I use the analogy with the previous series - it claims to support 64GB, but only 32GB in lpddrX memory mode always, for a number of reasons. Most likely, it's exactly the same here - on 2 slots with 2 х 128 modules (in the future, 384 with 192GB modules), and in the soldered version, no more than 128(64х2).
For normal work, this seems to be a monstrously large amount, but not for the operation of neural networks - for them 128GB is the minimum amount for a neural network with a serious level of complexity. If I guess correctly from the emerging trend, soon even 256GB will seem like a relic of the past, sort of like Bill Gates once said that 640Kb is enough for everyone. The question is what is enough for and for what tasks.
Complex local (autonomous) neural networks will quickly require tens of terabytes of operational for their structures.
Apparently, in the near future, mass production of multilayer RAM will begin, with a capacity that exceeds the current levels familiar to ordinary people by orders of magnitude...