The primary advantage of something like this over doing multiple die on a normal organic package is the increased bump density, allowing for more I/O, but it do

The primary advantage of something like this over doing multiple die on a normal organic package is the increased bump density, allowing for more I/O, but it doesn't really help you with power density at all. It's not that hard to be power-limited at well below even single-reticle sizes, though, in which case this doesn't really help at all. I could see AMD moving to something like this for integrating HBM in a future epyc/threadripper part, where they might want to add a lot of not-particularly-power-hungry silicon.