Intratumor transcriptional heterogeneity (ITTH) presents a major challenge in cancer treatment, particularly due to the limited understanding of the diverse malignant cell populations and their relationship to therapy resistance. While single-cell sequencing has provided valuable insights into tumor composition, its high cost and technical complexity limits its use for large-scale tumor screening. In contrast, several databases collecting bulk RNA sequencing data from multiple samples across various cancer types are available and could be used to profile ITTH. Several deconvolution approaches have been developed to infer cellular composition from such data. However, most of these methods rely on predefined markers or reference datasets, limiting the performance of such methods by the quality of used reference data. On the other hand, unsupervised approaches do not face such limitations, but existing methods have not been specifically adapted to characterize malignant cell states, focusing instead on general cell types. To address these gaps, we introduce CDState, an unsupervised method for inferring malignant cell subpopulations from bulk RNA-seq data. CDState utilizes a Nonnegative Matrix Factorization (NMF) model improved with sum-to-one constraints and a cosine similarity-based optimization to deconvolve bulk gene expression into distinct cell state-specific profiles, and estimate the abundance of each state across tumor samples. We validate CDState using bulkified single-cell RNA-seq data from five cancer types, showing that it outperforms existing unsupervised deconvolution methods in both cell state proportions and gene expression estimation. Applying CDState to 33 cancer types from the TCGA, we identified recurrent malignant cell programs, including epithelial-mesenchymal transition (EMT) and hypoxia as main drivers of tumor transcriptional heterogeneity. We further link the identified malignant states to patient clinical features, revealing states associated with worse patient prognosis. Finally, we find that alterations in genes such as KRAS or TP53, whose copy number or mutation status strongly correlate with specific malignant states, may play a key role in driving these cellular states. Overall, CDState provides a powerful and accessible approach to resolving intratumor heterogeneity using bulk RNA-seq data and emerges as a promising tool for advancing our understanding of malignant cell heterogeneity.