Even getting the data to that machine presents a problem that seems to undermine the value of it entirely: As you mentioned, the real value in this is linear processing of large sets of data, but the storage is ephemeral so your process has to be some variation of firing this instance up, copy TBs of data to the machine, and then do linear processing. Given that you have to get the data there, the value of the high aggregated gross-throughput seems secondary -- just stream process it, etc.

I'm having a tough time seeing where this type of instance fits.