The theoretical switching density isnt the practical...

The theoretical switching density isnt the practical manufacturable product density. Its a bit like a supercar versus a normal car. You have to do more hand configurable things and exotic things (carbon fibre, titanium parts, hand built engines) to get the performance and the cost goes up dramatically. For any given drop increase in chip density, the supporting components become more complex, there are more things to avoid, like crosstalk between paths and the tolerances become more important. The chance of variability sufficient to render things not working lowers the potential yield so you have to increase relative separation between components etc in the lower more compact design. All that costs, hence why the price goes up after a certain point.IBM can flip individual atoms around to make memory but the infrastructure to do that supported the memory cell is huge and costly and providing it would render a working chip completely uneconomic. Just because you theoretically can do 1nm or even do it in a lab doesnt mean its economic to so in production.

Now, there will always be a need for chips that push the boundaries and the money will be spent on higher density regardless of the added costs. And embedding memory at that density is useful. Having the technology in silicon oxide is an advantage, since it means the memory in the embedded situation wont need different masks or special layers for exotic materials to get the chip together from the other masked out components. You design the chip in its totality and there are effectively no extra steps in the process compared to the other components for the memory. I am not sure about Crossbar's additional steps when being embedded, nor if they have surmounted the additionaol costs at 10nm relative to the rest of the chip being made. Its only after this can you understand the relative advantages they may or may not have.