When studying the performance of distributed infrastructure in earthquakes, spatial variations in strong ground motion have a significant impact. Currently, prediction models for spatial ground‐motion variations in future earthquakes are calibrated using ground‐motion observations from densely recorded earthquakes. Although useful, that calibration process requires strong assumptions about stationarity and isotropy of correlations. This article reports results from conducting analogous spatial variation estimation using physics‐based simulations from the CyberShake platform. This platform contains simulated ground motions from hundreds of thousands of rupture realizations, at locations throughout southern California, providing a synthetic ground‐motion catalog that is much richer than we could ever hope to achieve from recordings. That richness allows significant relaxation of stationarity and isotropy assumptions, and provides new insights regarding the role of source and path heterogeneity on the spatial correlation of ground‐motion amplitudes. The results suggest that geological conditions, source effects, and path effects have significant impacts on spatial correlations. In addition, this work serves as a new dimension of ground‐motion simulation validation, because the estimated correlations can be compared to results from past earthquakes.