I think the plausibility comes from the fact that a specialized AI chip could theoretically outperform a general purpose chip by several orders of magnitude, at least for inference. And I don’t even think it would be difficult to convert a NN design into a chip or that it would need to be made on a bleeding edge node to get that much more performance. The trade off would be that it can only do a single NN (or any NNs that single one could be adjusted to behave identically to, eg to remove a node you could just adjust the weights so that it never triggers).

So I’d say it’s more accurate to put it as “the easiest/cheapest way to do an AI device is to use a standard SoC”, but the best way would be to design a custom chip for it.