SemiAnalysis Weekly: why 4-Hi HBM may win on inference economics
Metadata-only capture of SemiAnalysis Weekly Ep. 030: Myron Xie and Jordan Nanos on Rubin Ultra shipping 192GB HBM versus a 1TB preview, supply-driven decontenting, and why less memory per chip can still be the right call.
Abstract
Myron Xie and Jordan Nanos discuss NVIDIA's Rubin Ultra HBM configuration: a preview once pointed at 1TB per package while the part that ships is described as carrying 192GB. Show notes frame the cut as supply-driven rather than performance-driven, and argue that shipping less memory per chip can still be the correct economic choice — "for the first time in recent memory," a next flagship with less capacity than what it replaces. Related newsletter title: Long Live the Short King: Why 4-hi HBM Wins. Full transcript not processed this pass; theses and key math deferred until a transcript or ASR treatment.
The Open/Close · Research commentary, not investment advice. Positions may be held in securities mentioned.