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.
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.
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 EASED
† 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)
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)
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 arithmetic
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 expression
† 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 anchor
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
LONG sk-hynix — the cleanest and least-priced expression. Mechanism: hbm3e provided_by sk-hynix@0.62 and hbm4 provided_by sk-hynix@0.55 (but ~70% at NVIDIA), with sk-hynix supplies nvidia@0.6 — concentrated HBM4 share + sold-out-through-2027 + severalfold 2027 price hikes accrue to the yield leader. Valuation: ~6.3x fwd prices a glut that concentration prevents — this is a right-thesis-on-a-cheap-name, the opposite of priced-in. Caveat: it remains a cyclical commodity underneath; the re-rating is a multiple call as much as an earnings call, and demand (AI capex) must hold.
LONG micron — the US call option and #3 qualified name. Mechanism: qualified HBM4, sold out through 2027, ~6x fwd / ~10x FY27; best liquidity for US investors, supply-security premium. But Micron is the marginal share LOSER if concentration holds and is the most exposed to a 2027-28 glut — size below SK Hynix, treat as the liquid proxy not the thesis-purest leg.
RELATIVE: long sk-hynix vs short/underweight samsung — the variant IS that Samsung's HBM4 'catch-up' (mid-20s share) underdelivers vs the consensus convergence to ~30%. Mechanism: competes_with edge; Samsung is further diluted by a sub-scale foundry arm carrying the group multiple. Cleanest pure expression of 'concentration, not diffusion'.
LONG the HBM test/inspection enablers (advantest, teradyne, camtek, onto-innovation) as a lower-beta, share-agnostic expression. Mechanism: hbm4 enables these — HBM4's base-die-on-logic + taller stacks raise test/inspection intensity per GB, so they re-rate on HBM4 UNIT volume regardless of which memory maker wins the split. Flag: hbm4->advantest/teradyne/camtek/onto edges are weight_assumed 0.5 (unresearched) — magnitude uncalibrated.
AVOID expressing via nvidia: concentrated, sold-out HBM4 hands SK Hynix pricing power and raises NVIDIA's memory BOM, but at ~$5T cap NVIDIA is an AI-demand-durability bet (its own capex-cycle beta), not an HBM-tightness bet — the thesis barely moves the needle. Use NVIDIA only as the demand-anchor confirmation (hbm4->nvidia@0.9), not the trade.
SECOND-DERIVATIVE (contaminated): commodity-DRAM names (nanya ~5.2x fwd, winbond) benefit from HBM crowd-out tightening the whole complex — but they carry China/CXMT supply-wildcard risk and lack HBM4 exposure; a spillover play, not the thesis.
Second order
NVIDIA: hbm4 -> nvidia@0.9 (Vera Rubin's gating memory). Concentrated, sold-out HBM4 = supply secured but rising memory BOM and SK Hynix pricing power; NVIDIA accelerator output gated by SK Hynix allocation. Cascade: an hbm4 shock touches 179 downstream entities / ~$44T market cap.
SK Hynix / Micron / Samsung re-rate as the direct beneficiaries — but the SPLIT is the whole trade: SK Hynix > Micron > Samsung in torque; the variant says the market wrongly models convergence.
TSMC captures HBM4 base-die value: the base die runs on TSMC 12nm/3nm logic — a new memory-adjacent foundry revenue stream that deepens the sk-hynix–tsmc coupling and gives TSMC a share of every HBM4 stack.
Powertech (HBM back-end packaging partner of both sk-hynix and micron, partners_with edges) gains volume as HBM units scale.
Third order
Theme re-scoring: memory-supercycle (status: consensus) sharpens from 'cyclical up-cycle' to 'structurally concentrated shortage'. The variant reframes the theme's core debate (how much HBM de-cyclicalizes memory) — concentration, not just demand, is the de-cyclicalizer. Path: hbm4 concentration -> memory-supercycle re-scoring -> sk-hynix/micron multiple re-rate.
AI capex durability cross-feed: concentrated HBM supply makes SK Hynix allocation a SECOND physical ceiling on accelerator output alongside CoWoS. Path: hbm4 -> nvidia -> hyperscalers; reinforces advanced-packaging-bottleneck (HBM was already the 'H' in that theme). Two chokepoints, one owner-set.
Commodity-DRAM crowd-out: HBM wafer intensity (~half of 2027-28 wafer growth, HBM ~23-25% of DRAM wafers) crowds out commodity DRAM, tightening the whole complex. Path: hbm4 -> DRAM wafer allocation -> nanya / winbond / cxmt pricing. Cross-theme into china-indigenization: CXMT is the supply wildcard that could break tightness from below.
Chokepoint migrates upstream into logic: HBM4's base die depends on TSMC leading-edge logic wafers — a NEW dependency coupling HBM supply to foundry capacity (the same node feeding CoWoS). If TSMC base-die allocation is the gate, the memory chokepoint moves one layer into logic. Cross with advanced-packaging-bottleneck.
Sovereign / Stargate demand deepens the sold-out condition: price-insensitive sovereign-ai buyers reserving HBM years ahead extend the shortage horizon. Path: hbm4 -> sovereign-ai-buildout / stargate -> longer sold-out window, reinforcing the variant.
Indicators
name: HBM4 share concentration at NVIDIA · entity: sk-hynix · measure: SK Hynix % of NVIDIA Vera Rubin HBM4 allocation · current: ~60-70% SK Hynix, ~mid-20s Samsung, ~20% Micron (NVIDIA cert June 5 2026; SK Hynix reportedly secured ~70% of Rubin HBM4 orders) · consensus: Share converges toward ~50/30/20 as Samsung and Micron ramp qualified HBM4 supply — the point of all-three qualifying is diversification · variant: SK Hynix holds >=60% of NVIDIA HBM4 through 2027; the base-die-on-logic bar keeps volume concentrated at the yield leader · unit: % of NVIDIA HBM4 volume · source: https://www.semicone.com/article-385.html
name: 2027 HBM contract ASP direction · entity: hbm4 · measure: YoY change in HBM (esp. HBM4) contract price for 2027 supply · current: 2027 negotiations began Q2 2026; TrendForce expects prices to increase severalfold reflecting an undersupplied market · consensus: HBM pricing peaks in 2027 then declines as 2027-28 capacity lands; annual pricing mechanism caps upside and mean-reverts · variant: 2027 contracts print up sharply YoY (severalfold on HBM4); pricing power held through 2027 = no competing supply was added · unit: % YoY contract ASP · source: https://www.trendforce.com/research/download/RP260527UC
name: HBM sold-out horizon / supply-demand balance · entity: hbm3e · measure: How far forward HBM is sold out; supplier gap commentary · current: 2026 fully sold out; SK Hynix & Micron sold out through 2027; SK Hynix CEO: 2027 worst supply year in history, crunch to persist toward 2030 · consensus: Gap narrows as 2027-28 fab capacity (Samsung P5 Fab 1 by Jul 2027, Micron/SK Hynix new fabs) reaches volume · variant: HBM demand stays above supply through 2027; the wafer additions are eaten by HBM intensity, no glut arrives in the window · unit: sold-out horizon (years); qualitative gap · source: https://www.tomshardware.com/pc-components/dram/sk-hynix-says-2027-will-be-the-worst-year-for-memory-shortage-forecasts-crunch-to-last-until-2030-ceo-shares-grim-outlook-on-the-day-sk-hynix-gets-listed-on-nasdaq
name: Memory-maker forward multiple (the consensus tell) · entity: sk-hynix · measure: Forward P/E on SK Hynix / Micron and implied cycle stance · current: SK Hynix ~6.3x fwd, Micron ~6.0x (Micron ~10x FY2027 EPS); PEG ~0.10-0.17 — near-zero growth premium priced · consensus: Trough multiple is correct — 2027 is peak earnings before a cyclical down-leg; memory deserves a commodity multiple · variant: Multiple re-rates upward as tightness proves structural/concentration-driven rather than a demand spike that supply competes away · unit: x forward P/E · source: https://stockanalysis.com/quote/krx/000660/statistics/
name: HBM4 base-die yield / vendor execution gap · entity: hbm4 · measure: Relative HBM4 base-die yield and qualification quality across the three vendors · current: SK Hynix on TSMC 12nm (mainstream) / 3nm (premium) base die with yield lead; Samsung 4nm turnkey ramping; Micron ~20% allocation · consensus: Samsung closes the yield/qualification gap on the newer node; share converges as execution normalizes · variant: Base-die-on-logic keeps yield — and therefore share — concentrated at SK Hynix through 2027; Samsung's catch-up underdelivers vs expectations · unit: qualification/yield milestone by vendor · source: https://www.tomshardware.com/pc-components/dram/hbm-undergoes-major-architectural-shakeup-as-tsmc-and-guc-detail-hbm4-hbm4e-and-c-hbm4e-3nm-base-dies-to-enable-2-5x-performance-boost-with-speeds-of-up-to-12-8gt-s-by-2027
name: HBM wafer crowd-out of commodity DRAM · entity: hbm3e · measure: HBM share of DRAM wafer capacity and its claim on wafer growth · current: HBM ~23-25% of DRAM wafers in 2026; ~half of the +12%/yr 2027-28 memory-wafer growth consumed by HBM; non-HBM DRAM bit growth capped ~15% vs ~22% demand · consensus: Added wafer capacity eventually eases both HBM and commodity DRAM · variant: HBM's larger die and base-die-on-logic absorb the wafer additions and crowd out commodity DRAM, tightening the whole complex, not just HBM · unit: % of DRAM wafers · source: https://www.tomshardware.com/tech-industry/artificial-intelligence/samsung-and-sk-hynix-warn-ai-driven-memory-shortages-could-last-until-2027-and-beyond-as-hbm-demand-explodes-customers-already-reserving-supply-years-ahead-while-the-wider-dram-market-begins-to-tighten
Tripwires
Samsung's NVIDIA HBM4 share crosses 30% and rises — the diffusion consensus is right; cut the sk-hynix-vs-samsung leg and the concentration call.
A 2027 HBM contract print comes in flat-to-down YoY — pricing power is breaking ahead of the capacity wave; re-test whether tightness is real before the multiple re-rate.
TrendForce/SK Hynix show HBM moving off allocation (spot availability appears) before end-2027 — the glut is early; exit the memory longs.
AI capex air-pocket: a hyperscaler HBM order cut or NVIDIA Rubin push-out — demand, not supply, breaks the tightness; the whole memory complex de-rates regardless of concentration.
CXMT or a new entrant qualifies HBM at a Western accelerator maker — concentration breaks from below; re-score china-indigenization cross-feed.
Falsifiers
Samsung and Micron close the HBM4 share gap toward ~30/20 at NVIDIA in 2026-27 — qualification DID spread share, the base-die-on-logic bar was cleared by all, and the concentration mechanism is wrong.
2027 HBM contracts settle flat-to-down YoY and HBM comes off allocation in 2027 — the 2027-28 capacity wave arrived on time, supply competed the tightness away, and the 6x multiples were correctly pricing a glut.
SK Hynix forward multiple stays ~6x through the horizon even as earnings hold — the market never concedes structural, so the re-rate leg (the return, as opposed to the earnings) never pays regardless of who's right on flows.
An AI-capex digestion cuts HBM demand below supply — the tightness was demand-driven all along, not concentration-protected, and every memory long de-rates together.
Reasoning chain
Current-generation HBM (HBM3E) is a severe, SK-Hynix-concentrated bottleneck today; the model's next-gen reading has EASEDVALID
Samsung Q2-26 call: supply constraints expected MORE severe in 2027 than 2026, persisting through 2028.
why
Corroborated by 2026 HBM capacity sold out with makers refusing orders (TrendForce).
HBM4 — supply constraint moderate0.69 moderate
REACTIVE.
why
Same logic one generation on: the constraint reading is a demand claim. HBM4's customer-weighted pull is currently very high (134.6), which is why 'moderate' rather than 'severe' is the contested part. Anchors reproduce the hand-scored 0.7 today.
HBM supply tightness — value severe0.85 strong
40%
SK Hynix 2027 pricing discussions 'progressing smoothly on solid demand'; Samsung DRAM ASP +mid-40% QoQ.
why
Tightness is corroborated by pricing, not just commentary.
SK Hynix — 62% supply share of High-Bandwidth Memory (HBM3E / HBM4)0.60 moderate
62%
UNPRICED DEPENDENCY (probabilities.py flags this as deps=3).
why
The 62% share is a desk estimate, not disclosure — SK Hynix does not break out customer share. Scored low BECAUSE it is load-bearing and unsourced.
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
HBM40.99 strong
The technology exists and is qualified at NVIDIA.
why
Not in dispute.
SK Hynix — 55% supply share of HBM40.60 moderate
60%3x55%
Desk estimate.
why
Samsung's own guidance (HBM4 >60% of its HBM revenue mix, 3x QoQ) is evidence AGAINST a 55% SK Hynix share holding, so this is scored below the HBM3E equivalent.
Samsung Electronics — 25% supply share of HBM40.50 moderate
25%
Most contested number in the thesis.
why
Samsung's Q2-26 guidance implies it is gaining faster than 25% implies. Scored at a coin-flip because the disconfirming evidence is company disclosure and the estimate is not.
HBM4 — signal 2026-07-220.80 strong
Recorded desk signal; medium credibility, trade-press sourced.
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
Samsung/SK Hynix guidance and sold-out 2026 capacity.
SK Hynix — signal 2026-07-220.80 strong
SK Hynix CEO 'worst supply year' commentary; company source but forward-looking opinion, not a disclosure.
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 arithmeticVALID
premises
Hybrid Bonding0.75 strong
HBM5 stack heights would require hybrid bonding, and JEDEC has been RELAXING dimensional specs specifically to avoid forcing it (Raise Summit 2026, with SK Hynix's Hoshik Kim present).
why
A standards body bending physical constraints rather than adopting the obvious technical answer is a revealed statement that memory cannot yet do hybrid bonding at volume. Whoever gets there first sets the stack-height ceiling for everyone.
BE Semiconductor Industries (Besi)0.70 moderate
Bonding capability is an EQUIPMENT question before it is a memory one.
why
BESI, ASMPT and Hanmi supply the tools, and domestic share in bonding is near zero outside them — the same chokepoint bonding-not-litho identifies and Japan export-controlled on 1 August.
The HBM4 logic base die is manufactured on a leading-edge FOUNDRY process, not on a DRAM line.
why
That moves part of HBM4 competitiveness out of the memory makers' own fabs and into foundry allocation, where the queue is set by someone else's customer list.
Conventional DRAM (DDR5 / LPDDR)0.80 strong
3x
HBM bit-density per die is roughly ONE THIRD of standalone DRAM because of the TSVs, so a given HBM capacity consumes ~3x the wafers.
why
Share of HBM is therefore also a claim on general DRAM wafer capacity, which is what makes the consumer squeeze arithmetic rather than pricing.
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 expressionVALID
premises
SK Hynix — 55% supply share of HBM40.60 moderate
Contested desk estimate (see above).
SK Hynix — 62% supply share of High-Bandwidth Memory (HBM3E / HBM4)0.60 moderate
Unsourced, load-bearing.
SK Hynix supplies NVIDIA Corporation1.00 strong
Well established; SK Hynix is NVIDIA's lead HBM supplier.
SK Hynix — forward P/E 6.3x0.48 weak
MISPRICING IS THE WEAKEST LINK
6x
A 6x forward multiple is consistent with the market pricing peak-cycle earnings rather than mispricing durability, and the multiple ALONE cannot separate the two.
why
What can: whether the pull from SK Hynix's customers persists. So this premise is now REACTIVE — p is computed from derived:sk-hynix.demand_pull (customer-revenue-weighted YoY growth), which recomputes every pulse as those customers report. The anchors (25->0.25, 90->0.75) are deliberately set so today's ~55 reading reproduces the desk's hand-scored coin flip: they exist to let FUTURE movement move the number, not to re-rate the thesis today. Caveat, stated because it matters: demand_pull measures the LEVEL of pull, and a high level is itself consistent with a cycle peak. Its DECAY is the informative signal here, not its height.
SK Hynix — signal 2026-07-220.80 strong
Company commentary, forward-looking.
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 anchorVALID
premises
HBM4 supplies NVIDIA Corporation1.00 strong
Physical: Vera Rubin requires HBM4.
SK Hynix supplies NVIDIA Corporation1.00 strong
Established supply relationship.
HBM4 — signal 2026-07-220.80 strong
Desk signal, trade-press sourced.
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.