Let's say that a hard drive (or whatever storage device) has a failure rate of 1% over 1 year. With traditional hardware, for every 100 computers you have, you'll replace 1 hard drive. A new hard drive will cost a fraction of the machine's initial price, let's say 1/5. So maintenance costs for storage devices are 1/500 the initial costs per year, maybe a bit more if you factor in the cost of the labor.

Now if you make the same cluster of servers out of ARM hardware, let's say you'll need 4x the number of machines to get the same processing power. That's 400 machines. If the storage devices on these machines have the same failure rate, you'll need to replace 4 machines per year. However, since you're buying a new machine, you're don't get to pay for just the storage device. It costs you 4/400, or 1/100 the initial cost to maintain the storage devices for your ARM cluster per year.

A huge assumption here is that both types of hardware have similar initial costs. So the point here is that it'll always be cheaper to compartmentalize your losses, unless ARM devices are much cheaper than traditional commodity hardware. Intuitively, throwing away a whole machine when something breaks is going to be a lot more expensive than just replacing the broken part. I don't imagine that it's extremely difficult to make parts replaceable on ARM boards, and it would definitely save some money, so I don't see why it shouldn't be done.

(Also, I realize that the Beaglebone uses SD cards, which are cheap and replaceable. But imagine that instead of storage devices, I'd used memory for the example)