheses, we analyzed a multi-wave and multi-source dataset of 3,585 individuals nested in 800 GVTs. Our results largely support our hypotheses, showing that perceived cultural differences unfold their negative influence on team performance via team dynamics. The negative relationship is weaker when team members have a higher level of shared working language. We discuss theoretical and practical implications.