Neuroimaging studies point to neurostructural abnormalities in youth with anxiety disorders. Yet, findings are based on small-scale studies, often with small effect sizes, and have limited generalizability and clinical relevance. These issues have prompted a paradigm shift in the field towards highly powered (i.e., big data) individual-level inferences, which are data-driven, transdiagnostic, and neurobiologically informed. Here, we built and validated neurostructural machine learning (ML) models for individual-level inferences based on the largest-ever multi-site neuroimaging sample of youth with anxiety disorders (age: 10-25 years, N=3,343 individuals from 32 global sites), as compiled by three ENIGMA Anxiety Working Groups: Panic Disorder (PD), Generalized Anxiety Disorder (GAD), and Social Anxiety Disorder (SAD). ML classifiers were trained on MRI-derived regional measures of cortical thickness, surface area, and subcortical volumes to classify patients and healthy controls (HC) for each anxiety disorder separately and across disorders (transdiagnostic classification). Modest, yet robust, classification performance was achieved for PD vs. HC (AUC=0.62), but other disorder-specific and transdiagnostic classifications were not significantly different from chance. However, above chance-level transdiagnostic classifications were obtained in exploratory subgroup analyses of male patients vs. male HC, unmedicated patients vs. HC, and patients with low anxiety severity vs. HC (AUC 0.59-0.63). The above chance-level classifications were based on plausible and specific neuroanatomical features in fronto-striato-limbic and temporo-parietal regions. This study provides a realistic estimate of classification performance in a large, ecologically valid, multi-site sample of youth with anxiety disorders, and may as such serve as a benchmark.