Not all data is used equally: some is read constantly, while much is kept only in case it is ever needed. Storing everything on the fastest, most expensive storage wastes money, so systems separate hot data, accessed often and kept close at hand, from cold data, rarely touched and moved to slower, cheaper storage.
The saving is large because cold data is usually the bulk of it. The judgement is in predicting what will be needed quickly, since fetching cold data takes longer, so misjudging leaves someone waiting for something assumed to be forgotten.