From what i understand the training data and careful curation of it is the hard part. Everyone wants training data sets to train their own models instead of producing their own.