SemiAnalysis analyst covering semiconductors, memory markets, AI infrastructure, and data center buildout dynamics. Returns to podcast after busy period working on deep-dive models and industry analysis.
no scored calls yet — needs a stated position or a categorical verdict, with a matured window vs SPY
Doug argues H100s will become worthless as models grow too large (Kimmy K3 needs B300/MI355X), but Dylan counters that data center retrofit friction (power, cooling, footprint) prevents ripping out Hopper clusters; true test is B200 vs B300 pricing divergence.
SK Hynix missed consensus because they shifted more volume to long-term agreements (LTAs) at lower prices; memory prices 3x'd last year but will only rise 30-50% next year, causing levered investors to face margin calls as the second derivative of price turns negative.
Frontier labs are racing to final pre-train models to IPO and raise final mega-rounds; but even if revenue hits $500B, a $5T cumulative capex implies 10-year payback on revenue (not profit), creating a narrow path where cash flow timing — not technology — becomes the existential risk.