← research bank

HBM4 Qualification Concentrates Share, Not Spreads It

hbm4-share-concentration · conviction medium · status open · horizon 2027 · as of 2026-07-22

HBM4 qualification concentrated volume at the yield leader (SK Hynix ~70% of NVIDIA Vera Rubin) rather than spreading it across the three qualified vendors — the CORE, fully-supported claim. The knock-on that HBM stays structurally tight through 2027 (2027-28 capacity does not commoditize HBM, contracts rise severalfold, ~6x forward multiples re-rate) has WEAKENED as of 2026-07-22: the model's own hbm4 supply_constraint eased from severe to moderate, so structural tightness now rests on supplier guidance (SK Hynix CEO's 'worst supply year' call) running ahead of the model rather than on a model-confirmed severe constraint. Concentration holds; the tightness leg is now contested.
Rests on shares nobody discloses. 1 of 2 derived inputs move materially when the undisclosed supply weights are redrawn across their plausible range. The argument may still hold — but these figures are ranges, not points. Computed from evidence at most 17 days old (oldest input: alphabet).
hbm4.demand_pull — 102.9 to 152.3

Exhibits

Exhibit 1Relative performance, indexed to 100How the names in this thesis have traded against SOXX.
86051201MU 739000660.KS 573005930.KS 331SOXX 225NVDA 12312mo, indexed to 100 at start · dashed = SOXX benchmark

Series available as data/hbm4-share-concentration.csv

Exhibit 2Who pays Conventional DRAM and NAND supply (HBM crowding-out), and who keeps the moneyCapturers average 56.8% operating margin against payers' 32.4% — the owners of the scarce thing capture the rent, as expected.
Micron Technology80.4%NVIDIA Corporation64.0%SK Hynix58.6%Samsung Electronics24.2%Apple Inc.32.4%

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.

Exhibit 3What the conviction is actually made ofEach premise and the number it composes to. A conjunction of plausible premises is far weaker than any of them.
Current-generation HBM (HBM3E) is a severe, SK-Hynix-concentrated bottleneck today; the model's next-gen reading has EASEDHigh-Bandwidth Me… — supply constraint seve… †100.0%HBM4 — supply constraint moderate †68.5%HBM supply tightness — value severe †85.0%SK Hynix — 62% supply share of High-Bandwid…60.0%COMPOSED (and)60.0%

† 3 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.

One gating premise, so the conclusion is exactly as strong as it. The figure is an ordering device, not a calibrated probability — see how the numbers are made.

Weakest link: SK Hynix — 62% supply share of High-Bandwidth Memory (HBM3E / HBM4) at 0.60 — UNPRICED DEPENDENCY (probabilities.py flags this as deps=3). The 62% share is a desk estimate, not disclosure — SK Hynix does not break out customer s

HBM4 qualification concentrated share rather than spreading it (CORE claim — fully supported)HBM499.0%SK Hynix — 55% supply share of HBM460.0%Samsung Electroni… — 25% supply share of HB…50.0%HBM4 — signal 2026-07-2280.0%COMPOSED (and)23.8%

24% if the 4 gates are independent, 50% 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: Samsung Electronics — 25% supply share of HBM4 at 0.50 — Most contested number in the thesis. Samsung's Q2-26 guidance implies it is gaining faster than 25% implies. Scored at a coin-flip because the disconf

Whether tightness is STRUCTURAL through 2027 is now contested, not model-confirmed (WEAKENED)HBM4 — supply constraint moderate100.0%High-Bandwidth Me… — supply constraint seve…100.0%SK Hynix — signal 2026-07-2280.0%COMPOSED (or)100.0%

100% if the 3 gates are independent, 80% 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 — signal 2026-07-22 at 0.80 — SK Hynix CEO 'worst supply year' commentary; company source but forward-looking opinion, not a disclosure.

HBM4 share is decided OUTSIDE the memory layer — by bonding capability, base-die foundry access, and wafer arithmeticHybrid Bonding75.0%BE Semiconductor Industries (Besi)70.0%Taiwan Semiconductor Manufacturing Company70.0%Conventional DRAM (DDR5 / LPDDR)80.0%COMPOSED (and)29.4%

29% if the 4 gates are independent, 70% 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: BE Semiconductor Industries (Besi) at 0.70 — Bonding capability is an EQUIPMENT question before it is a memory one. BESI, ASMPT and Hanmi supply the tools, and domestic share in bonding is near z

SK Hynix is the levered, mispriced expressionSK Hynix — 55% supply share of HBM460.0%SK Hynix — 62% supply share of High-Bandwid…60.0%SK Hynix supplies NVIDIA Corporation100.0%SK Hynix — forward P/E 6.3x †48.2%SK Hynix — signal 2026-07-2280.0%COMPOSED (and)28.8%

† 1 premise 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.

29% 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 (see above).

The tightness plugs into the main accelerator artery, giving it a demand anchorHBM4 supplies NVIDIA Corporation100.0%SK Hynix supplies NVIDIA Corporation100.0%HBM4 — signal 2026-07-2280.0%COMPOSED (and)80.0%

80% if the 3 gates are independent, 80% 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: HBM4 — signal 2026-07-22 at 0.80 — Desk signal, trade-press sourced.

The variant

Consensus

Memory is a cyclical business at a peak: SK Hynix and Micron trade ~6x forward earnings (Micron ~10x FY2027 EPS), because the market expects the 2027-28 capacity wave (Samsung P5, Micron/SK Hynix fabs) plus HBM4 being the first generation where all three vendors qualified at the start of the ramp ('battle pivots to HBM4', Samsung catching up) to spread share, add competing supply, and roll HBM pricing over into a glut. Tightness is real now but temporary and mean-reverting.

Variant

HBM4 qualification concentrated share rather than spreading it. Even with SK Hynix, Samsung and Micron all certified for NVIDIA Vera Rubin (June 5 2026), SK Hynix holds ~70% of NVIDIA HBM4 because the base-die-on-logic architecture (TSMC 12nm/3nm vs Samsung 4nm) raised the qualification bar and made the split a yield question, not a binary pass. Combined with HBM's wafer intensity crowding out commodity DRAM (~half of 2027-28 memory-wafer growth), effective HBM supply stays concentrated and short through 2027 — SK Hynix's own CEO calls 2027 the worst supply year in industry history — so 2027 contracts print severalfold higher and the trough multiples re-rate.

Differentiator

Consensus counts qualified VENDORS (3 = looser) and capacity announcements (glut in 2027-28). The variant watches the SHARE SPLIT among the qualified and the 2027 contract ASP: multi-vendor qualification with concentrated volume AND severalfold price hikes proves qualification added no competing effective supply. The tell the market is missing — passing NVIDIA's HBM4 cert is necessary but not sufficient; the base-die-on-logic step converted 'who qualified' into 'who yields', and yield concentrates.

Stance
Own the concentration, not the cycle: HBM4 qualification raised the bar (base die on logic) and kept ~70% of NVIDIA's volume at SK Hynix, so the tightest supply year on record (2027) arrives with pricing power intact — long SK Hynix (cleanest, cheapest), long the share-agnostic HBM test enablers, and fade the Samsung 'catch-up' convergence; the risk is the 2027-28 capacity wave and any AI-demand air-pocket.

Recommendations

Second order

Third order

Indicators

Tripwires

Falsifiers

Reasoning chain

Current-generation HBM (HBM3E) is a severe, SK-Hynix-concentrated bottleneck today; the model's next-gen reading has EASED VALID
premises

Today's bottleneck is real and concentrated. The conjunctive form is deliberate: this conclusion needs all four to hold, and the weakest link is the unsourced 62% share estimate.

HBM4 qualification concentrated share rather than spreading it (CORE claim — fully supported) VALID
premises

The CORE claim is only as strong as the share weights, and two of the four premises are contested desk estimates facing company guidance that points the other way. This is why the composed number lands well below what 'fully supported' in the original conclusion text implies.

Whether tightness is STRUCTURAL through 2027 is now contested, not model-confirmed (WEAKENED) VALID
premises

OR because tightness persisting needs only ONE of the generations to stay constrained — HBM3E severity alone sustains it even if HBM4 loosens. Note the counter-evidence held in view: Jefferies has 3Q26 hikes at 15-20% vs 25-30% expected with CE buyers pushing back, which caps how far this runs.

HBM4 share is decided OUTSIDE the memory layer — by bonding capability, base-die foundry access, and wafer arithmetic VALID
premises

THIS THESIS HAS BEEN ARGUED ENTIRELY INSIDE THE MEMORY LAYER, and that is its weakness rather than its focus. Every prior premise references hbm3e, hbm4, sk-hynix or hbm-supply-tightness. But the variables that actually decide HBM4 share sit elsewhere: WHO CAN BOND (equipment, and export-controlled since 1 August), WHO GETS FOUNDRY ALLOCATION for the logic base die (foundry, and not the memory maker's decision), and HOW MANY WAFERS A BIT COSTS (materials, at a 3:1 penalty). A memory analyst can see the first set. Only a model that spans layers can see the second, which is the entire reason to hold one. HELD AT AND: all four must bind for share to stay concentrated. THE FALSIFIER IS CLEAN — if a second maker achieves production hybrid bonding at memory volumes, or secures equivalent base-die foundry allocation, the concentration argument loses its mechanism regardless of what current share data says.

SK Hynix is the levered, mispriced expression VALID
premises

The recommendation is the weakest conclusion in the thesis, and the composition shows why: it stacks two contested share estimates on top of a mispricing claim the desk cannot substantiate. A low composed number here is the honest read, not a failure of the thesis.

The tightness plugs into the main accelerator artery, giving it a demand anchor VALID
premises

The demand anchor is the best-supported part of the thesis — it rests on physical necessity and an established relationship rather than on estimated shares.

Write-up

Pre-filled skeleton: hbm4-share-concentration.md