The AI-data category spans a steep moat gradient that its collective label obscures. At the top sit toll booths — proprietary data fused to sticky professional workflows a court just affirmed cannot be relabeled around. In the middle: scarce but non-exclusive corpora and stock-photo libraries licensing to their own displacement. At the bottom: annotation labor with low switching costs, the layer most exposed to in-housing. The distinction holds: a unique corpus is a stock you own; annotation capacity is a flow you rent. Scarcity compounds the former far more durably than the latter.