Some wise person in the 90's looked at the 16-32 bit transition, and side-eyed the 64 bit machines that were already available. He noted how differently we write software on 32 bit hardware (flat memory addressing, virtual memory, etc), and wondered what sorts of things we'd do differently on 64 bit. It took a while, but I think the answer is virtual machines, and multiplexing virtual addresses in a single process. There are some key techniques for highly concurrent garbage collection that use features meant for VMs to accelerate mark and sweep by switching the address space of objects that have been swept, and using some page faulting tricks to handle objects that are being written and swept at the same time.

We won't have to go to 128 bit processors for address space reasons. So I wonder what the justifications will be for doing so? And what features people will actually use it for when they have it? I suppose in some ways SIMD has skipped mostly over 128 to go to 256 and 512 bits though, hasn't it?

But I also know of 2 top-tier JITed languages that can use pointer packing to use 32 bit pointers for heap sizes under 32GB to save tons of memory. So 128 bit words would make pointers obnoxiously expensive.