The unbearable slowness of tapes

It is still common for the lowest layer of data storage to be 'tape'. While individual tape devices apparently offer very good performance, this is only achieved in a limited set of circumstances, and the overall throughput can easily become very poor very quickly. How can this be avoided, if at all? Is it sufficient to just add more and more equipment? These difficulties will be discussed in the context of CERN's current CASTOR HSM version, and the need to 'carry LHC data forward' for many years.