Semiconductor fabs usually not make ultraclean materials themselves, but bought from many smaller specialized entities. As example exist few companies, producing ultraclean silicone as crystal cylinders. When fab bought such crystal they could slice from it few wafers and made test runs with their product line designs. For example, they could try i7-7700, i7-7700hq and i7-7700k. If they seen good output with 7700k, it is best possible, as these chips have very similar die size, but 7700k is most expensive; if they seen moderate output with 7700k, but good output with 7700, it's also ok, for them also good demand; but if only 7700hq have good output, things are not good, because notebook chip is not so easy to sell, so have to do some marketing (usually these seen as some exceptionally high prices for some chips, because of low silicone output).
Unfortunately, semiconductors have limits on how large could be one die. Companies like Cerebras use some tricks, usually they just disable parts of die with jumpers, as I know, same tricks used Nvidia before 30th series, but approx same time appeared whole industry movement for chiplets standards, and looks like Nvidia used 2.5D technology to avoid trap with too large die.
As I know, in ~1990th HBR written article about constantly under-loaded semiconductor fabs and concluded, it is unprofitable to tightly couple them to R&D.
Counter-argument was, that Intel used ties to fabs, to achieve extreme level of scalability, to fill market demand fastest, so marketing won.
What really happen, appeared few new specialized classes of semiconductors - signal, accelerators, high-power (high-current), and low-power (energy effective), and independent fabs made universal pipeline, to fill all market demand, but Intel stuck on desktop CPUs and failed all other classes (as example, Intel was unsuccessful in try to got niche on smartphones SoCs - still have not made cellular modem and nearly failed on GPUs).
To be more exact, Nvidia in reality is most software company from hardware companies, and AMD with their GPU division constantly competes to Nvidia literally head-to-head.
And what was gamechanger - when AMD struck to limits of reliable transistors on one die, they decided to switch to chiplets - they made 2.5D multiple-die design with silicon interimposer, while Intel used their manufacturing superiority to make huge dies with all included.
- Once appeared, with chiplets, AMD could achieve much better performance on weaker but much cheaper technology and won.
So my point - Intel suffered from too tightly couples with fabs, so once they have to adapt their designs and marketing to semiconductors, when AMD successfully avoided this trap. BTW, for this exist much better example - similar problem once killed Atari and Commodore.