State of the Memory Supercycle
HBM/DRAM/NAND pricing and the cycle 2nd-derivative, the HBM oligopoly, the CXMT glut trigger, and the through-cycle margin floor — size the rent for the roll, not the wedding.
Size the rent for the roll, not the wedding: own the through-cycle memory winners (MU + WFE toll-booths AMAT/LRCX/KLA, ASML via ADR) at cycle-aware entries; treat the China commodity-DRAM glut (CXMT) as the live falsifier, not a tail.
The HBM oligopoly is real, sold out through 2026, and printing 70%+ memory-division operating margins (SK Hynix 72%, Samsung ~73% est., Micron 74.9% GM Q2-FY26) — the most violent up-leg the industry has seen, with a Goldman-pegged ~4.9% DRAM supply gap, the worst in 15 years. But DRAM/NAND remain the most mean-reverting businesses in tech: the same sold-out-record-price euphoria has historically sat near the cycle top. The durable rent is the WFE toll-booth at the bottom (every memory bit needs AMAT/LRCX/KLA/ASML capacity, gated by 12-month EUV lead times and a €38.8B ASML backlog) and the one US-listed pure memory maker (MU) bought cycle-aware, not the Korea makers at peak multiples — with CXMT's ~350kwspm 2026 DDR5 exit and YMTC NAND as the explicit glut-trigger falsifier to monitor, not dismiss.
Two distinct dominators. On pricing power TODAY: the 3-maker HBM oligopoly — SK Hynix (~62% HBM share, first to High-NA EUV in DRAM), Samsung (~17%), Micron (~21%) — sets memory prices and is capturing all-time-record margins, but this dominance is cyclical and partly an artifact of disciplined capacity restraint (SK Hynix is deliberately slowing HBM4 to keep DRAM tight). On structural durability: the WFE toll-booths (AMAT/LRCX/KLA + ASML's EUV monopoly, ENTG/MKSI materials, Camtek/Onto HBM inspection) — a tighter, more defensible oligopoly than the makers, gated by 12-month+ lead times, a €38.8B ASML backlog, and decades of process IP. The makers' dominance survives only until the next glut; the toll-booths' survives the glut because down-cycle capex cuts hit them later and leading-edge/HBM/High-NA migration keeps WFE intensity rising even as bit growth slows. The most durable rent is the toll-booth (and ASML's EUV monopoly above it), not the maker.
Full-stack value chain
The memory value chain from AI/HBM end-demand down to the wafer-fab equipment and materials that gate supply. The marker shows where the HBM-led supercycle is a tailwind, headwind, or mixed for each layer’s margin pool. Tap any layer for the full read.
The toll-booths that gate memory supply — HBM packaging/TSV, the WFE makers, the 3-maker DRAM oligopoly — defend margin through the cycle, while commodity DRAM and NAND mean-revert. Own the rent; size for the roll.
Shift-point register
Ranked by margin-at-stake × demand-durability × evidence-strength. The flagged rows clear the bar and are promoted to a deep-dive.
Wafer-conversion squeeze
HBM cannibalizes commodity-DRAM wafers — the conversion tax that re-rates the whole memory P&L
An HBM bit consumes ~3x the wafer area of commodity DDR5 (TSV penalty + die-stacking yield loss), and HBM revenue/wafer runs 3-5x DDR5 ($60-100/HBM3E module vs $5-10 equivalent DDR5). As the big-three convert lines to HBM (HBM now ~23% of DRAM wafers per tech-insider.org Q2-2026), the residual commodity-DRAM wafer pool shrinks against rising datacenter+mobile bit demand — the global market flipped from slight oversupply to a projected ~12% DRAM shortage in 2026. The scarcity, not HBM itself, is what drove DDR5 contract +90-95% QoQ in Q1-2026 and a further +58-63% in Q2, and dragged even DDR2 up 55-60%. This is the single mechanism that simultaneously re-prices commodity DRAM AND funds the HBM mix-shift — it sits upstream of every other shift point.
Opportunity board
21 names sorted into three buckets — durable compounders, undervalued / high-potential, and short / avoid. Each tile shows its conviction; open one for the thesis, catalyst, and falsifier.
Durable compounder
8Undervalued / high-potential
11Short / avoid
2Views & the Voices
The strongest bull and the strongest bear case, the US-listed ETF expressions of the theme, and where the tracked QAI Voices roster nets out — each stance stamped trackable vs inferred.
Bull case
the longThis is the first AI-driven memory cycle where the up-leg has a structural floor: HBM is a sold-out, capacity-rationed oligopoly that pulls wafers out of commodity DRAM, so even the 'commodity' leg is in its worst shortage in 15 years. Own the through-cycle winners (Micron, the memory-WFE toll-booths) into a cycle that the supply side cannot relieve until late-2027 fabs ramp.
- Pricing is at a historic extreme AND broadening: TrendForce prints DRAM contract +90-95% QoQ in Q1 2026 and +58-63% in Q2, with NAND accelerating to +70-75% in Q2 — and crucially DDR5 commodity profitability is now forecast to surpass HBM3e, so the up-leg is no longer HBM-only (TrendForce, Q1/Q2 2026).
- Micron printed the thesis: FQ3'26 record $41.5B revenue (+346% YoY) at 84.9% gross margin, DRAM 76% of revenue with ASPs up low-60s% QoQ, HBM4 already >$1B shipped, and guided FQ4 to ~$50B revenue / ~86% GM — margins a commodity business was never 'supposed' to reach (Micron, June 2026).
- The HBM oligopoly is real and rationed: SK Hynix ~62%, Micron ~21%, Samsung ~17% of HBM in Q2'26, all three sold out through 2026, with HBM consuming a rising share of DRAM wafers (~23%) — every HBM wafer is a commodity-DRAM wafer removed, manufacturing the shortage in the rest of the market (presenc.ai/Astute, TrendForce, 2026).
Bear case
the short / avoidDRAM and NAND are the most violently mean-reverting businesses in tech, and the cure for high prices is high prices: at 85% gross margin every maker is incentivized to add capacity, China (CXMT/YMTC) is adding subsidized bits into exactly the commodity segment that is most of the wafers, and you are buying peak earnings on peak margins — the cycle has not been repealed, only delayed, and delay makes the correction more violent.
- China is the live glut trigger, not a tail: CXMT reaches ~17% of global DRAM wafer capacity by 2027 (from 13% in 2025) with bit-share rising 9%->12%, now shipping DDR5-8000/LPDDR5X, while YMTC NAND share has jumped to 13% (tied with SanDisk/Micron) targeting 15% by late-2026 — subsidized, sanctions-indifferent capacity aimed straight at commodity DDR5/LPDDR and NAND (SemiAnalysis/Tom's Hardware/TechNode, 2026).
- You are buying peak-of-peak: Micron at 84.9% GM guiding to ~86% is the definitional cycle top for a business whose mid-cycle GM is ~35-40% and whose trough GM has been negative — every prior DRAM cycle ended with the same 'this time is different' narrative right before margins collapsed (Jim Handy's structural-commodity frame; Micron, 2026).
- HBM is partly cannibalizing the bull case: HBM is converted commodity-DRAM wafers, so when AI HBM demand normalizes those wafers flood back into a commodity market that has meanwhile lost pricing discipline — and HBM4 ASP premiums compress as Samsung/Micron close the gap on SK Hynix and yields improve.
bullish
Dispersion: moderate-to-high
Strong consensus that 2026 is a genuine memory supercycle — the pricing prints (DRAM +90-95% then +58-63% QoQ) and Micron's 84.9% gross margin are not in dispute, and the data houses (TrendForce, Counterpoint) and Street (Rasgon, Hosseini) plus the WFE buy-side (Laffont's 13F) are aligned long. The real dispersion is on DURABILITY and CYCLE POSITION, not direction: the structural-break camp (Aasholm, Castellano, Patel) argues HBM has rewired the cycle into lasting tightness; the cycle-historians (Handy, Irrational Analysis) insist it is still a commodity that will mean-revert violently. The swing variable everyone converges on is CXMT/China — Ray Wang frames it as the bifurcation point (HBM tight vs commodity glut). Net: own the through-cycle winners but treat the China commodity-glut and a hyperscaler-capex air-pocket as the live falsifiers, not tails. The cleanest US-first expression is Micron plus the memory-WFE toll-booths (AMAT/LRCX/KLA), which capture the capex wave with less direct exposure to commodity-bit price reversal.
US-listed ETF expressions
7The QAI Voices
stance · trackable / inferredPrediction matrix
Directional calls across Sep'26 / Dec'26 / Jun'27 / Jun'28, confidence decaying high → low over the horizon. Each cell is the call; click a row for the full reasoning, leading indicator, and falsifier.
As of 26 June 2026 the memory super-cycle is mid-cycle and accelerating, not topping: DRAM contract prices ran +90-95% QoQ in Q1 and +58-63% in Q2, NAND +70-75%, HBM is sold out through 2026 with ~$22B of pre-paid customer deposits, and maker inventories are at multi-year lows (SK Hynix DRAM ~2-3 weeks). The reported peak in fundamentals is roughly 3-9 months out (Sep-Dec 2026 prints will be the violent ones) as the price 2nd-derivative rolls from explosive to merely strong; spot DDR5 is already softening at the margin, the first faint tell. The dangerous horizon is Jun27-Jun28, when the 2027 greenfield fabs (SK Hynix, Samsung, Micron ID1) ramp into a possibly-moderating AI-capex pulse and CXMT's ~350 kwspm of commodity-DRAM lands — the live falsifier. Own the through-cycle structural winners (Micron, the WFE toll-booths, advanced-memory packaging) at cycle-aware entries, fade pure commodity-DRAM beta into 2027, and treat the China glut as the thesis-killer to monitor, not a tail.
Valuation scenarios
Every target is scenario-conditional with a probability; the verify produced zero outright buys. Tap a name for its full bull / base / bear ladder.
Regime calls
The four cross-cutting forces and when each bites across the Sep'26 → Jun'28 horizon. Tap any force or modulating risk to read the full call.
Regime A
Prices still rising QoQ, sold-out HBM, ultra-low inventory (SK Hynix DRAM ~2-3wk), record maker margins. The 1st-derivative is positive and large; the trade is owning makers + WFE toll-booths, buying weakness. Risk is complacency about the 2nd-derivative already rolling.
Regime B
Price LEVELS stay high but QoQ rate-of-change goes flat-to-negative; spot softens before contract; estimates plateau. Stocks de-rate on the rate-of-change even as earnings print records — the classic 'good news, lower stock' phase. Rotate from commodity-DRAM beta toward HBM/packaging/process-control quality.
Regime C
2027 fabs (SK Hynix/Samsung/Micron ID1) ramp into possibly-moderating AI capex while CXMT's ~350 kwspm commodity bits land; DRAM/NAND roll over hard (history: -50%+ in a few quarters). HBM premium compresses but the oligopoly keeps the trough far shallower than 2023. This is the bear case the house-call sizes against.
Regime D
The cycle splits. Makers with the richest HBM mix (SK Hynix, then Micron) and the packaging/inspection toll-booths outperform; pure commodity-DRAM beta and CXMT-exposed low-end suffer. The bull-bear hybrid most consistent with the structural HBM rent thesis.
Regime E
A hyperscaler capex pause or AI-demand wobble hits before new supply even arrives — demand-side glut, the fastest and most brutal version. Inventory rebuilds, prepays get renegotiated, the whole complex de-rates regardless of supply discipline. Low probability near-term given sold-out books, but the highest-severity tail; the falsifier is a contract-price DECLINE printing ahead of schedule.
How they modulate the book
CXMT commodity-DRAM glut (the live falsifier, China — analysis-only)
CXMT is targeting ~350kwpm by end-2026, which would make it close to the third-largest DRAM supplier by wafer capacity, with ~60% of capacity on DDR5/LPDDR5 and the rest LPDDR4X, and near-total exposure to commodity DRAM (Tom's Hardware / SemiAnalysis / Digitimes, 2025–26). It ships DDR5-8000 and LPDDR5X-10667 despite US export curbs. This is the single biggest threat to the whole long side: a commodity-DRAM oversupply collapses the DDR5 price leg that is currently more profitable than HBM3e, and the Big-3's 3:1 HBM/DDR5 conversion logic only holds while commodity DRAM stays tight. China-listed — NEVER a recommendation per geography policy; it lives here as the cross-chain falsifier that the entire board is sized against. Pairs with YMTC on NAND (China NAND-supply response is the symmetric NAND-leg glut risk).
Late-cycle top / month-30 upcycle exhaustion
Weighted cycle peak clusters ~Q1 2027 (CI H2'26–H2'27); TechInsights forecasts a 2027 memory downturn; no prior memory upcycle exceeded ~30 months and this one is at that threshold (useLuminix / blocksandfiles, 2026). The bull base case (Micron/SK Hynix: upcycle through 2028, 'the cycle has changed' on HBM-contract-smoothed FCF) is unproven against history. Every name on this board — makers, WFE, packaging, test — de-rates together if the peak arrives on schedule. The board's discipline (cycle-aware entries, buy-the-dislocation, avoid the FQ4 peak chase) is the direct hedge; the falsifier is symmetric — a genuinely amplitude-shifted longer cycle would punish the under-exposed. The June-2026 Korea national-fab plan (Samsung+SK Hynix, 2x capacity in 5yr) is a structural supply accelerant landing into exactly this 2027 downturn window.
HBM/AI-capex digestion air-pocket
HBM is the demand engine: 2026 HBM is sold out with ~$22B Micron prepayments, but the entire long side assumes the hyperscaler AI-capex pull continues. A digestion pause in datacenter/GPU capex (the demand driver, not a board layer) would (a) soften HBM directly, and (b) — per memory-maker warnings — push HBM wafer capacity back into conventional DRAM, manufacturing a commodity glut from the supply side even without CXMT. The 3:1 conversion ratio is a double-edged sword. This is the macro-context risk that links the AI-compute trade to the memory cycle; it hits HBM-pure names (Camtek, SK Hynix ADR) hardest and the diversified toll-booths (AMAT, KLA) least.
2027 WFE capex cliff (toll-booth concentration)
The WFE toll-booth thesis (~$135B cal-2026; LRCX ~1/3 memory, DRAM record 23% of systems; Micron capex +45% to ~$20B; SK Hynix $8B EUV commitment) is itself the cyclical variable. WFE bookings turn before memory prices — makers cut tool orders at the first sign of a glut, so AMAT/LRCX/KLA/ASML and the small-caps (MKSI/ACLS/FORM/ONTO/CAMT) can roll over ahead of the commodity. The most memory-concentrated and most-levered names (Axcelis, MKS, Camtek) lead the de-rate; diversification (AMAT, KLA, SOXX) is the only structural cushion. Owning toll-booths does not escape the cycle — it changes its shape.
Export-control + Korea/Taiwan supply concentration
The supply side is geographically concentrated and policy-exposed: HBM and leading-edge DRAM sit in Korea (SK Hynix, Samsung) and Taiwan-packaged, gated by a single EUV vendor (ASML, Netherlands). US export curbs already shape CXMT/YMTC's trajectory and could tighten or loosen either way; a curb-relaxation accelerates the China glut, a curb-tightening can disrupt the incumbents' China-fab output. The SK Hynix Nasdaq ADR (Jul 10 2026) creates a US-listed expression but also concentrates Korea-macro/FX/governance risk (SK Hynix+Samsung >40% of KOSPI) into the book. Geography policy keeps recommendations US-first with thin Korea/EWY weight precisely to bound this; the analysis of CXMT/YMTC/Samsung/SK Hynix is unrestricted but the dollars are not. Korea doubled down: in June 2026 Seoul unveiled an ~800T-won (~$518B) national semiconductor plan — Samsung + SK Hynix each building two new fabs, HBM packaging in Chungcheong, approvals streamlined to pull fab construction forward ~12yr and aiming to DOUBLE memory capacity within 5yr (Tom's Hardware / CNBC / CNN, Jun-2026). A 'Korean Stargate' for memory: it deepens the Korea concentration and front-loads the very capacity a demand air-pocket would turn into a glut — supply-side confirmation that cuts both ways.
Premise pressure-test
The six named 2026-Q3 catalysts the thesis rests on, probability-weighted. Click any premise for the if-true / if-false split.
AI/HBM demand pull stays structural, not a bubble, through 2027
holdingThe HBM oligopoly (SK Hynix/Samsung/Micron) holds — high barriers, rational pricing
holdingCommodity DRAM/NAND remain violently mean-reverting — this is NOT a permanent shortage
holdingCXMT stays a commodity/legacy-DRAM glut trigger, not a leading-edge or HBM threat, near-term
at-riskWFE toll-booths collect regardless of which memory maker wins — capex funds the buildout
holdingWe are mid-cycle (Jun 2026), with the fundamental peak 3-9 months out, not behind us
holdingHBM packaging intensity (TSV → hybrid bonding) is a secular content step-up, not just cyclical
unprovenUS export controls on advanced memory tooling stay in force and effective through the horizon
at-risk