I assume the data accessed is a heavily skewed pareto distribution.

Given that, it's maybe still cheaper to build their own serving / caching layer in front to save egress costs than to have constructed the whole storage solution themselves.