Abstract Craving is a core feature of substance use disorders and a strong predictor of relapse. It is linked to polysubstance use, overeating, pathological gambling, and other maladaptive behaviors. Despite its utility, craving measures are based on self-report, with associated limitations related to introspective access and variability across sociocultural contexts. Objective biological markers of craving, which could reveal the neurophysiology of craving, are lacking. For example, it remains unclear whether craving for drugs and food involve similar or separable mechanisms. Across three studies (N = 101), we combined fMRI with machine-learning to identify a brain-based marker of cue-evoked drug and food craving, resulting in a cross-validated multi-system pattern (or brain signature), including ventromedial prefrontal and cingulate cortices, ventral striatum, temporal and parietal association areas, mediodorsal thalamus, and cerebellum. This signature predicted self-reported craving intensity (p < 0.002, within-person r = 0.50) and discriminated high from low craving with 78% accuracy. It also discriminated drug users versus non-users based on brain responses to drug (75% accuracy), but not food, cues. Predictive brain models trained on drug cues also transferred to food cues, and vice versa, suggesting shared neurophysiological mechanisms for drug and food craving. In conclusion, fMRI can be used to predict craving across different drug and food cues, and separate drug users from non-users. Future studies are needed to assess whether the brain signature of craving responds to clinical intervention and can predict long-term clinical outcomes.