Larger scale-up domains challenge simple HBM scarcity trades
Bubbleboi highlights a Chinese hardware thesis that larger scale-up domains compensate for less HBM per accelerator. Bubbleboi’s separate skepticism about Nvidia adopting four-high HBM prevents treating that design as a universal roadmap. Jukan adds a possible hyperscaler-backed Intel–Hynix memory venture. The new distinction is architecture and financing: neither a substitute-memory thesis nor exploratory talks establish delivered capacity.