As a compromise you could have it be more about availability and rate-limiting - if you've got a given piece of data split across 4 currently-live mirrored storage disks and you've got 32 front-end machines able to serve it up to users, you're using operational capacity for that which could be used for other higher-demand data. Assuming you have a robust backup/recovery strategy, low-demand data could be stored across 2 or 3 live disks (note: you'd still want backups somewhere for emergency recovery) instead, and have a smaller number of front-end machines dedicated to providing data to users.
In that case if the low-demand data front-end machines get too much load, you'd be entering a queue to be served.
Not sure that's realistic here though, it sounds like the main problem is just the volume of data so whether the storage for the data is hot, warm or cold isn't necessarily going to make a big difference and having enough frontends to serve the content isn't a big problem (in comparison) either