If you want to build something like youtube, you could store the data from a start of a video, and data defining a slower resolution stream, and more popular videos on the lower latency parts of tapes in your storage arrays, and while the video plays you could go get the rest from the higher latency places.
You could also predict What is the next video the user will want.
Also, the 2 minute seek time is from one end of the tape to the other. But you could also move accross tracks, vertically so that gives another route for short seek times for some of your data.
So it might be possible to build something with a reasonable performance.