That was the gist of the response, as far as I remember. It was repeated access from the same set of machines, so it was more likely that one bit flip persisted in a config and was subsequently rolled out to others. So the study grossly overestimated the number of actual bit flips occurring by like 20000x or something.