In the treatment of advanced non-small cell lung cancer (NSCLC), immune checkpoint inhibitors have shown remarkable results. However, not all patients with NSCLC respond to this drug treatment or receive durable benefits. Thus, patient stratification and selection, as well as the identification of predictive biomarkers, represent pivotal aspects to address. In this framework, metabolomics can be used to support the discrimination between responders and non-responders. Here, metabolomics was used to analyze the sera samples from 50 patients with NSCL treated with immune checkpoint inhibitors. All the samples were collected before the beginning of the treatment and were analyzed by NMR spectroscopy and multivariate statistical analyses. Significantly, we show that the metabolomic fingerprint of serum acts as a predictive “collective” biomarker to immune checkpoint inhibitors response, being able to predict individual therapy outcome with > 80% accuracy. Metabolomics represents a potential strategy for the real-time selection and monitoring of patients treated with immunotherapy. The prospective identification of responders and non-responders could improve NSCLC treatment and patient stratification, thus avoiding ineffective therapeutic strategies.