When using the H2OPrincipalComponentAnalysisEstimator object to calculate PCA components of data, the first components don't have zero mean as expected when categorical variables are present. If those variables are dummified upfront to 0/1 variables per variable category, the result is as expected (i.e. all zero means).
This looks like there might be a bug in PCA for categorical features. Reproducible code snippet below
Where the means look like
for categoricals and
for manual one-hot encoding the numbers look like: