Hardware

No, you are limited by:

Compute Performance, you will need 10,000%+ more compute than was available per chip, and those PCIe accelerators don't have the ability to compute the way they do now. You are going to have to rely on CPUs which is worse

Lack of scalabality of interconnecting chips to behave as one, increasing IO requirements dramatically.

Lack of memory pooling (yes you qualified it), memory bandwidth and memory sizes (we are talking in megabytes), imagine waiting for 1 billion parameter model calculations to load and store in each layer of a neural network using floppy disks.