I always found this a bit of an omission in Bostrom’s argument in Superintelligence - on one hand he’d say we should avoid giving AI a goal like paperclip maximisation while overlooking that the method of transistor maximisation used to attempt to build the Superintelligence was the same kind of goal. Same goes for whatever system he fantastically proposed to run his fanciful “mind uploading” Cartesian scenario.

The difference is the demand for paperclips is already met. Each additional paperclip is almost pure waste of resources.

The demand for silicon, even outside AI is not, and advances in silicon are going to benefit pretty much every industry in the long run, even if during the short term they are distorted by the AI demand.