What makes you say that ? - assuming a reasonable number of variables I would think you'd only end up needing that order of magnitude if you split every combina

What makes you say that ? - assuming a reasonable number of variables I would think you'd only end up needing that order of magnitude if you split every combinations of variables into it's own test group to do A/B testing. Presumably the reason that you'd apply a machine learning algorithm in this case is that you could extrapolate from a much smaller test group.