Rolling contact fatigue of rolling element bearings is a statistical phenomenon that is strongly affected by the heterogeneous nature of the material microstructure. Heterogeneity in the microstructure is accompanied by randomly distributed weak points in the material that lead to scatter in the fatigue lives of an otherwise identical lot of rolling element bearings. Many life models for rolling contact fatigue are empirical and rely upon correlation with fatigue test data to characterize the dispersion of fatigue lives. Recently developed computational models of rolling contact fatigue bypass this requirement by explicitly considering the microstructure as a source of the variability. This work utilizes a similar approach but extends the analysis into a 3D framework. The bearing steel microstructure is modeled as randomly generated Voronoi tessellations wherein each cell represents a material grain and the boundaries between them constitute the weak planes in the material. Fatigue cracks initiate on the weak planes where oscillating shear stresses are the strongest. Finite element analysis is performed to determine the magnitude of the critical shear stress range and the depth where it occurs. These quantities exhibit random variation due to the microstructure topology which in turn results in scatter in the predicted fatigue lives. The model is used to assess the influence of (1) topological randomness in the microstructure, (2) heterogeneity in the distribution of material properties, and (3) the presence of inherent material flaws on relative fatigue lives. Neither topological randomness nor heterogeneous material properties alone account for the dispersion seen in actual bearing fatigue tests. However, a combination of both or the consideration of material flaws brings the model’s predictions within empirically observed bounds. Examination of the critical shear stress ranges with respect to the grain boundaries where they occur reveals the orientation of weak planes most prone to failure in a three-dimensional sense that was not possible with previous models.