SK Hynix captures the single largest slice of the AI-memory scarcity rent — roughly 62% of high-bandwidth memory, about two-thirds of NVIDIA's HBM4, sold out and pre-paid into 2027 — and earns it at a 72% operating margin and a ~61% return on equity. And yet it trades like a Korean memory-cyclical at a single-digit forward multiple. The rent shows up in the earnings, not the multiple. Management is so convinced of the mispricing that it took the company to Nasdaq in July 2026 specifically to force a re-rating against Micron — a rare case where the best fundamentals in the cluster come with the most explicit, dated catalyst. BULLISH, conviction MEDIUM: the lead is yield-and-trust rather than a monopoly — Samsung out-qualified it on HBM4 timing yet won only a quarter of the allocation — and two-thirds of HBM riding one customer is the risk that caps conviction. The structural driver underneath the share numbers is a physics gap: compute scales roughly 3x every two years while memory bandwidth scales only about 1.6x, so each accelerator generation needs proportionally more HBM than the last — 80GB on H100, 192GB on Blackwell, more again on Rubin. That is what makes this a rent rather than a cycle. The bottleneck widens with every generation instead of clearing. (Merged in from the retired "HBM Suppliers Are the Picks-and-Shovels of AI" call, 2026-08-03.)
Note, 9 Oct 2026. The Nasdaq listing this call named as its catalyst, and as the second half of "What would prove this wrong", completed on 10 Jul 2026 (Form 424B4). That half of the falsifier can no longer fire. From here the call is tested on its first half alone: whether Samsung's yields converge and NVIDIA rebalances HBM4 allocation.
TSMC
Broadcom
Micron