I'm not surprised that you''re befuddled, since you seem to not see the difference between storing data and doing data analysis.
The data sets is in tens of PB, so SSD is way out for cost reasons. But that doesn't matter much, since a small cheap server with 16 7k2 rpm drives can saturate 10Gbit/s networking, in real life use with the storage software written in Java (this position). And we need lots of these servers to make the volume up, so the aggregate bandwidth is pretty large.
The analysis software is up to my users, the LHC experiments, and is mostly C++, but with parts in python and probably some fortran libraries too.