Even AMD realizes the jig is up; they dumped their fabs because they realized they didn't need them anymore. It's not about having the best damned process available anymore. It's about having the lowest power design, the smallest design, the widest/most-parallel design.
That'd be likelier to be because AMD's past history fabbing things is so miserable they couldn't get their hands on the $10s of Bs needed. Intel'd disagree with you - they keep their mastery and the process tech lead they've commanded through most of the microchip's existence. They've done so well because that's always been their first priority, and they put a ton of leading engineers into it, unlike other chip companies.
Fab investment and construction stay strong, just concentrated into fewer and changing actors, like every other industry. Chip density, which's what Moore's law predicts, rather than clock speed, still continues strong. Intel's new chip efforts are about using that density to put more and more stuff into the same chip, and about putting more cores per chip. And, it's about a bigger share of PC costs going to Intel instead of overall PCs getting more expensive.
Now, these multicore chips are a pain for getting gains from our traditional programming models, but neural simulations are the most trivially parallelizable thing out there.