I agree. There's a wide range of possibilities here.
In the most basic implementation, you'd have 6 features, each with 14 possibilities, and you choose one option from each and combine them. That gets you 7.5 million combinations.
A more sophisticated implementation might pull from a larger number of features and options and apply complex transforms to produce nonformulaic output.
My guess is that what's actually cutting-edge here is not the algorithm or the designs, but the ability to print custom labels at all. My guess, though, is that it's only cost-effective as a publicity stunt. Printing 7 million identical labels is likely far less expensive.