> All of this is without tackling the other elephant in the room, which is the end of Moore's Law. Moore's Law has been on its death bed for over a decade now. We're seeing only limited improvements in computing and storage performance, mainly from parallelism.

> (it's one of those goddamn bubbles: to the limited extent that LLMs are useful, we'll inevitably see a shift towards using pre-trained models running on local hardware)

I don't completely disagree with his main theses, but this part seems contradictory. You cannot state at the same time that Moore's law is over - so the compute capacity of consumer PC will roughly stay the same from now on - but also that we will all just casually run GPT-5 locally on those same PCs in the future and therefore the data center buildout is a waste of money.

(I also think it's a bubble, but more because there might be a lack of demand, not because some technological miracle might make all the infrastructure obsolete)

Doubling capability every 18 months while halving in price, is not the opposite of 'compute capacity staying roughly the same from now on'.