“Effectively the amount of HBM cubes that are harvestable compared to eight high could more than double.”
โ @JordanNanos ยท SemiAnalysis Ep. 030 ยท Long Live the Short King: Why 4-hi HBM Wins (Memory)
Stack height is a supply-chain variable. Almost nobody is treating it that way.
The dominant framing around HBM is capacity per chip โ how much memory sits on a single accelerator. Jordan Nanos at SemiAnalysis flips that lens entirely: the real question is how many harvestable cubes you can extract from a given wafer run.
That reframe matters enormously. If 4-hi stacking can more than double the harvestable cube count versus 8-hi, then the constraint on AI infrastructure supply is not simply fab capacity โ it is stack architecture.
Map of AI lens โ the Enabler layer. HBM sits deep in the AI stack as an Enabler: it does not serve end-users directly, but without it, the compute layer cannot scale. A 2ร improvement in harvestable cubes is not a chip spec upgrade โ it is a throughput unlock for the entire layer above it. That changes how you model AI infrastructure scarcity.
“The real bottleneck might not be how fast you can build fabs โ it might be how you stack the dies inside them.”
FourWeekMBA Analysis
Why The Knock-Ons Matter
Nanos’s argument โ as attributed in the episode โ is not just about the cube count. The implications run downstream: DRAM, logic wafers, substrates, and PCBs all feel a stack-height decision. That means a single architectural choice ripples across multiple supplier tiers simultaneously.
This is a Product Overhang in reverse: a capability already latent in the supply chain, waiting to be unlocked by a design preference shift. The constraint was always architectural, not physical.
If the 4-hi argument holds, the companies to track are not only the HBM fabs โ they are the substrate and PCB vendors who would absorb or benefit from a configuration shift. Stack height is a design decision with a very long supply-chain tail.
The Bottom Line
The AI infrastructure debate is dominated by GPU counts and fab capacity. Nanos’s 4-hi framing is a reminder that supply is also a function of how you architect the stack โ and that a well-placed design choice can double effective output without a single new fab coming online.
Clip via the episode โ @JordanNanos on why going 4-hi HBM isn’t just less memory per chip โ it can more than double harvestable HBM cubes vs 8-hi, with knock-ons for DRAM, logic wafers, substrates, and PCBs. / @JordanNanos @SemiAnalysis_ / SemiAnalysis โ Ep. 030 – Long Live the Short King: Why 4-hi HBM Wins (Memory).
This is FourWeekMBA editorial analysis of a published podcast clip โ not investment advice. All views attributed to @JordanNanos represent his argument as expressed in that episode, not verified forecasts or established fact.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
