a16z Chief Calls Samsung, SK Hynix as Critical as Nvidia in AI - 조선일보

a16z's framing exposes a pricing error: the AI capex narrative has been monopolized by Nvidia, with HBM valued as a commodity input. But physics doesn't follow narratives—GPU compute ceilings are ultimately bandwidth-bound. Samsung and SK Hynix don't sell 'memory'; they control the flow valve of AI compute. Technically, HBM4 pushes TSV stacks to 12-16 layers with hybrid bonding replacing micro-bumps—a qualitative shift in DRAM-process and advanced-packaging coupling. Samsung's 2nm GAA will feed HBM4 base-die performance; SK Hynix's 1c-node yield determines 2026 supply elasticity. This is a race to break physical limits, not incremental improvement. Risk-wise, HBM's inclusion in US export controls creates structural tension for both Korean players. More critically, effective global HBM capacity splits roughly 50% SK Hynix, 35% Samsung. A single yield event halts GB200 shipments—and unlike logic chips, HBM has no substitution path. This is more rigid than the 2021 shortage. Strategically, Nvidia is shifting from procurement to capacity lock-in via prepayments and co-development. AMD's MI350 demands 40% higher HBM4 bandwidth density, forcing accelerated iteration. Micron leverages domestic-manufacturing politics for second-source share, but yield barriers aren't capital-solvable. 12-24 month call: the cyclical-stock paradigm for memory breaks. PE anchoring shifts from historical averages to Nvidia's compute growth curve. The long-tail effect: HBM4 hybrid bonding converging with 3D NAND will spawn 'storage-defined compute' architectures—potentially eroding Nvidia's moat from below.