Acute myeloid leukemia (AML) is an aggressive disease with complex and heterogeneous biology. Although several genomic classifications have been proposed, there is a growing interest in going beyond genomics to stratify AML. In this study, we profile the sphingolipid family of bioactive molecules in 213 primary AML samples and 30 common human AML cell lines. Using an integrative approach, we identify two distinct sphingolipid subtypes in AML characterized by a reciprocal abundance of hexosylceramide (Hex) and sphingomyelin (SM) species. The two Hex-SM clusters organize diverse samples more robustly than known AML driver mutations and are coupled to latent transcriptional states. Using transcriptomic data, we develop a machine-learning classifier to infer the Hex-SM status of AML cases in TCGA and BeatAML clinical repositories. The analyses show that the sphingolipid subtype with deficient Hex and abundant SM is enriched for leukemic stemness transcriptional programs and comprises an unappreciated high-risk subgroup with poor clinical outcomes. Our sphingolipid-focused examination of AML identifies patients least likely to benefit from standard of care and raises the possibility that sphingolipidomic interventions could switch the subtype of AML patients who otherwise lack targetable alternatives.