... like drop-in coprocessors like the ol' FPUs, no shuttling weights about like modern GPU/NPUs, memory bottleneck gone, speed at inference per clock cycle, no big brother monitoring service, no subscription payments, offline/airgapped utility at battery power, useful at the consumer level (rather than data-center levels)... just pay for a new chip if you want the next model... maybe that can be the business model instead of perpetual subscriptions.

I wonder if that would even be technically feasible.

It is probably unreachable, as we would surely fall into something like intel upgrade paths, where every generation of processors is incompatible with the next... at least if it were pushed for now.