[7777fps]

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.

[vidarh]

Putting a caching layer in front of AWS is often very cost effective even without much skew in the access pattern. It tends to take a very low hit rate before it pays for itself.