The imprecision of anatomical alignment methods commonly limits the spatial resolution and sensitivity of conventional fMRI analysis based on statistical parametric mapping. Recently proposed machine-learning methods aim to circumvent the cross-subject (XS) alignment problem by computing a linear projection of the fMRI signal from each subject's anatomical space to a common albeit abstract "functional" space . The success of these "hyperalignment" methods is often attributed to a spatially and functionally specific (linear) correspondence between the fMRI signal in different subjects under similar stimulation conditions. Cross-subject PCA of averaged fMRI data from repeated movie-viewing experiments reveals smooth globally distributed fMRI signal components that facilitate robust cross-subject classification by Linear Discriminant Analysis (LDA). Such global cortical network activity may contribute to the success of fMRI hyperalignment strategies.