I think the lesson came in earlier in the NUMA and MPP machines where they kept trying to cram more stuff on boards that were themselves pluggable into the larger system. This convergence has happened from several directions. It's not all the different from the earlier one that started in the 1960's where they fought cost and inefficiency by getting as few components per box sharing as much as possible. Moores Law temporarily reversed it (transistors and memory are free!) then reality check hits that this seems to be a fundamental principle.

My design a while back was to put it all on PCI cards on a PCI backplane. I saw backplanes that basically look like motherboards full of PCI slots that load into racks. I wanted to make the cards nothing but CPU and memory whose software communicated over efficient networking (not TCP/IP) through PCI DMA. My design had IO/MMU functionality in the backplane or PCI cards. At least one card having full-featured stack for management and at least one I/O card for external interface. I figured the backplane itself could be extended for that, too, with a dedicated port like motherboards do integrated GigE. Management and I/O could come through remote DMA over dedicated wires like many servers do with Ethernet so all the PCI slots could be dedicated to compute.

Dumbest thing about Facebook's model is them destroying drives. The first thing to notice, due to Ross Anderson's Security Engineering, is that those pieces still contain a lot of data if they weren't degaussed first. Next is to remember the fastest way to destroy data: use clustered, encrypting filesystems so that secrets never touch the drive. Then, you just have to delete the keys to loose the secrets. No need to trash the drives at all. The crypto can happen at the storage manager or at hardware interface with HW acceleration available for both types. I'm surprised they haven't already built this with all the smart people they have working on big-data stacks.