Esophageal squamous cell carcinoma (ESCC) is a worldwide malignancy with high mortality rates and poor prognosis due to the lack of effective biomarkers for early detection. Exosomes have been extensively explored as attractive biomarkers for cancer diagnosis and treatment. However, little is known about exosome metabolomics and their roles in ESCC. Here, we performed a targeted metabolomic analysis of plasma exosomes and identified 196 metabolites, mainly including lipid fatty acids, benzene, amino acids, organic acids, carbohydrates and fatty acyls. We systematically compared metabolome patterns of exosomes machine learning from patients with recrudescence and patients without recrudescence and demonstrated a marker set consisting of 3'-UMP, palmitoleic acid, palmitaldehyde, and isobutyl decanoate for predicting ESCC recurrence with an AUC of 98%. These metabolome signatures of exosomes retained a high absolute fold change value at all ESCC stages and were very likely associated with cancer metabolism, which could be potentially applied as novel biomarkers for diagnosis and prognosis of ESCC.