I feel like early "AI PCs" will be just that - early. 10 years from now people will be trying to run interesting models on anemic processors to see what they can do, meanwhile contemporary systems will fly through modern quantized models.

There's basically nothing in the way except high memory prices, and I hope the intense demand ignites new memory factories.

It's just another evolution of PC technology. First it was all about the CPU and cache, then later memory and disk and other I/O (network), and now it will be all about parallel token processing.