Highly tumorigenic cancer cell (HTC) populations have been identified for a variety of solid tumors and assigned stem cell properties. Strategies for identifying HTCs in solid tumors have been primarily empirical rather than rational, particularly in epithelial tumors, which are responsible for 80% of cancer deaths. We report evidence for a spatially restricted bladder epithelial (urothelial) differentiation program in primary urothelial cancers (UCs) and in UC xenografts. We identified a highly tumorigenic UC cell compartment that resembles benign urothelial stem cells (basal cells), co-expresses the 67-kDa laminin receptor and the basal cell-specific cytokeratin CK17, and lacks the carcinoembryonic antigen family member CEACAM6 (CD66c). This multipotent compartment resides at the tumor-stroma interface, is easily identified on histologic sections, and possesses most, if not all, of the engraftable tumor-forming ability in the parental xenograft. We analyzed differential expression of genes and pathways in basal-like cells versus more differentiated cells. Among these, we found significant enrichment of pathways comprising "hallmarks" of cancer, and pharmacologically targetable signaling pathways, including Janus kinase-signal transducer and activator of transcription, Notch, focal adhesion, mammalian target of rapamycin, epidermal growth factor receptor (erythroblastic leukemia viral oncogene homolog [ErbB]), and wingless-type MMTV integration site family (Wnt). The basal/HTC gene expression signature was essentially invisible within the context of nontumorigenic cell gene expression and overlapped significantly with genes driving progression and death in primary human UC. The spatially restricted epithelial differentiation program described here represents a conceptual advance in understanding cellular heterogeneity of carcinomas and identifies basal-like HTCs as attractive targets for cancer therapy.