They're both difficult to grok but for different reasons. (Well, the same underlying reason - bandwidth - just accounted for differently).

Glacier, as the article points out, is difficult to grok because there are provisioned bandwidth costs.

Nearline is difficult to grok because it simply caps your bandwidth at 4 MB/s per TB stored. It starts at 1 MB/s.

Basically, both are optimized to make retrieval slow and/or expensive because the cost optimizations they're doing internally are incentivizing very cold data, very big data, or both.