hbm-moat-vs-dram-commoditization · conviction medium · status playing-out · horizon 2026-2028 (China DRAM-share and HBM-qualification are multi-quarter measurements) · as of 2026-07-29
hbm3e.demand_pull — 37.11 to 48.8Series available as data/hbm-moat-vs-dram-commoditization.csv
Green/blue = model marks it as CAPTURING the rent (unbound and supplies the scarce good); faded = PAYING it (bound severe or moderate). Operating margin, live.
† 5 premises marked supporting — shown and arguable, but the conclusion does not depend on them, so they are not multiplied into the composed figure. Citing a filed figure should not cost conviction.
84% if the 2 gates are independent, 85% if they move together. They are claims about one industry, so the truth is between and nobody can say where. Treat this as an ordering device rather than a calibrated probability — the ranking of premises is the information, not the level.
Weakest link: Incentive × Capacity — the indigenization / margin-compression generator at 0.85 — The prior is medium-conviction and explicitly not backtested.
† 2 premises marked supporting — shown and arguable, but the conclusion does not depend on them, so they are not multiplied into the composed figure. Citing a filed figure should not cost conviction.
47% if the 4 gates are independent, 60% if they move together. They are claims about one industry, so the truth is between and nobody can say where. Treat this as an ordering device rather than a calibrated probability — the ranking of premises is the information, not the level.
Weakest link: SK Hynix — 55% supply share of HBM4 at 0.60 — Contested desk estimate. Samsung's Q2-26 guidance — HBM4 >60% of its HBM mix, 3x QoQ — argues SK Hynix's share is eroding faster than 55% implies.
93% if the 4 gates are independent, 95% if they move together. They are claims about one industry, so the truth is between and nobody can say where. Treat this as an ordering device rather than a calibrated probability — the ranking of premises is the information, not the level.
Weakest link: Memory Super-Cycle at 0.95 — The cycle is observable in ASPs and company guidance.
SK Hynix's Q2'26 higher-HBM/lower-DRAM mix caused a revenue and blended-ASP 'miss' (KRW 79.3T vs ~84T; DRAM 73% of revenue vs 78% prior). Read as a negative: higher-DRAM-mix memory players capture more of the near-term DRAM price surge (contract DRAM +30% QoQ), so they benefit while HBM-heavy SK Hynix's blended ASP lags the market.
The 'miss' is a mix signal, not a weakness signal — and the sign flips once you price the China asymmetry. HBM is the durable, higher-moat, shortage-protected leg (hbm3e supply_constraint severe; SK Hynix ~#1 at ~55% of HBM4, deepening via custom base-die co-design) and China CANNOT enter it at scale on a ~1-2yr timescale (CXMT HBM3E only targeted 2027, 1-2 generations behind, <2% of capacity, export-controlled). Commodity DRAM is exactly where China DOES enter at scale (~9% share, path to ~11%+), so DRAM-mix players face a commoditization wall the HBM leg does not. Allocating scarce wafers into HBM and away from DRAM — the cause of the 'miss' — is therefore SK Hynix choosing the China-proof, more durable business. The miss is bullish long-run.
The desk does not trade the headline miss. It SEPARATES the scoreable from the narrative: China's global DRAM-share trajectory, China's qualified-HBM shipment timeline, and SK Hynix's HBM revenue-share trajectory are dated, sourced, falsifiable measurements that carry weight; the reflexive 'a headline revenue miss means the business is weaker' reading is flagged unfalsifiable-in-advance and carries zero weight. The edge is holding a scored probability on the DRAM-vs-HBM durability asymmetry that a one-quarter ASP-optics reading ignores.
ChangXin Memory Technologies0.99 strongNot in dispute.
ChangXin Memory Technologies — china dram scale entry at-scale0.90 strongScale entry in commodity DRAM is well corroborated.
ChangXin Memory Technologies — china hbm scale entry in-development0.85 strongAn absence-of-evidence claim, so discounted.
China HBM market entry — value in-development0.85 strongChina (CXMT) global DRAM revenue share — value scaling0.90 strongPricing-power incumbent op margins (Finnhub TTM) — value fat0.95 strongSamsung DRAM ASP +mid-40% QoQ and NAND +high-60% confirm fat margins directly.
Incentive × Capacity — the indigenization / margin-compression generator0.85 strongSeven premises, all must hold. Note this is the asymmetry the whole thesis rests on, and the two weakest legs are both absence-of-evidence claims about China's HBM capability — which is exactly the kind of claim a qualification announcement flips overnight.
SK Hynix0.99 strongHigh-Bandwidth Memory (HBM3E / HBM4) — supply constraint severe0.89 strongHBM3E tightness is a claim about DEMAND against fixed supply, so it is governed by what HBM's customers are growing at — recomputed each pulse. Anchors set so today's 43.6 reproduces the hand-scored 0.9. If customer pull collapses, 'severe' stops being supportable and this premise should fall before the narrative does.
HBM supply tightness — value severe0.85 strongSK Hynix — 55% supply share of HBM40.60 moderateSamsung's Q2-26 guidance — HBM4 >60% of its HBM mix, 3x QoQ — argues SK Hynix's share is eroding faster than 55% implies.
HBM4 supplies SK Hynix1.00 strongCustom HBM (custom base die)0.79 strong'Real but early' is exactly the kind of judgement that should track evidence rather than sit fixed. Custom-HBM's customer pull is the observable: if the names adopting it stop growing, 'a direction, not yet a differentiator' becomes the whole story.
The moat leg is stronger than the share leg. The single 0.6 premise is doing most of the damage to the composed number, and it is the same contested share weight that weakens hbm4-share-concentration — the two theses share a load-bearing estimate, which is itself a concentration risk in the book.
Memory Super-Cycle0.95 strongSK Hynix0.99 strongChangXin Memory Technologies0.99 strongSK Hynix supplies ChangXin Memory Technologies1.00 strongThe edge is true but coarse — it does not encode that the competition is segment-specific.
The conclusion inherits the weaknesses above. Note the modelling gap the last premise exposes: a single competes edge cannot express 'competitors in one product line and not another', which is precisely the distinction this thesis trades.
Pre-filled skeleton: hbm-moat-vs-dram-commoditization.md