Considerable phenotypic variation under identical culture conditions limits the potential of stem-cell-based embryo models (SEMs) in basic and applied research. The biological processes causing this seemingly stochastic variation remain unclear. Here, we investigated the roots of phenotypic variation by parallel recording of transcriptomic states and morphological history in individual structures modeling embryonic trunk formation. Machine learning and integration of time-resolved single-cell RNA sequencing with imaging-based phenotypic profiling identified early features predictive of phenotypic end states. Leveraging this predictive power revealed that early imbalance of oxidative phosphorylation and glycolysis results in aberrant morphology and a neural lineage bias, which we confirmed by metabolic measurements. Accordingly, metabolic interventions improved phenotypic end states. Collectively, our work establishes divergent metabolic states as drivers of phenotypic variation and offers a broadly applicable framework to chart and predict phenotypic variation in organoids and SEMs. The strategy can be used to identify and control underlying biological processes, ultimately increasing reproducibility.