Three rivals, three different fights. Samsung fights for the same customers. Micron sets the multiple ceiling. CXMT attacks the commodity floor. Updated Jul 2, 2026.
| Company | HBM position | Latest generation | Stacking tech | Profitability | Valuation | US access |
|---|---|---|---|---|---|---|
| SK Hynix | ~58-60% share — the leader | HBM4 shipping at scale · HBM4E 12-layer sampled Jun 18 (first) | Advanced MR-MUF — best heat + warpage control | 72% op. margin (Q1 26) — industry best | ~8× fwd | Nasdaq ADR Jul 10 · Frankfurt HY9H.F · KRX 000660 |
| Samsung | ~high-20s-30s% — chasing | HBM4 mass production since Feb 26 · "Super-Gap" roadmap | TC-NCF — fights warpage with force; weaker thermals | Lower memory margins; conglomerate blend | discount to peers | No ADR yet — investor inquiries surging after our filing · London GDR SMSN |
| Micron | #3 — HBM4 36GB in volume for Rubin | HBM4 ramping; no HBM4E sample announced | TC-NCF | FQ3 revenue $41.5B (+346%), GM 84.9% — cycle confirmed. But its new deals cap prices at Q2-26 levels; SK's have no cap | 15-29× fwd | US-listed (MU) — the benchmark our re-rate is measured against |
| CXMT (China) | None — no credible HBM | Commodity DDR/server DRAM · 3% → 8% world DRAM share in a year | n/a (no stacked HBM) | State-backed; price weapon, not margin story | STAR IPO pending | None · $2.94B (20B+ yuan) Tencent server-DRAM deal, 3-5 yrs |
The threat: HBM4 shipping since February, capacity expanding, "Super-Gap" roadmap aimed at reclaiming leadership. Tripwire: Samsung winning a majority HBM4E slot at Nvidia or Google — that would hit L2 (our share) directly. Until then, our MR-MUF stacking keeps a real physics edge on heat and warpage — the exact failure mode that just killed Nvidia's 4-die Rubin Ultra package.
New (Jul 3): Anthropic is reportedly in talks with Samsung to develop custom AI chips — another AI exposure layer forming — and Samsung shares in the ₩800T national fab program. Both strengthen the case below.
The opportunity (diversification plan): Samsung is the designated first buy after the SK re-rate fires — cheaper multiple, same shortage, lagged beneficiary of the cycle. Vehicle: SMSN (London GDR — works on CommSec; KRX 005930 doesn't). New kicker: overseas-investor inquiries about a Samsung ADR surged after our Nasdaq filing — if Samsung follows us to New York, the diversification target inherits its own re-rate catalyst. BHP remains the secondary, uncorrelated hedge option.
FIGHT: HBM4E sockets · WATCH: quarterly share, ADR filingMicron is what a US-listed memory maker is paid: 15-29× forward earnings, while we sit near 8×. That gap is the L4 mispricing the ADR attacks. Their just-reported quarter (revenue up 4×+) independently confirms the cycle — good for us, since we're bigger in HBM. Discipline: the Micron ceiling caps our multiple (~10×), not our price — growing profits at a flat multiple still compounds the share price. Their debut history is also our playbook: a 10-15% listing-week drawdown is normal, not thesis-breaking.
FIGHT: none · USE: valuation anchor + debut precedentChina's DRAM champion tripled its world share to ~8% in a year and just signed a 3-5 year, $2.94B server-DRAM deal with Tencent ahead of its IPO. Apple is reportedly lobbying to source from them too. What it touches: commodity DRAM — roughly 30% of our revenue, where prices are set at the margin. What it can't touch (yet): HBM — no credible product, and DDR5 yields still lag badly. This is exactly the 2027-28 structural threat in the bear file, and the main reason the engine's Scenario C exists. Tripwires: a credible CXMT HBM sample at any hyperscaler, or >15% world DRAM share.
FIGHT: commodity price floor · WATCH: quarterly share, yield claims, any HBM claim| Company | Bottleneck | State of play | Valuation | Framework verdict |
|---|---|---|---|---|
| Nittobo (TSE 3110) | T-glass cloth — ~90% world share; every AI substrate needs it | +210% in a year, peaked ¥6,580 (May 7), now ¥3,625 — −45% off ATH in the AI rout. Tripling capacity; relief mid-2027. ¥4.4B-ish market cap, beta 2.0 | ~15.8× trailing (flattered by one-off asset-sale gains; core ~20-25×) | Watchlist only. The cheap entry existed a year ago; today it's a crowded, high-beta proxy for the same AI-capex bet we already hold at 100% |
| Mitsui Kinzoku (TSE 5706) | HVLP copper foil — one of ~5 non-Chinese suppliers for AI-class boards | Record FY (NI ¥91.3B); Malaysia foil exports +91% YoY; MS PT ¥56,314, GS Buy. Conglomerate: foil sits beside zinc smelting and auto parts | ~26-30× — vs peers 19× and industry 14× | Fails the low-PE rule — same class as the rejected energy/gas hedges. Use its export data as a demand gauge, not a position |
| SK Hynix (SKHY) | Micron (MU) | |
|---|---|---|
| HBM share (Counterpoint, Q1) | 58% | ~21% (Samsung ~21%) |
| Latest quarter | Q1: rev ₩52.6T (+198%), OM 71.5% | FQ3: rev $41.5B (+346%), GM 84.6%, EPS $25.11; guiding $50B |
| Forward P/E | ~5.4-6.1× | ~6.7× (FactSet) |
| HBM4 status | Mass production since Sep 2025; ~70% of Nvidia HBM4 allocation (chatter) | Now in high-volume shipment for lead customer; HBM4E volume cal-2027 — the gap is narrowing (bear file #10) |
| Demand engine | Nvidia Q1 FY27: revenue $81.6B (+85%), data centre $75.3B. Baker: HBM = 30-40% of ALL hyperscaler capex by 2027; no fourth supplier exists. | |