Are data sets the new server rooms?

Explores whether proprietary datasets serve as a competitive moat for ML-based companies, analogous to owning server infrastructure. The key insight is that more data isn't always better — companies need a 'sweet spot' where data collection is hard enough to create a barrier but not so hard it's impractical. Three model behaviors are described: convergence too early (data has little value), convergence at a suboptimal level (wrong data distribution), and requiring impractically large amounts of data. Sweet spots for data moats include fraud detection, loan default prediction, and crime detection from security footage.