SK hynix makes the memory every AI chip needs and can’t get enough of. It is sold out through 2027, its CEO says the shortage runs to 2030, and the shares cost under 3× next year’s earnings while the AI peer set trades at 46–58×. Our path: ₩5.5M end-2026 → ₩8.0M end-2027 → ₩8.4M end-2028, from ₩1.65M today. What could hurt it: the 2029–31 capacity wave, Samsung taking share (50/33/18 last quarter), and AI spending stalling. Verdict GREEN 8.3/10. Everything below is the evidence.
THE THESIS IN 60 SECONDS · 5 SEP 2026
WHAT WE OWN
The bottleneck of AI: 50% of HBM revenue (Q2; 58% Q1) with Samsung at 33%, 50–67% of NVIDIA’s Rubin HBM4 (see L2 watch) — sold out through 2027, CEO says short through 2030.
PRICE vs PROFIT
₩1.65M = 2.8× next year’s clean earnings. Peers trade at 46–58×. Entry ₩986K (+67%).
Buyer: short through 2028 · Supplier CEO: through 2030 · Micron: 50% gap · contracts now 3–5 years · SpaceX +10GW outside the $1.3T.
THE PATH (year-end)
₩5.5M → 8.0 → 8.4 → 7.9 → 8.5 (2026→30) on 9.5–10.6× next-year EPS. 12-mo: E[V] ₩3.92M, fan median ₩5.12M.
WHAT THE BET ACTUALLY IS
Same ~10× multiple as the street. They price a 2027 plateau (~₩300T run-rate held flat; consensus FY27 EPS ~₩326K vs our ₩583K) → PTs ₩3.0–3.2M. We price growth with the customer (+70% guided) → ₩5.5–8.0M. That is the whole disagreement.
WHAT BREAKS IT
2029–31 capacity wave (D&A peak) · Samsung share (Q2 print: 50/33/18 — pricing-power nibble now, volume a 2029 problem) · circular financing / token prices · rates (10Y >4.90% re-opens the multiple). Worst 1-in-20: ₩2.15M. Disruption check (4 Sep): no chip or memory development in silicon reduces memory demand — candidates shift form factor or value split, and we supply every form; the 2028+ watch is HBM value capture. Checked 2 Sep: the "NVIDIA got favorable pricing" read is an estimate; the record is floors without ceilings (Q2 ASP +30%). Watch, not wound.
Green band = each year’s range. The 2029 dip is the capacity-wave pause (multiple 10.2× → 8.8×), now under review after the CEO’s through-2030 statement. Full derivations → Forecast · signals → Thesis · demand ledger → Demand.
CHART, MOVING AVERAGES, OPTIONS · 5 SEP 2026
WHAT THE CHART SAYS, SIMPLY
The stock is building a base after a −49% correction (₩2.92M on 22 Jun → ₩1.50M on 26 Aug) that never closed below the 200-day average — the long-term uptrend is intact. Price sits above the 20-day (turning up) and 26% above the 200-day, but 11% below the falling 50-day at ~₩1.84M. The 20/50 bearish cross of 24 Jul is unresolved. The reversal completes on a close above the 50-day, which is converging with the ₩1.744M regime gate — so the signals stack in the ₩1.74–1.84M zone. A break below ₩1.558M points to the 200-day at ₩1.31M as the next test. RSI 50 = neutral; realised volatility 74% = big moves either way.
OPTIONS (SKHY ADR, CBOE) — WHAT THEY MEAN
18 Sep monthly: 116K calls vs 92K puts, and 45K calls parked at the $200 strike (+14% from $176). Above ~$190 the sellers of those calls must buy to hedge — a squeeze trigger; below it, the strike acts as a lid. Max pain sits at $157.5, so expiry gravity pulls lower unless the $200 wall is reached. Weeklies are put-heavy (put/call 1.8–4.0) and 23.6K deep puts at $105 expire 25 Sep — tail hedges: the holders are protected, not naked-long, which means less forced selling on bad days. ADR premium ~45% over Seoul ($176 × 10 × ₩1,356 ≈ ₩2.39M vs ₩1.647M): US demand is outrunning Korean; the line to watch is whether the premium compresses or Seoul catches up.
LIKELY NEXT MOVES, SIMPLY · next 4–6 weeks, into the 27 Oct print · chart-structure odds, not engine outputs
Base case (~50%): a range, ₩1.56–1.84M. The stock chops between tested support and the falling 50-day while the buyback absorbs supply; Seoul catches up to Friday’s US move on Monday, then the ADR drifts toward its $157 max-pain into the 18 Sep expiry. Boring is bullish here — a base that holds above the 200-day is how the next leg starts. Upside break (~30%): a close above ₩1.744M and the 50-day. Likely triggers: another realised-price data point (HBM export values), Fed relief, or the ADR testing its $200 call wall. First target the July shelf at ₩2.0–2.1M, then the ₩2.3M high-volume zone. The ₩5.5M year-end path needs the earnings to print, not the chart — but this is how the road to it opens. Downside test (~20%): a close below ₩1.558M. Likely triggers: a stronger won plus a Fed hike, or a bad Samsung-share headline. Target the 200-day at ₩1.31M — which is also where the bear-branch floor sits. The buyback (through 19 Nov) and the tail hedges in the options argue against a crash; a dip there is a buy-zone in this model, not a thesis break.
THINGS TO LOOK OUT FOR · tracked and updated every session
Levels: ₩1.744M regime gate · 50-day ~₩1.84M (falling toward price) · ₩1.558M support · 200-day ₩1.31M · ADR $190–200 (gamma) and $157 (max pain). Dates: Mon 7 Sep Korea alone (US holiday) · Tue 8 Sep both tapes — first clean read on the ~45% ADR premium · Fed meeting (Sept hike odds ~59%) · 18 Sep SKHY monthly expiry · 24 Sep Trump–Xi (tariff leverage) · 25 Sep $105 put expiry · 27 Oct Q3 (est.) — the day the numbers move. Signals: foreign net flows (Friday +₩527B) · buyback pace vs 19 Nov · KRW/USD (1,346 — every 1% of won strength ≈ 1% off reported KRW profit) · monthly HBM export unit value (last $76.13) · server DRAM contract prices · Counterpoint Q3 HBM share (~Dec) · any close above the 50-day = reversal confirmed.
Chart log (newest first): 5 Sep — base above the 200-day; MA20 ₩1.61M turning up; MA50 ₩1.84M falling; 20/50 cross unresolved; RSI 50; vol 74%; SKHY 45K calls at $200, max pain $157.5; ADR premium ~45%; verdict on the chart: constructive base, reversal unconfirmed.
Forecast overlay (full chart file): the modelled path to ₩5.5M by end-2026 with its ₩4.6–6.1M range, the street’s ₩3.21M, and the 12-month instruments (E[V] ₩3.92M · fan median ₩5.12M · worst 1-in-20 ₩2.15M). The chart is a map of where price is versus where the numbers say it should be; the numbers move on 27 Oct, the chart moves every day.
27 AUG — THE CUSTOMER REPORTED. THE BIG PICTURE, IN PLAIN WORDS:
1 · The buyer is growing faster, not slower. NVIDIA growth accelerated a fourth straight quarter (+106%, $96B/qtr) and it guided two years ahead at +70% — a company signals that only when the order book already exists. 2 · The pool of buyers is widening. Non-hyperscale customers (+138%/yr) are on pace to overtake Big Tech within quarters — demand no longer hangs on four boardrooms. 3 · Our price rises are being paid, not fought. NVIDIA margins absorbed "memory cost headwinds" and it raised system prices ~15% — memory is the toll bridge on the only road. 4 · Custom memory arrived early — as our product. NVHBM puts NVIDIA-spec memory, built by memory suppliers, inside Amazon Trainium. The escape route the bears drew now runs through our fab. 5 · What changes here: demand envelope $1T → ~$1.2T · 2027 top-5 capex $1.3T (18% above street) · Rubin HBM4 shipping since early August · verdict 8.2 → 8.4. Every figure with its falsifier → the Demand page. 6 · A fourth mega-buyer, and the reframe (28 Aug). SpaceX targets ~10GW in 2027 (~$300–500B capex, SemiAnalysis) — outside the top-5 $1.3T, NVIDIA-exclusive, ≈1.5 exabytes of HBM at rack density; Microsoft signed 10GW YTD. Inference earns >$100B/GW/yr against ~$12B/GW/yr of cost — demand funds itself. The frame: we are a play on AI memory-bandwidth demand itself, not a derivative NVIDIA trade — Rubin, Jalapeño, TPU and Trainium all buy it. 7 · Company side (28 Aug). Ground broken on the $4B Indiana HBM-packaging + R&D complex — Washington co-funds ~24% ($958M CHIPS package); ops 2H 2028, HBM4E volume Q3 2029. The tightest bottleneck in AI memory now gets a US address next to the customer (cleanroom Oct 2028, first Made-in-USA HBM4E from Q3 2029). No near-term supply relief — discipline intact. And at the ceremony the CEO put the window on the record: "We expect the supply shortage to continue until the end of 2030" — with Micron guiding a 50% demand-supply gap. Three principals now bracket scarcity: buyer through 2028, biggest supplier through 2030, Micron at a 50% gap — and the contracts are going long: 3-5-year industry tenors, NVIDIA multi-year DRAM+HBM agreements reported (Edgewater). The 2029-pause review is armed for Oct 27; its first trigger (LTA tenors) fired 28 Aug via secondary sourcing — logged, held for primary confirmation. 8 · Trim reversed — it misapplied the warning (28 Aug). Patel’s compression argument targets what trades at 30-58×; his crash-endpoint ("everything at 2-3×") is our entry multiple. His exact scenario was already the engine’s largest bear branch (10% weight, ₩2.6-3.9M) — the central-case haircut counted it twice. His rate guess was self-flagged "highly speculative" and has since faded (10Y 4.74→4.66%), while the buyer warns shortages persist through 2028. Multiples restored to 9.5 / 10.6 / 10.2 / 8.8 / 9.0 → path ₩5.5M → 8.0 → 8.4 → 7.9 → 8.5 (+234% / +386% / +410% / +380% / +416% from the ₩1,647K close). Both decisions stand in the audit trail — same process, updated information. Trigger: Nancy’s challenge (unease record 3/3). Narrative-table auxiliary columns and the 12-mo scenario-bar values (a 4 Aug engine run — percentages re-based to the ₩1,647K close) are retained as found; both get a full re-run on Q3 actuals, 27 Oct.
SK Hynix Command Deck
One position, watched like a hawk. 100% SK Hynix via Frankfurt GDR (HY9H.F). Updated Sep 5, 2026 — central case; anchor ₩1,647K, the 4 Sep close (+3.2%, the share print absorbed with foreigners the top buyer). Verdict GREEN 8.3, held. Verdict GREEN 8.3 · showing our central estimate, with 45% of probability mass held in seven bear scenarios and every tripwire live.
The conclusion — what all the evidence says, in plain words
DATA BLOCKPrice ₩1,647,000As of Fri 4 Sep close · +3.2%(first session on the Q2 share print — absorbed; foreigners #1 net buyer ₩527B)Verdict GREEN 8.3(Model B)Central case · 44% bear-regime mass · tripwires live
OUR BEST ESTIMATE — WHAT WE THINK ACTUALLY HAPPENS
The world ordered more AI memory than can be built. SK Hynix makes most of it — and costs 2.8× next year's profit.
YEAR
REVENUE
PROFIT / SHARE
SHARE PRICE
WHY
Today
—
—
₩1.647M
The market still prices this like a commodity chipmaker
2026
₩360T
₩270K
₩5.5M
+226%
H1 already banked ₩132T. Q3 prices +20%, Q4 again. Shortage is printed, not forecast.
2027
₩816T
₩583K
₩7.0M
+361%
Already sold out at published prices: HBM4 at $32-40/GB vs $15.5 today. SK removed its price cap — Micron kept theirs.
2028
₩1,102T
₩753K
₩7.4M
+388%
Peak tightness. No rival fab ships a chip before Dec 2028. Robots and cars arrive on top.
2029
—
₩828K
₩6.9M
pause
New fabs finally land. Profits still rise — the multiple pauses, not the earnings. Contract floors stop a crash.
TWELVE MONTHS — 500,000 SIMULATIONS
Likeliest ₩5.12M+211%·57% chance of at least doubling·worst 1-in-20 ₩2.15M — still above today·3% chance of a loss
WHY IT'S THIS CHEAP, AND WHAT WOULD STOP US
Why cheap: memory has always been priced as a boom-bust commodity. This time 60-70% of the book sits under 5-year contracts with price floors, buyers prepay, and SK removed the ceiling that used to cap the upside. Intel's CEO just publicly abandoned the "memory is a commodity" view. The discount rests on a belief that is dying.
What would stop us: Q3 revenue under ₩98T (late Oct) · October HBM capture under 55% · early-2027 market share under 34%. Seven bear scenarios carry 44% of the probability — China flooding cheap DRAM, Samsung catching up, a credit freeze, an inventory bust.
A detailed external critique of this site landed on 15 Aug. Five points were valid. All five were applied, and they moved the numbers down.
1 · The revenue forecast was arithmetically impossible
H1 actuals were ₩131.9T. Our old ₩380T needed H2 at ₩248T — every quarter +56% above Q2. Rebuilt bottom-up: FY26 ₩359T, FY27 ₩816T, FY28 ₩1,102T.
2 · Our own tripwire was set below our own forecast
The path needs Q3 ≈ ₩102T; the demotion trigger sat at ₩90T. A rule that fires 30% below the requirement governs nothing. Raised to ₩98T.
3 · The probability engine could not be made bearish
100% survival meant the "hostile" cases were all still inside one shortage frame. Rebuilt with six structurally different bear regimes — share loss, decode substitution, multiple compression, inventory correction, Korea/FX, model-error — now 38.5% of probability mass, up from 9.3%.
4 · Retiring a tripwire after it fired broke governance
Correct, and the fix is theirs: a retired rule is now preserved and scored out of sample rather than deleted, so governance changes leave a trail.
5 · Web traffic is not token demand
Cloudflare's figure includes search crawlers, uptime checkers and ad networks. It is demoted to supporting context, not a revenue input, until the full conversion chain (agents × actions × calls × tokens × model size × batching × cache reuse × quantisation × memory tier) is built.
Net effect on the headline: median 12-month outcome ₩5.12M → ₩5.12M, probability of at least doubling 91% → 66%. The case is still strong. It is no longer overstated.
📊 THE WHOLE THESIS IN PICTURES
HOW THE TAILWINDS STACK INTO THE PRICE (12-month build, ₩M)
TAILWIND METER — what each force means, and how hard it pushes
🔒 Supply fixed to 2029
orders <70% filled · a rival fab needs 400 EUV tools from a world building 60/yr → sellers set price
🌊 Flood pricing live
two desks print Q3 +20% · Apple implies +125-250% · market ×4 → every quarter re-prices UP
🤝 Nvidia lock, prepaid
70% of HBM4 through 2030 with advances → 2027 revenue is contracted, not hoped
🤖 Robot wave loading
humanoids +272%/yr · AVs 300GB each · fleet training hits HBM 2027 → second demand wave before the first is fed
💵 Buyers get richer
Goldman: hyperscaler cash flow at records by 2029 → rising prices stay payable
💰 Cash returns near
returns framework due Aug–end-Sep · net cash is ₩69.4T printed (cash ₩88.0T − debt ₩18.6T), not yet the ₩100T marker → buyback bid real but smaller than headlines suggest
🏦 Compute is now an asset class
Nvidia + Apollo, Blackstone, BlackRock, Brookfield, Goldman, KKR mobilising $500B; SEC waived Dodd-Frank risk-retention → funding pool is now structured credit, not balance sheets
🏭 Rivals concede the point
Intel's CEO drops "memory is a commodity" · says no relief till 2028 → the discount's premise is collapsing
⚠️ HBM per task may fall
Etched ships 144GB of HBM per chip · Taalas fits only an 8B model · Samsung's zHBM is still DRAM, phase 3 of a roadmap · Intel's XBM/ZAM land 2029-2030+ → every bypass still carries DRAM
📈 Index flows queued
ADR (2nd-largest US listing ever) → Nasdaq-100 window Dec → passive money must buy
3.9% below today5.4% under 2×56.3% land at 2-3×34.7% more than triple→ 60% AT LEAST DOUBLE
THE SCENARIO LADDER — every rung above today
Street bear (BNK)
₩1.48M−10%
Skeptic floor (6-8×)
₩3.3M+91%
Fallback (Oct <55%)
₩3.6M+119%
Flood-weighted 12-mo
₩5.2M +216%
24 months
₩7.0M+325%
Musk path
₩5.3-7.0M +231-338%
🔬 WHY NEW SUPPLY CANNOT ARRIVE EARLY — the EUV machine count
And the most advanced tool just slipped: Samsung and TSMC both pushed High-NA EUV to ~2030 — Samsung says it is "not yet mature enough". Both will squeeze existing tools with multi-patterning, which means more passes per wafer and slower real capacity growth. Important limit on this argument: it gates leading-edge supply. Much DRAM is built with DUV immersion + multi-patterning, and China now makes its own DUV (5 units 2026, 20 in 2027, CXMT among the first customers). The China channel is not gated by ASML's order book — which is why a dedicated China bear regime now carries 7% of the probability mass.
⚠️ WHAT THIS CHART DOES NOT SAY — READ BEFORE CONCLUDING "NO CRASH EVER"
The machine-count wall stops newcomers — Terafab, Intel, anyone starting from zero. It does not stop the 2029-31 wave, because that wave was never going to come from newcomers. It comes from four things already in motion:
🏗️ INCUMBENT FABS
Yongin (Feb-27), M15X, M17, Samsung P4/P5, Micron — funded, building, and already at the front of ASML's queue. Korea's programme adds $585B more.
📐 NODE SHRINKS
1a → 1c → 1d yields 15-25% more bits per wafer with no new cleanroom. Historically the main source of gluts — needs zero construction.
🔄 WAFER RE-ALLOCATION
HBM eats ~3× the wafers of DDR5. If HBM growth merely slows, those wafers flood back into conventional DRAM within months. Fastest glut mechanism there is.
📉 DEPRECIATION
Today's capex becomes tomorrow's cost: D&A rises from ~₩31T (2026) toward ~₩120T (2030) — it eats margins even if prices hold.
This is precisely why the path shows 2029 pausing at ₩7.9M rather than collapsing — floors and robot demand soften it, incumbent capacity and depreciation still bend it. The EUV wall protects 2026-28 beautifully. It does not cancel 2029-31. Now bracketed by the principals themselves: the CEO says the shortage runs through end-2030 ("no clear signs of a downturn"), the buyer says through 2028, Micron sees a 50% gap — if corroborated, the 2029 step-down is early; review armed for the Oct 27 re-run.
🏭 THE FAB CALENDAR — when new supply actually arrives (not when buildings open)
M15X, Cheongju
Shipping now — wafers deploying, HBM ramp
This is the supply inside our 2026-27 numbers
Yongin Y1
Feb 2027 = first cleanroom · equipment from Q2 2027 · six phases
360,000 wafers/mo by H1 2030
M17, Cheongju
First cleanroom Dec 2028 (NAND)
₩19.1T approved 7 Aug
Yongin Y2
First cleanroom June 2029 (DRAM + HBM)
₩35.2T approved 7 Aug
Two readings, both true. Conviction: SK just committed $38B to plants that cannot ship before December 2028 — a company only does that when it expects the shortage to outlast them. Caution: that same capacity is the 2029 wave, which is why the path plateaus rather than climbing that year. New fabs do not raise 2027 — they confirm 2029.
🚪 SELL-WINDOW SCORE — the half of the model that was missing
Everything above answers "what is it worth." Nothing answered "when has the market already paid for it." For a position with no exit rule, that is the bigger hole. This closes it.
COMPONENT
NOW
READING
Memory cycle deteriorating
0 / 2
Micron: "no end in sight". Samsung adds 1.3% capacity.
Valuation at a premium
0.5 / 2
3.0× forward, 87% below peers, and still under the July high.
Earnings estimates rolling over
0 / 2
Opposite — BofA resumed at ₩3.0M, UBS at $204.
Price/volume distribution
1.0 / 2
₩1,744,000 rejected twice in four sessions — a real failed-auction signal, but at a low level, not a high. Half credit.
Positioning & flows
0.5 / 1
Margin credit +20.9%; ~20% of the Citadel book still to clear.
AI demand deteriorating
0 / 1
Opposite — capex past $1tn, Nvidia raising prices 17%.
What the top will actually look like — and it is not bad news. Good news + rising price = hold. Fantastic news + price refuses to rise = pay attention.Fantastic news + failed breakout + estimates flatten + supply catches up = the sell window. Semiconductors peak while earnings are still climbing, because the market starts pricing the rate of improvement falling.
First watch zone: late November 2026 → H1 2027. Not a forecast of the top — it is when the buyback ends (19 Nov), 2027 HBM pricing is known rather than negotiated, and deceleration becomes something the market can price. Honest caveat: an independent backtest of the mechanical version of this method showed a 32% win rate and losses in five of nine years. It is regime-dependent. This is a watch layer over the fundamental model, not a trading system.
📏 THE MULTIPLE LADDER — what we assume, and why it is not aggressive
The forecast assumes SK re-rates from 3.0× forward earnings today to a peak of 10.6×. Here is every comparable, so the assumption can be argued with rather than taken on trust.
SK hynix today, forward on FY27
3.0×
the market's answer right now
Our end-2026 assumption
9.5×
Our peak assumption (end-2027)
10.6×
Micron's historical ceiling
Micron, trailing, today
18.6×
75% above our peak
Western Digital · SanDisk
32.6× · 41.5×
US semiconductor industry · Seagate
46× · 61.7×
a hard-drive maker at 6× our peak
Even at our peak assumption, SK trades at a 77% discount to Micron's multiple today. And the whole path is 4.9× from here — not the "10×" Patel warns memory stocks won't repeat. His "everything at 2 or 3× multiples" describes the broad market coming down toward where SK already sits at 3.0×.
A correction to our own method, recorded. We briefly trimmed these multiples ~6% for the higher-rate argument, then reversed it. The reason: the risk was already carried twice — the 10% "multiple compresses" scenario implies 3.5-5.2× on FY28 earnings, which is that risk, and the 30-year Treasury hit a 19-year high on 18 August, so the current rate regime is already inside the ₩1,730,000 anchor. Cutting the forward multiple as well counted the same macro three times.
📡 OPTIONALITY TRACKER — carried at zero, watched for the moment they stop being free
Everything below contributes ₩0 to the forecast. Each row has a specific trigger that would move it into the numbers — because optionality you do not track is just a story you tell yourself.
OPPORTUNITY
STATUS NOW
TRIGGER TO ADD IT
LPDDR6-PIM
processing-in-module
Scheduled 2028. No named customer. Described as "specialised applications"
A named design win — any hyperscaler or accelerator vendor committing volume
CXL pooled memory
CMM / memory-as-a-resource
Demonstrated publicly; no disclosed volume
A hyperscaler deploying CXL pooling at rack scale in production
High Bandwidth Flash
with SanDisk
First spec unveiled at FMS 2026. Not before 2030 — needs new media and an agreed standard
A finalised JEDEC-style standard plus a first accelerator adopting it
MRAM · RRAM
SK holds leading patents and prototypes (TrendForce). No product, no customer
A product announcement with a shipping date — or Terafab licensing SK IP
Co-packaged optics
photonic interposer
Roadmap published in Nature Electronics (20 Aug). Research, not product
A named co-development partner — an accelerator vendor building to this architecture
Orbital memory
First Starmind satellite Q4 2027. Supplier unnamed
SpaceX naming an HBM supplier, or disclosing memory per orbital rack
3D DRAM
On the 2029-31 roadmap. This is defensive — it is how DRAM keeps scaling, not new revenue
Treated as a requirement to stay in business, not an opportunity. No trigger.
Allocation power
terms, not products
Already visible — no price caps, prepayments, take-or-pay. The realistic version of "memory as a service"
Equity or revenue-share terms appearing in a supply agreement — the Nvidia playbook, one layer upstream
Two of these are closer than they look.Allocation power is already happening — the removed price cap, the prepayments and the take-or-pay structures are SK monetising scarcity as terms rather than products. And custom HBM4E has already crossed over: named buyers in Nvidia, Microsoft and Broadcom, samples shipped 18 June. It is no longer optionality — it sits inside the HBM assumptions, which is why it is not on this list.
The rule this table enforces: nothing moves from ₩0 into the forecast on a roadmap slide, a research paper or a partnership announcement. It moves on a named customer, a shipping date, or a disclosed volume. That is the same standard that kept the 900,000-wafer Stargate figure out of the model — and it is the standard that will keep this page honest when one of these finally does start earning.
🎯 THE SHAPE OF THIS THESIS — worth stating once
None of the exciting parts are load-bearing. Processing-in-memory, the orbital hierarchy, custom silicon optionality, a memory-as-a-service future — all of it is carried at zero in the numbers. The case rests on the dull part: capacity sold out through 2027, contracted prices roughly doubling, and 3.0× forward earnings.When a thesis needs its moonshots, that is the warning sign. This one treats them as free extras.
And the market is asking the wrong question about it. Everyone is debating whether the AI boom is "real". For SK specifically that is close to irrelevant — the 2027 revenue is contracted, prepaid and floored. The question that actually matters is narrower and harder: do the buyers stay solvent through 2028? That is why the financing tripwire above earns its place, and why it is the one bear thread worth genuine respect.
One contrarian note to hold onto: the best outcome from here may be a boring six months. Every melt-up pulls the re-rate forward but also pulls the rate spiral closer. Earnings compounding quietly at 3× forward while nobody is watching is how the large wins are actually built.
🏦 FINANCING STRESS TRIPWIRE — the gap in the warning system
Every tripwire on this page was operational — Q3 revenue, HBM capture, market share. All of them measure whether SK is selling. None measured whether the buyers can still pay. That was the hole, and it is the one place the sharpest bear argument actually lives.
TRIGGER
WHY IT MATTERS
Hyperscaler IG spreads widen >75bp
The cheapest borrowers getting dearer is the first tremor
Data-centre ABS spreads widen >100bp
Currently near year-to-date tights while corporates widened — if that breaks, the securitisation channel is transmitting stress
Any AI-related bond deal >$1B pulled or failed
Issuance windows shutting is how a credit accident starts
A neocloud covenant breach or missed payment
The second tier breaks before the giants do — CoreWeave's GPU-backed facility runs at ~11%
TLT breaks below $80
The long end repricing is the mechanism behind the whole rate argument
And here is the response, written down now rather than during. A financing wobble hits the share price long before it hits a single wafer order — because SK's 2027 revenue is contracted, prepaid, and floored. So if these wires fire, the question is not "is the thesis broken" but "has anything changed in the order book?"
Price stress and order stress are different events. Only the second one changes the model. The expected outcome of a credit wobble is a frightening drawdown in a thesis that is still entirely intact — and knowing that in advance is the difference between selling the bottom and buying it.
Why the asymmetry makes this survivable: if the rate world arrives, the multiple compresses but earnings still triple — the position "only" doubles. If it does not, you get the earnings and the re-rate. When the bear case is a double, the sizing has already done its job.
⚠️ TWO CHALLENGES TO THIS PAGE, ANSWERED
1 · "₩816T is about $580B — roughly 3-4× TSMC's entire revenue. Can that be right?" It sounds impossible, and the comparison is the problem. TSMC sells fabrication services; SK sells the finished product. An Nvidia GPU retails near $30-40K while TSMC receives perhaps $3-5K of that for making it. So SK's revenue line is not comparable to TSMC's — Nvidia's is. The check that actually matters is market share: $589B against a memory market of roughly $1.98T in 2027 = 30% share, against SK's ~36% of DRAM and 57% of HBM. It passes — but it rests on the memory market roughly doubling next year, which is the single load-bearing assumption on this page. That is exactly why the ₩98T Q3 tripwire outranks every opinion, including ours.
2 · "The earnings leg survives Dylan's rate world. The re-rate leg is where it hurts." Correct, and worth showing rather than arguing. Below is the whole path re-run at compressed multiples — not our central case, a named scenario:
YEAR END
CENTRAL (9.5-10.6×)
RATE-COMPRESSED (7×)
SEVERE (5×)
2026
₩5.5M
₩4.1M
₩2.9M
2027
₩8.0M
₩5.3M
₩3.8M
2028
₩8.4M
₩5.8M
₩4.1M
Gain from ₩1.73M
+388%
+235%
+137%
Even at a 5× multiple — below anything memory has traded at in an up-cycle, and below Micron's 18.6× today — the position still more than doubles. That is the point worth holding: the thesis survives on earnings alone. The re-rate is the upside, not the case.
And the migration insight, which deserves recording. Terafab, OpenAI's Jalapeño, Anthropic on TPUs, orbital compute — every one diversifies who buys memory, and not one of them makes memory. Musk can build a fab for logic; nobody is attempting DRAM, because the EUV and TSV walls apply to everyone equally. Custom silicon is a war among SK's customers — and the arms dealer does well out of wars.
📊 WORLD AI COMPUTE, CONVERTED INTO WAFERS
World AI compute runs about 65 GW in 2025, then adds +30, +50 and +70 GW — reaching roughly 215 GW by 2028. At 170,000 DRAM wafers per gigawatt, here is what that means for the world's fabs.
YEAR
NEW COMPUTE
DRAM WAFERS
SHARE OF WORLD CAPACITY
2026
+30 GW
5.1M
22%
2027
+50 GW
8.5M
37%
2028
+70 GW
11.9M
52%
By 2028 the incremental AI build alone requires 52% of every DRAM wafer on Earth — before a single phone, laptop, car or enterprise server is supplied. That is the physical case for the shortage running through 2028, drawn from the compute forecast rather than asserted.
💡 THE VALUE-CAPTURE LENS — memory is 6% of what it unlocks, falling to 2%
Every argument on this page has been about shortage — how tight supply gets, how high prices go. This is a different question: how much is memory worth to the buyer? It has not been modelled until now.
One gigawatt of AI compute costs
~$50B of datacentre capex
Memory's share of that
~30% = $15B
That gigawatt earns
~$50B/yr → $250B over five years
So memory is…
6% of the value it unlocks
Now apply the generational gain. A new watt is 3-5× more performant than last year's watt — but memory content per gigawatt stays roughly constant at 170,000 wafers. The same memory spend now sits underneath three to five times the output.
2.0%
at 3× performance
1.5%
at 4× performance
1.2%
at 5× performance
The buyer could pay three to five times more for memory and be exactly where they are today. That is the headroom — and it is why this looks nothing like a commodity negotiation. Nobody fights hard over an input worth 2% of what it produces.
And SK compounds it from both ends. HBM4E delivers +20% energy efficiency with a cooling layer cutting heat over 30%. In a power-limited data centre, more efficient memory means more compute per watt — so SK is contributing to the very performance gain that widens its own pricing headroom. The supplier of the bottleneck is also improving the thing being bottlenecked.
Used honestly, this is a floor argument, not a ceiling one. Headroom is not a forecast — Samsung, Micron and CXMT compete, and contracts set the realised price. What it explains is why memory pricing should prove durable rather than mean-revert: the buyer's willingness to pay is rising faster than the price is. Our model does not raise prices on this. It holds them with more conviction.
⚖️ THE WHOLE THESIS IN ONE NUMBER — 170,000 WAFERS PER GIGAWATT
Per gigawatt of Vera Rubin compute: 170,000 DRAM wafers against 55,000 leading-edge logic wafers. 3.1× more memory silicon than logic silicon. This is a physical property of the machine, not a forecast anyone can revise.
NEW AI BUILD PER YEAR
DRAM WAFERS NEEDED
SHARE OF ALL WORLD CAPACITY
20 GW
3.4M
15%
30 GW
5.1M
22%
40 GW— implied by $2tn capex in 2028
6.8M
30%
50 GW
8.5M
37%
At $2 trillion of capex, roughly 30% of every DRAM wafer on Earth goes into AI data centres. That is why consumer memory rose 90% in a single quarter — it isn't a squeeze, it's an eviction. If every wafer in the world went to AI, total output would build only 134 GW a year. That is the ceiling on how fast AI can physically scale, and it is made of memory.
Two figures that look contradictory, reconciled: "AI consumes ~70% of global memory production" and "~30% of wafers" are both true — 70% by value, 30% by volume. HBM sells at $32-40/GB against commodity DRAM at a fraction of that, so AI takes a third of the wafers and two-thirds of the money. That gap is the margin.
🧮 THE 1.5-3× RULE — measured, not assumed
SemiAnalysis spent $3M and 2MW of continuously running compute benchmarking real agent workloads across 1,000+ chips at one million tokens of context. The finding that matters here:
"Long context conversations stress the HBM capacity for KV cache and necessitate offloading KV tensors to different tiers of memory (DRAM, SSD)… most effective when the amount of DRAM available for offloading is significantly bigger — a multiple of 1.5 to 3 — than HBM KV cache capacity."
B300, 384 agents, 3TB DRAM
91% HBM hit rate · working set 43M tokens
B200, 196 agents
73% HBM hit rate — leans on DRAM for ~20% of hits · working set 22M tokens
H200
"cannot compete with the newer SKUs in high throughput scenarios due to lack of HBM" — capacity, not compute, separates the generations
The agent economy does not choose between HBM and ordinary memory. It requires HBM and 1.5-3× as much DRAM beside it, plus SSD as a third tier — all three products SK sells. This is why Nvidia tripled the Vera CPU to 1.5TB of LPDDR: not a cost dodge, but the physics of serving long-context agents.
And it is happening now, not later: agentic workloads "now dominate traffic for production inferencing", with OpenAI's enterprise agentic spending overtaking ChatGPT spending in April 2026. Counterweight kept: Qwen3.5's GatedDeltaNet has "constant state storage" instead of attention's linear growth — architectures that shrink KV cache are real and being adopted. Logged against the decode-substitution scenario, which stays at 7%.
Every served request has three stages. Only one of them is what the new fast chips take.
1 · Prefill
Read the prompt. Compute-bound. Runs on GPUs — HBM.
2 · KV cache
Hold the whole conversation for the session. Pure capacity. HBM and pooled DRAM.
3 · Decode
Emit tokens one by one. Bandwidth-bound. This is the only stage Cerebras takes.
Decode was the bottleneck. Remove it and the other two bind — and both live in memory. SemiAnalysis states the scaling law directly: KV cache grows in proportion to concurrent users × context length. At NVIDIA's own figure of 40GB per user for a 70B model at 128K context:
4 TB
100 users
40 TB
1,000 users
400 TB
10,000 users
44 GB
one CS-4 wafer, total
Cerebras cannot absorb the demand it unlocks. A wafer holds 44GB — barely one user's cache, with no room for model weights. So the faster it makes decode, the more users get served, and the more of that 400TB has to live in HBM and pooled DRAM beside it. HBM per token falls. HBM per cluster rises.
We have already seen this play out once. Nvidia's TileRT software made GPU decode 1.9-3.6× faster three weeks ago. It did not reduce HBM demand — it made batch-1 inference viable, and batch-1 reads 21-42GB of HBM per token, the hungriest mode there is. And the price signal points the same way: SemiAnalysis reports premium-priced "fast modes" — users pay more for lower latency. Speed is not a cost saving handed to buyers; it is a new product tier sold at a premium, which is how cheaper intelligence ends up consuming more of everything.
💰 MEMORY IS NOW 62% OF AN NVIDIA AI CHIP
UBS teardown of a $39,000 Vera Rubin Superchip:
HBM4 on the Rubin GPU
$4,943
53% of the GPU's entire cost
SOCAMM2 on the Vera CPU
$19,355
96% of the CPU. Without memory the CPU costs $704.
Total memory share
62%
up from 53% in Grace Blackwell · memory cost 2.5× in one generation · 74.7 TB per rack
So when Nvidia announces servers going up more than 15%, the decode inverts the Apple one. Apple's memory was 12-15% of cost, so its price rise implied memory up +125-250%. Nvidia's is 62%, so a 15% rise implies memory up only ~24%. And that is the more powerful fact: at 62% of cost, the most profitable company in the chain cannot absorb an increase — it can only pass it on. Pricing power has moved upstream, to the people who make the memory.
🔍 IS IT MORE MEMORY, OR DEARER MEMORY? — the honest decomposition
HBM capacity per GPU did not rise at all. Both GB300 and Rubin carry 288GB per package; both racks hold 20.7TB of HBM. SemiAnalysis states it plainly: "HBM capacity remains flat from GB300, while HBM bandwidth scales ~2.8×." The generational gain came from bandwidth — double the bus width per stack at 10.8 GT/s — not from more memory.
GRACE BLACKWELL
VERA RUBIN
HBM per rack
20.7 TB
20.7 TB flat
CPU memory each
480 GB
1,500 GB +213%
Total rack memory
38.0 TB
74.7 TB +97%
So the 2.5× memory cost decomposes as roughly 2.0× more memory × 1.27× dearer memory — in this teardown. Most of it is volume, not price. The growth all came from the CPU side, where Nvidia tripled capacity and switched from server DDR5 to LPDDR5X.
The genuine bear point inside this: Nvidia is economising on HBM specifically — Rubin CPX ships 128GB of cheaper GDDR7 instead of HBM4 for the prefill stage. Deliberate HBM-intensity discipline, and it belongs on the record.
Three things that answer it: the teardown uses today's prices — Cantor has 2027 HBM4 contracted at $32-40/GB against the $17/GB implied here, so the price leg is still ahead. Flat capacity means HBM growth is unit-driven, which is exactly what our +40-50% bit assumption models. And the next generation reverses the flatness entirely: Kyber racks carry 340.4TB — 4.6× a Vera Rubin rack, at a reported $41.6M each.
🇹🇼 THE TAIWAN DEPENDENCY — the largest tail risk on this page
Yes, there is a direct link, and it is new. HBM4's base die — the logic layer at the bottom of the stack — is fabricated by TSMC on a 12nm process. Previous generations built the base die on SK's own DRAM process, so this single-point dependency arrived with HBM4. Samsung Foundry and Intel could in principle substitute, but requalification with Nvidia would take roughly a year.
But that is the small problem. TSMC also makes the accelerators themselves — Nvidia's Rubin, AMD's Instinct, Google's TPUs, Amazon's Trainium, Broadcom's ASICs — and runs CoWoS, the advanced packaging step that physically joins the GPU to the HBM stacks on an interposer. CoWoS was the binding constraint on Nvidia shipments in 2023-24, ahead of both wafer starts and memory supply. In a Taiwan event the customer disappears before the supplier does. There is nothing to attach HBM to, and no line capable of attaching it.
And there is a second exposure most people forget: SK's Wuxi fab in China makes 30-40% of its DRAM output. Any Taiwan conflict puts that plant inside the same crisis — seizure, export control, or forced idling.
How this is handled here, stated plainly. A Taiwan conflict is a market-wide event, not an SK-specific one — global equities, not one Korean stock. It sits inside the 3% "thesis broken" scenario (₩0.8-1.6M) alongside the 4% Korea and currency scenario, so roughly 7% of probability covers geopolitical and model failure together. That weight is a judgement, not a calculation — nobody can compute the odds of an invasion, and pretending otherwise would be false precision. It is named here rather than buried so the exposure is visible: this is the one risk that no contract floor, no buyback and no shortage protects against.
🤖 AN UNOPINIONATED MODEL AGREES — Simply Wall St's automated DCF
Simply Wall St runs a discounted cash flow on every stock in the world, every day, with no thesis and no analyst attached. On SK Hynix it returns a fair value of $519.56 against a last close of $163.41 — a 68.5% discount, or ×3.18 upside.
Our 12-month median
₩5.12M
×3.11
Simply Wall St DCF fair value
₩5.50M
×3.18
Our end-2027
₩8.0M
×4.86
It lands between our twelve-month figure and our end-2027 figure. That matters because the model has no narrative to defend — it cannot be talked into a conclusion, and it arrives near ours anyway.
On their 7.8× P/E: that is trailing. Ours is 2.8× forward on 2027 earnings — ₩1,647,000 ÷ ₩222K trailing versus ÷ ₩583K forward. Both are correct on different bases, and the forward figure is the one that matters if the earnings arrive. Their peer comparison is the striking part: 7.8× against a peer average of 58.4× and a US semiconductor industry average of 46× — an 87% discount to peers on trailing earnings, before any of the growth lands.
💵 WHAT PRICE RISE IS ACTUALLY EMBEDDED — audited
FY26 ₩360T → FY27 ₩816T is ×2.22. With bits at +24%, the embedded price assumption is +79% — not the "+55%" this page carried until now, which was a label left behind when the bottom-up rebuild raised the chain. The number was right; the description of it was stale.
HBM · 32% of revenue
+119%
$15.5/GB → $32-40/GB contracted for 2027
Conventional DRAM · 44%
+55%
Contracts printed +90-95% then +58-63% QoQ in H1, and the cap is gone on renewal
NAND · 24%
+38%
+55-60% then +70-75% QoQ in H1; eSSD ramp
Weighted
+71-79%
Cross-checks against the ×2.22 revenue step. Consistent.
And Nvidia's +17% is confirmation, not an addition. Nvidia is repricing servers because of the HBM4 contracts at $32-40/GB — the same contracts already inside the +119% above. Counting the pass-through again would book the identical price rise twice: once at the seller, once at the buyer.
🏭 PRODUCTION TRACKER — every source of output, and where it lands in the forecast
Baseline: SK's total DRAM capacity is roughly 450,000 wafers a month (industry estimate — the company does not disclose it).
SOURCE
WHAT IT IS
WHAT IT ADDS
WHERE IT SITS IN THE MODEL
1c node ramp
20K → 175K wafers/mo by end-2026 · 4% → 39% of the fab
Conversion, not new wafers
~22% more bits per wafer → +8.6% blended bits
Inside the conventional bits +6-12% input — at the top of that band. Also gates HBM4E, which needs 1c core dies.
M15X, Cheongju
Wafers deploying since Feb-2026, ramping through 2027
New capacity, substantially HBM
~40K wpm ≈ +8.9% new wafers
Supports the HBM bits +40-50% input. HBM burns ~3× wafer area per bit, so these wafers yield fewer bits — at far higher prices.
TSV / packaging
Drilling, thinning, bonding, test
HBM's real ceiling
The binding constraint — HBM takes >70% of industry TSV capacity
HBM output grows with the stacking line, not the fab. Adding wafers cannot lift HBM past this.
Yongin
Cleanroom Feb-2027, equipment Q2-2027
360K wpm only by H1-2030
Nothing in this window
Excluded from 2026-2029 output. Appears on the supply-wave side from 2030.
Does the ramp exceed what the forecast assumes? No — it confirms it. Node conversion delivers +8.6% against a +6-12% conventional assumption, and SK's own guidance of +10% DRAM shipments quarter-on-quarter annualises to about +46%, sitting inside the +40-50% HBM assumption. Both check out.
Why this table exists. The 1c ramp was previously netted into the bit assumption rather than shown, on the grounds that counting it twice would inflate the forecast. That reasoning was right, but the presentation was wrong: an assumption you cannot see is an assumption you cannot audit. Every output source is now a visible line, checked against each quarterly print.
📈 THE PRODUCT-SHIFT WINDFALL — now built into the numbers above
Two of the three asymmetries have been modelled in. The third was excluded to avoid counting it twice.
Tighter supply
Samsung adds 1.3% of ordinary capacity in 2027, not the 12% we assumed industry-wide. The gap widens ~5 points; SK has no price cap so it captures it. Worth +3-4% on turnover.
HBM4E content
36GB → 48GB per stack, Rubin Ultra at 192–384GB/GPU depending on stack height (8-high HBM4 = 192GB per SemiAnalysis 1 Sep; 12-high 288GB; 16-high 384GB) — the FY28 content line re-derives at the Oct 27 run; note the chain is wafer-gated, so fewer bits per GPU means more GPUs per wafer at higher price, not less revenue. Modelled conservatively — only half the FY28 volume moves across: units +25% × content +17% × price +10% = HBM lane ×1.60.
1c ramp
Shown separately in the production tracker above. The 20,000→175,000 wafer conversion delivers +8.6% bits, which is the existing bit assumption — so it is tracked as a line item rather than added twice.
The windfall, in money: FY27 turnover +₩29T and FY28 +₩94T; profit per share +₩21K then +₩74K. End-2027 moves ₩7.0M → ₩7.8M and end-2028 ₩7.4M → ₩8.3M.
And the check that kept it honest. The raw HBM4E build implied FY28 at ₩1,150T — but that would put SK at 37% of the entire world memory market, above what its DRAM and NAND shares support. Held at ₩1,102T, SK sits at 35% of a market growing to ~$2.2T: defensible against ~36% of DRAM and 57% of HBM. The upside was capped by arithmetic, not by caution. Bear mass 44%; every tripwire still live.
🔬 HOT CHIPS 2026 — what the two rivals actually showed
SK hynix
shipping now
12-layer HBM4E samples shipped ahead of schedule. Per-stack capacity 36GB → 48GB (+33%), pin speed 16 Gbps (from 11-13), efficiency +20%, first use of 1c DRAM, TSMC 3nm base die, MR-MUF packaging, plus a cooling layer cutting heat over 30%.
Samsung
architecture roadmap
zHBM — true 3D integration of accelerator and memory stack, removing the interposer entirely. Four stacks beside a 1,200W GPU: more bandwidth while saving ~100W. Needs wafer-on-wafer and hybrid copper bonding. Phase three of a roadmap, not a product — and note what it still is: DRAM, stacked differently.
The HBM4E step also answers last week's question. HBM capacity per GPU was flat from GB300 to Rubin at 288GB, which looked like content discipline. 48GB (12-high) stacks lift Rubin Ultra toward 288–384GB per GPU — though NVIDIA is reportedly weighing 8-high (192GB) configurations where HBM4E is scarce, which cuts bits per GPU, not our wafer-gated revenue — the flatness was a one-generation pause, not a trend.
And the supply figure worth watching: Omdia has Samsung adding just 1.3% of ordinary memory capacity next year against SK's ~10%. The number-two producer is choosing mix over volume — pushing HBM, barely adding conventional DRAM. Our model assumes conventional capacity grows 12% in 2027. If that is right, our supply assumption is too generous and the gap is wider than modelled. Single-sourced, so logged as an asymmetry to confirm at the October print rather than a revision.
🌶️ THE CUSTOM-CHIP THREAT ARRIVED — and it buys HBM4
ACCELERATOR
HBM PER GPU
PER RACK
MEMORY
Nvidia Rubin
288 GB
20.7 TB
HBM4 @ 9.6 Gbps
AMD MI455X
432 GB
31 TB
HBM4 — 50% more than Rubin
OpenAI Jalapeño
—
128 ASICs
HBM4 @ 10 Gbps · 15.4 TB/s per package — the highest shipping
Every serious new architecture runs HBM4, and AMD carries half again as much as Nvidia. Competition between accelerator vendors pushes memory content up. The custom-ASIC threat everyone feared turned out to be another HBM4 customer — with the highest bandwidth per package in the industry.
The risk, corrected and sized: SemiAnalysis writes Jalapeño's HBM is "likely provided by Samsung" — but OpenAI signed memory LOIs with both Korean makers in Oct 2025, so this is one socket, not a defection. A share risk, not a demand risk — and Jalapeño's next goal is 100 MW, which is 17,000 DRAM wafers, 0.075% of world capacity. Even at 2 GW it is under 2%. Filed as evidence for the existing 8.5% share-loss scenario, not as a new one. Trigger to revisit: a confirmed award, or Jalapeño past 1 GW.
🧾 WHO ACTUALLY BUYS THE ROADMAP — and when
The roadmap items split sharply into revenue with named customers and optionality with neither. Worth keeping them apart.
PRODUCT
BUYERS
WHEN
Custom HBM4E
customer logic on the base die
Nvidia · Microsoft · Broadcom — named and ordering. Requests received from across the Magnificent Seven. Ecosystem ties to Broadcom and Marvell.
Samples shipped 18 Jun 2026
48GB · 16 Gbps · mainstream 2027
Standard HBM4
SK holds two-thirds of Nvidia's HBM4 orders; also supplies Microsoft's Maia 200
Shipping now, ramp H2-2026
LPDDR6-PIM
processing-in-module
No named customer. Described as being for "specialised applications"
2028
High Bandwidth Flash
None. Needs all-new media and a spec agreed with SanDisk and other NAND makers
Not before 2030
So: custom HBM is real, sold and near. The custom HBM market is forecast to grow from $18.2B in 2024 to $130B by 2033, and SK is described as set to lead it on the strength of Nvidia, Microsoft and Broadcom orders. PIM and HBF are optionality — no named buyers, 2028 and 2030 respectively. They belong in the thesis as free upside, not in the revenue lines, and our model carries them at zero.
And the contest is real. Samsung has added 250 engineers to custom projects targeting Google, Meta and Nvidia, moved its custom logic die to a 2nm process, and supplies 60%+ of Google's TPU HBM3E. Micron is reported as trailing because it kept its existing DRAM process to hold costs down — "a structural disadvantage in the custom HBM race."This is where the 8.5% share-loss scenario actually gets decided.
🧬 SK'S OWN ROADMAP TO 2031 — the answer to "what replaces HBM"
Every architecture threat on this page — MRAM, RRAM, 3D DRAM, processing-in-memory — is already on SK's published roadmap. TrendForce: SK and Samsung "both hold leading patents and prototypes" in MRAM and ReRAM.
2026-28
HBM4 16-Hi · HBM4E 8/12/16-Hi · custom HBM4E with the controller moved onto the base die (with TSMC — frees compute area, cuts interface power) · LPDDR6 · SOCAMM2 · MRDIMM Gen2 · CXL LPDDR6-PIM · 245TB+ QLC enterprise SSD
2028
LPDDR6-PIM — processing-in-module, i.e. SK entering near-memory compute itself
2029-31
HBM5 · HBM5E · custom HBM5 · GDDR7-next · DDR6 · 400+ layer NAND · 3D DRAM
Terafab is targeting MRAM, RRAM and 3D DRAM — the exact technologies SK already holds leading patents in. And custom HBM with per-customer controller silicon deepens the moat rather than commoditising it. UBS expects SK at ~70% of HBM4 for Rubin — above its 57% overall HBM share.
And one door just closed. Musk is "exclusive to Nvidia". Nvidia's Rubin generation runs on HBM4/HBM4E from the three qualified suppliers, and Terafab's own memory is MRAM/RRAM/3D DRAM arriving "as early as 2028". Musk cannot bypass HBM inside this window even owning a fab. What remains from that quarter is share risk, not substitution risk.
🧪 WILL SOMETHING REPLACE HBM? — the four candidates, examined
Etched (Sohu)
It's a customer. Carries 144GB of HBM3E per chip; its whole speed advantage comes from HBM bandwidth. Every Sohu shipped is 144GB of HBM sold.
Taalas
Real, but tiny. Etches weights into metal — genuinely no HBM. But it holds 8 billion parameters (~4GB). A frontier model is 200× larger and changes every few months. Works for small frozen models at huge volume; not the frontier.
Photonic memory
Doesn't exist. Photonic chips do compute, not storage — light doesn't sit still. Optical interconnect is real, and it moves data between HBM systems faster.
Intel XBM / ZAM
Credible — and dated 2029-2030+. ZAM is a nine-layer stack of conventional DRAM; XBM is a DRAM stack that ditches the interposer for UCIe. ZAM 2029-2030 · XBM "start of the next decade." Intel has exited memory five times: RDRAM, Optane, HMC, MCDRAM, and its NAND business — sold to SK hynix.
The pattern: every serious attempt to supersede HBM either buys HBM, fits only a frozen small model, doesn't exist commercially, or is still DRAM in different packaging. Bandwidth to a large working set is physics, not fashion — and it keeps returning the same answer.
And Terafab's own roadmap says it too: its job postings name MRAM, RRAM and 3D DRAM — not HBM — with operations "as early as 2028". The forecast on this page runs to 2029. None of these arrive inside it.
🔧 WHY SK CAN'T JUST BUY CHINESE DRAM AND STACK IT INTO HBM
The obvious question, and the answer is the moat. HBM is not made from DRAM chips — it is made from dies that were purpose-built for stacking at the wafer stage.
The holes are drilled first
Through-silicon vias — thousands of vertical copper channels — are etched and filled during fabrication, not afterwards. A finished commodity DRAM die has none and cannot be given them. SemiAnalysis: TSV tooling is "the main bottleneck when it comes to converting regular DDR wafer capacity to HBM capacity."
The bottleneck isn't dies
A normal DRAM package takes ~700 process steps. HBM needs 19 additional steps on entirely different equipment — TSV formation, thinning each die to ~30 microns, precision stacking, thermal-compression bonding. None of that equipment can do conventional packaging. HBM will consume over 70% of the industry's total TSV capacity this year. More dies into a saturated stacking line produces nothing.
Yield compounds
Four serial steps — die fab, TSV, bonding across 8-12 layers, stack test. One bad die ruins a whole stack, which is why HBM burns ~3× the wafer equivalent of DDR5 per bit. Foreign dies of unknown uniformity would wreck the economics before qualification even began.
The layout is the secret
"How a company arranges its TSV layout is one of the core proprietary differentiators between memory vendors. It directly determines yield and performance, which is why nobody publishes it." Outsourcing the core die means handing over the moat.
HBM4 goes the other way
The base die now moves to a foundry logic process and carries customer-specific logic — so Nvidia's HBM4 and AMD's HBM4 are physically different products. Memory is becoming semi-custom, not more interchangeable.
The twist worth holding: China expanding commodity DRAM is partly useful to SK. It serves the ordinary memory market that SK is deliberately abandoning to build HBM — which lets SK specialise further without leaving customers stranded. That is precisely why the China scenario in the model attacks the commodity lane at 7% weight and leaves the HBM lane untouched. And the industry does buy specialist parts from each other — SK's own HBM4 base die is made by TSMC. Just not the memory dies, because those are the product.
📊 THE WHOLE SECTOR RE-RATED. SK DID NOT.
2026 RETURN
YTD
TRAILING P/E
WHAT THEY MAKE
SanDisk
+653%
41.5×
Flash memory
Seagate
+262%
61.7×
Spinning hard-disk drives
Micron
+255%
18.6×
DRAM · 21% of HBM
Western Digital
+211%
32.6×
Storage
SK hynix
slightly negative
110.2×
57% of the HBM market · highest margin in the industry
Seagate makes spinning platters and trades at more than three times SK's multiple. The market is paying most for the parts of the storage stack furthest from AI acceleration, and least for the company that makes the memory every accelerator requires.
Why — and it is a flow story, not a fundamental one. July's pullback is attributed to "the collapse of AI-focused hedge fund Situational Awareness", whose forced selling hit Korean names hardest. Cramer's own trust bought Micron "during a pullback tied to a selloff among South Korean semiconductor stocks."US investors used Korean forced-selling to buy US-listed memory instead. Flow gaps close; fundamental gaps do not.
One honesty note: some sources quote SK's P/E at 7.8×. That figure is flattered by the ~₩63T Kioxia investment gain in reported profit. On clean earnings the trailing multiple is nearer 110.2×. The 3.0× forward used on this page is built on clean earnings throughout — but 7.8× should not be quoted as evidence.
⚔️ SAMSUNG'S OWN ACCOUNTS, SIDE BY SIDE — same quarter, same shortage
SK HYNIX
SAMSUNG CHIPS
Q2 revenue
₩79.3T
₩127.5T
Q2 operating profit
₩60.5T
₩89.2T
Operating margin
76.3%
70.0%
HBM market share
57%
21%
Price ÷ annual operating profit
5.3×
7.3×
SK earns a higher margin, holds nearly three times the HBM share — and trades at a 27% discount to its own domestic rival. Samsung's shares ran ~150% over the year, so its record quarter was already priced in and the stock fell 6.9% on the print. SK's year-to-date return is slightly negative. Same shortage, same country, same customers, same quarter — one was re-rated, the other was not.
And the detail that says everything: Samsung's Galaxy phone division posted its first operating loss ever (₩0.7T) on "elevated component cost pressures". Samsung's memory business is bankrupting Samsung's phone business. The world's second-largest handset maker cannot absorb the price of the memory its own colleagues are selling.
⚖️ THE COMPARISON MICRON'S OWN SHARE PRICE MAKES
SK HYNIX
MICRON
Latest quarterly revenue
$57.3B
$41.5B
Share of the HBM market
57%
~21%
Market value
~$931B
~$1,300B
Price ÷ annual revenue
4.1×
7.8×
SK earns 38% more revenue, sells nearly three times the HBM — and carries 28% less market value. Micron trades at 1.9× SK's revenue multiple. Priced on Micron's own multiple, SK would be ₩3.34M today — 93% above Friday's close, before any of the 2027 earnings growth arrives.
Micron's CEO on the same day: customers want 50% more supply than we can commit; "all our customers will buy everything that we make"; 16+ take-or-pay contracts backed by $22B of customer cash deposits; next quarter guided +21%. The shortage is described identically by the competitor — the market simply prices the two companies as though it were not.
⚠️ WHY JULY FELL 55% — AND THE LESSON IN IT
It was not the business. A $45B fund running ~4× leverage held SK Hynix as one of its largest longs — roughly $7B taken at the July 10 US listing. When AI infrastructure sold off, its $16B book lost 67% in one month, prime brokers called the margin, and on 30 July the entire portfolio went to Citadel in a single block.
30 July
Book transfers. SK bottoms at ₩1,320,000 — the same day — then rises ~30% the next session.
August
Citadel distributes: 100+ block trades, $4B+, single-session records in ten names. This is why rallies kept stalling.
21 August
Griffin's letter: more than 80% unwound. Roughly 20% of the overhang remains.
18 August
Separate event — a rates shock, 30-year at a 19-year high, uniform across the whole complex. Not this.
The lesson is the valuable part. That fund's thesis was ours — AI scaling forcing an unprecedented buildout in chips, memory, data centres and power. It was up 439% and right about everything. 4× leverage still destroyed it, and the assets it was right about recovered without it. Same thesis, same stock, different survival — the only variable was leverage. This is why the discipline here reads "the sizing is not the risk, the path is," and why we do not add at a maxed position.
🏦 THE FUNDING POOL JUST CHANGED SIZE — and that was the bears' best argument
What the bear case assumed
The buildout is limited by hyperscaler cash flow and corporate bond capacity. Capex already ran at 94% of operating cash flow, and spreads were widening.
What just happened
Nvidia signed Apollo, Blackstone, BlackRock, Brookfield, Goldman Sachs and KKR to mobilise $500B of third-party capital. Huang: "the first time technology chips have become an investable asset class."
The enabler
Four days later the SEC waived the Dodd-Frank risk-retention rules for this debt — the post-2008 requirement that sponsors keep skin in the game.
Proof it works
Morgan Stanley: data-centre ABS spreads near year-to-date tights even while corporate spreads widened. Issuance $2.4B → $15.5B, JPMorgan sees $30-40B/yr by 2027.
The catch, stated plainly
This is 2008's machinery rebuilt. The BIS warns private-credit opacity makes systemic exposure unmeasurable, IT investment is ~5% of US GDP — above the dot-com peak — and the collateral depreciates: 2027 brings the first GPU lease-renewal cliff.
Net: the runway is far longer than we modelled — the buildout is no longer capped by what buyers earn. But a named systemic fragility now lands in 2027-29, the same window as the supply wave. Less near-term funding risk, more late-cycle risk. The runway lengthened; the late-cycle risk is now named.
🚗 PHYSICAL AI, SIZED — cars and robots, and where they sit in the model
Memory per car
90GB → 278GB (Micron) — 3.1× as cars move to centralised computers
Global auto demand
90M vehicles × ~$350 ≈ $31B/yr · the content jump adds ~$21B ≈ 2.4% of the memory market
Humanoid robots
Shipments +272%/yr · 10× a car's memory each (Micron CEO) · Jetson Thor carries 128GB
In the model?
Yes — both. Three physical-AI lanes (devices · vehicles · fleet training) sit inside the +24% bit-growth input and the year path. This confirms the sizing rather than adding to it.
What the car story really proves is inelasticity. Ford is absorbing $1B of extra memory cost rather than pausing production, and average new-car prices are heading from $50,000 toward $60,000. A car ships with its memory or it doesn't ship. That is the 0.25 elasticity assumption in this model, demonstrated in someone else's accounts.
📌 WHY 2029 PAUSES RATHER THAN FALLS — and what the pause is actually made of
The 2029 step-down is entirely a valuation-multiple effect, not a demand forecast. Profits keep rising right through it: ₩828K → ₩894K → ₩948K per share across 2029-31, because agent traffic, humanoids and autonomous fleets are all still scaling. What steps down is what the market pays for those profits, from 10.2× to 8.8×, as supply finally lands and depreciation peaks.
Three demand curves begin in this window rather than ending — Micron's CEO puts the multi-decade cycle starting "in the latter part of this decade", Cloudflare's 1,000× traffic build lands around 2031, and Intel's CEO has publicly retired the "memory is a commodity" premise that compression depends on. So the wave becomes a plateau, not a dip. It is not removed, because the supply genuinely arrives, D&A climbs toward ₩120T, and markets de-rate cyclicals ahead of the event regardless of who is right about demand.
💬 THE BELIEF THAT DISCOUNTS THIS STOCK IS DYING IN PUBLIC
"I used to be, do not invest in memory because it is a kind of commodity business — but now it has become different."
— Lip-Bu Tan, CEO of Intel, Aug 2026 · he also says there is no relief on memory supply until at least 2028
SK trades at PEG 0.03 versus 1.08 for the chip industry for exactly one reason: memory is priced as a commodity. When the industry's own veterans stop calling it that, the discount has a mechanism to close — that is the re-rating block in the waterfall above.
"Customer demand has doubled this year… next year will be the worst year for memory shortages… it's somewhat like a war."
"I'm not saying there are no cycles, but the shape of the cycle must change — it will become longer. This is a turning point, a structural shift."
— Chey Tae-won, Chairman of SK Group, CNBC, Aug 13 2026 · he also showed a wafer inscribed by Nvidia's CEO: "Please make more."
🧾 WHAT BUYERS ACTUALLY PAY NOW — printed receipts, not forecasts
OVHcloud RAM invoice yr to Jun-26
6×
…same buyer, next year
12×
Consumer RAM this quarter alone
+90%
DDR5 modules since Sep-25
+307%
DDR4 since Sep-25
+158%
HBM4 $2/Gb → 2027
$4-5/Gb
Apple products implied memory line
+125-250%
OVHcloud estimated 5-10% increases in March. Five months later it is raising rents up to 87% and expects the squeeze until 2028. The flood outran a professional buyer's own forecast tenfold. Our model prices SK's contract-weighted blend — far below these spot receipts — which is why the chain still reads +30-45%, not +500%.
🏦 WHAT THE BANKS SAY — ADR price targets vs today ($, Aug 2026)
Six banks initiated after the Nasdaq listing. Every one calls for a double or more. 39 desks now cover it: 38 Buy · 1 Hold · 0 Sell. Our number sits above them because the flood pricing is in our chain and only partly in theirs.
REVENUE & PROFIT TRAJECTORY (flood-weighted)
380
2026
816
2027
1102
2028
Revenue ₩T
270
2026
583
2027
753
2028
Clean profit per share ₩K — today's price = 3.2× the 2027 bar
📅 THE PROOF CALENDAR — what lands, when, and what it settles
AUG 26
🟢 Nvidia earnings
$91B guided — the demand read
OCT 5
⚡ Compute futures list
first forward curve = live demand read
Q4 2026
📜 2027 pricing settles
Edgewater: not final until late 2026
AUG–SEP
💰 Returns framework
cash >40% of mcap by '27
LATE OCT
⚠️ Q3 · tripwires + share
<₩98T or <55% → step down
DEC
📈 Nasdaq-100 window
passive funds must buy
FEB 27
🏭 Yongin opens
pulled forward 3 months
CONCLUSIONS AT A GLANCE
🔒 The shortage is real & dated
<70% of orders filled · CEO: worst in 2027, demand > capacity past 2030
🌊 Prices are exploding, provably
Apple decode +125-250% · market ×4 → $890B · spot +77% over contract
🏳️ The buyers have surrendered
Musk builds · AMD buys · hyperscalers prepay $38B · Apple hoards $11B · Nvidia prepays SK
📜 Contracts changed the game
Perpetual 5-yr terms · price floors + open ceilings · 70% of Nvidia HBM4 locked to SK
🏦 The smartest money is in
BlackRock: SK = #1 position > Nvidia · Baker triple-digits · Goldman models the harvest
😴 The stock hasn't noticed yet
3.1× forward · PEG 0.03 vs industry 1.08 — a 97% relative discount
🛡️ The bad endings are covered
Worst 1-in-20 = ₩2.15M > today · two written demotion tripwires · 2029 wave checkpointed
📅 The proofs have dates
Aug 26 Nvidia · late Oct Q3 + tripwires · Q4 2027 pricing · Dec returns + index
EXECUTIVE SUMMARY · AUG 12, 2026 · ENGINE v9.2 · YEAR PATH v5
The world ordered more AI memory than can be built. SK Hynix makes most of it — and costs 2.8× next year's profit.
Evidence says ₩4.7-5.6M within a year. Market says ₩1.647M. A second wave — robots — is loading behind the first.
🧾 OVHcloud pays 6× → 12× for RAM · rents +87%🏦 $500B Nvidia + six capital managers · compute becomes an asset class⚡ WATCH: 30-yr yields at 19-yr high · securitisation = 2008 machinery rebuilt💻 Consumer RAM +90% in one quarter · Lisa Su: no relief until 2028🛢️ AI capex $765B — passed oil & gas ($681B) for the first time🍎 Apple decode: memory +125-250% implied📈 Market ×4 → ~$890B this year📜 Contracts: floors + open ceilings, perpetual 5-yr💰 $38B prepaid · $50-100B more offered🏦 BlackRock: SK = #1 position (7.64% > NVDA)🧠 Baker 13F: SK/MU/SNDK all +100%+🚀 Musk: 200% demand vs 20% supply → hires DRAM engineers🏭 CEO: worst shortage 2027, demand > capacity past 2030📊 Citi: HBM/system +434% · fulfilment <70%🏆 SK: 76% operating margin · 60%+ HBM share🤖 Robotics: humanoids +272% y/y · 10× car memory · multi-decade cycle (Micron CEO)
Five giants drove free cash flow from a $125B/qtr peak to ~zero — choosing the AI buildout over their own cash — and Goldman models record FCF by 2029 as the AI harvest lands. The spending is proven; the payback is modeled; memory is inside every dollar of both.
GAVIN BAKER · ATREIDES ($7B) · ALL-IN POD
"The most supply-constrained semiconductor market I have ever seen." Four reasons the crunch persists — devices can't even run AI locally; TSMC discipline kills the glut. Google & Microsoft offering $50-100B prepayments for 3 years of output. SK's stance: "I won't guarantee supply — I'll just raise prices."
Receipts: added SK Hynix, Micron, SanDisk late-2025 (13F) — every one +100%+ in 2026
TONY KIM · BLACKROCK ($15T) · HEAD OF GLOBAL TECH
"The RAMpocalypse.""The human brain is memory-intensive; today's AI is compute-intensive. Going forward, memory in concert with compute becomes even more important." Capex: $1T this year, $10T over five.
Receipts: SK Hynix = #1 position in BlackRock's flagship tech fund, 7.64% — above Nvidia
ELON MUSK · THE BIGGEST BUYER
"Demand +200%/yr vs supply +20%/yr. Economics 101 would suggest the price increases. It does not decrease." His prediction, expressed in actions: 10 GW Nvidia-exclusive buildout — then hired DRAM engineers rather than trust the market.
Buyer → pleader → builder: the strongest duration testimony an insider can give
FWD P/E 2027E
3.1×
FWD P/E 2028E
3.1×
PEG
0.03
FWD PEGY
0.03
INDUSTRY PEG
1.08
SK −93% vs it
NOW
₩1.647M
2.8× next-yr profit
SKEPTIC FLOOR
₩3.3-4.4M
history's harshest rule
FLOOD-WEIGHTED · 12 MO
₩4.7-5.6M
fallbacks: 3.9-4.2 / 3.5-3.7 armed
24 MONTHS
~₩8.0M
end-27 path: ₩753K × ~10.6× (₩6.7–8.9M)
MUSK PATH
₩5.3-7.0M
if +30%/qtr persists
WORST 1-IN-20
₩2.15M
above today (on calibrated tails — see engine notes)
97%
of simulated futures end higher
91%
end at least double
+211%
to the median ₩5.12M · 66% of futures at least double (post-recal) · doubling claim survives 100% of stress configs
8.2
confidence · central case · 45% bear mass
YEAR-END PATH
'26
5.5
→
'27
8.0
→
'28
8.4
→
'29 pause
7.9
→
'30
8.5
₩M · plan: repositioned before '29
NEXT DATESAug 26 · Nvidia $91B guidelate Oct · Q3 = THE PROOFDec · returns + indexany day · DART on the $2.8 rumor
FROM AGENTS TO SHARE PRICE — THE FIVE-STEP TRANSLATION (each step carries its printed receipt)
1 · DEMAND
Chairman: customer requests DOUBLED this year · agents ×4 by '28 on Goldman's floor (Cloudflare wires ×1,000)
→
2 · SUPPLY
Bits +17%/yr — Chey: "no memory companies are yet ready to scale up" · his own $720B = 14.7%/yr
→
3 · ELASTICITY
25pt gap ÷ 0.25 elasticity = +100% theoretical — but contracts cap realized at +42%
Memory is ~12-15% of a MacBook's build cost. Apple repriced whole products +18-25%. Passed through dollar-for-dollar, that implies Apple's memory line rose +125% to +250%. Cross-checks agree: the memory market is quadrupling to ~$890B this year, spot trades +77% over contract, contracts rose +50%+ per quarter in H1, and Cook says he'll pay more next quarter. The flood-weighted chain above simply believes these printed numbers.
So yes — the 30-45% flood pricing is the 2027 engine: that year's +71% revenue growth decomposes as ~+37% price × ~+22% volume. The Nvidia lock doesn't add on top; it documents the funding of the center already adopted.
📐 THE ELASTICITY GAP — why realized prices rise far less than spot, and why that protects the downside too
DEMAND +42%
Chairman: customer requests doubled. Discounted hard for double-ordering and hoarding.
SUPPLY +17%
"No memory companies are yet ready to scale up." His own $720B adds just 14.7%/yr.
SK removed the price cap (TrendForce, 2 Jul) — spot rises now flow fully into contracts. Micron kept its cap. Same shortage, different capture.
The 58-point gap between theoretical and realized is the single most important number here — and it cuts both ways. Contract ceilings stop us capturing the mania on the way up; the same contracts' price floors stop the collapse on the way down. That trade is why 2029 is modelled as a plateau, not a crash — and why this is a lower-variance bet than a spot-price bet on the same shortage.
PRICE & PATH — ₩M, 2024 → 2030
White = actual · green dotted = likeliest path with shaded range band · amber = regime flip level · blue = your entry. Overlay above the whole path if +30%/qtr persists (Musk rung: ₩5.3-7.0M).
VALUATION vs PEERS & INDUSTRY · street approximations Aug-2026 for relative scale; SK = this site's clean-EPS model
NAME
FWD P/E '27E
EPS GROWTH
PEG
PEGY
READ
SK HYNIX
3.1×
+51%
0.03
0.03
the anomaly
Micron
~9×
~+45%
0.20
0.20
same trade, 3× the price
Samsung
~12×
~+40%
0.30
0.28
conglomerate discount
Nvidia
~23×
~+45%
0.51
0.51
the customer
TSMC
~23×
~+30%
0.77
0.74
the escapee benchmark
AMD
~38×
~+42%
0.90
0.88
paying for the story
Broadcom
~36×
~+28%
1.29
1.22
fully priced
Semi industry (SOX)
~27×
~+25%
1.08
1.05
the market's yardstick
One line: SK grows 2× the industry rate at 1/7th the industry multiple — PEG 0.03 vs ~1.0, a ~97% relative discount, and 65% below even the cheapest peer running the identical memory trade.
PRICE TARGETS & UPSIDE · from ₩1,647,000(row percentages last computed on the 21 Aug ₩1.730M basis — full refresh at the Oct re-run)
SCENARIO
TARGET
UPSIDE
GATE
Street bear (BNK)
₩1.48M
+4%
even the bear target is above the price
Skeptic floor (history's 6-8×)
₩3.3-4.4M
+62-111%
no re-rating needed
Citi (held through the crash)
₩3.1M
+118%
"buy the dip"
Street consensus
₩3.32M
+133%
the average desk
Fallback 12-mo (if Oct <55%)
₩3.5-3.7M
+146-160%
analyst assumptions only
Flood-weighted 12-mo
₩4.7-5.6M
+169-194%
printed flood prices + peer-set multiple
Korea Investment (raised Aug)
₩4.7M
+231%
the countertrend desk
Rung 3 — Musk path
₩5.3-7.0M
+273-392%
+30%/qtr persists; kill <+15%
TECHNICAL MAP · flow chart, not thesis
All-time high (mid-Jun)
₩2.19M
−35% drawdown from peak = historical shakeout scale
Regime tripwire (200-EMA 4H)
₩1.744M
a close above flips the dashboard GREEN (+22.6%)
Now — the thesis line (P4)
₩1.600M
kissed to 0.14%, not breached; weekly close decisive
Support 2 (200-day vicinity)
~₩1.30-1.35M
the long-term trend's last rail
Structure read
capitulation volume Thu · two-wave margin unwind late-stage · ₩1.744M tested 28 Aug and not held — regime flip unconfirmed, the gate stands · ₩1.58M (31 Aug low, Warsh shock absorbed by the buyback) = tested support above the ₩1.500M capitulation low · a weekly close above ₩1.744M is the signal · ADR: IPO base wedge, breakout ~$172
OPTIONS & FLOW DESK · mechanics, approximate
Nearest expiry
Aug 26
last dated derivatives-pressure window; monthly KRX settlement
Vol regime
IV spiked
post −10% day fear premium — rich put skew; pros sell it, retail buys it
Gamma pin zone
₩1.40-1.45M
heavy round strikes; price magnetised near them into Thursday
Clearing tells
IV crush after expiry · put/call ratio normalising · margin balance <~₩4T · green close on rising volume
ADR channel
SKHY options live
US-listed hedging/expression now exists — deepens the arb that closes the anomaly gap
▾ Tap any section below to expand the full reasoning.
1 · What is being built — the demand is physical and verified
The four tech giants are spending $732 billion this year on AI infrastructure (they spent $301B in just the first half), passing $1 trillion next year; Goldman Sachs counts $7.6 trillion through 2031. New buyers joined this month: SpaceX is building ~10 gigawatts of computing by end-2027 — verified not by promises but by analysts literally counting the power turbines at its sites (27 → 69 in five months) — and Microsoft signed 10 gigawatts of binding contracts this year. Every gigawatt needs ~500,000 Nvidia chips carrying ~$2.2 billion of the memory SK makes.
2 · Why prices must keep rising — simple supply and demand
Wanted demand is exploding — Elon Musk says 200% a year for his companies ("Economics 101 — the price increases. It does not decrease."). But factories can only grow output ~18-20% a year, and you cannot sell chips that don't exist — so all the demand that can't be built turns into price and into time: prices rising +20-30% a year (September contracts just printed ~+30%), and the shortage running years longer than analysts model. And here is the part that was fixed: last quarter SK captured only ~48% of those price rises, because its old contracts had price ceilings — that is exactly why the quarter felt like an underperformance. The ceilings are now gone: new contracts lock the volumes but let the price float. Same factories, same customers — a bigger share of every future price rise flows straight to SK. The proof arrives with the October results: if capture prints at 60-75% instead of 48%, the earnings and share-value numbers below all step up. Industry detail: the contracts now also carry price floors — quarterly declines capped, rises of 10-20%+ allowed — across all three makers, with ~$38B of customer prepayments behind them and five-year terms that roll forward perpetually. Heads SK wins; tails SK doesn't lose. Footnote: the rumor behind the latest share-price fall — HBM4 for Nvidia at $2.8/Gb, called \u201cdisappointing\u201d — is ~50% above the $1.9/Gb these forecasts actually assume. The market panicked toward a number that would raise this page. Samsung is meanwhile quoting $4.5-5 and refusing to undercut.
3 · Who pays SK — the customers, by name
Nvidia is the biggest: ~₩44 trillion flowing to SK next year from Nvidia alone (SK supplies roughly 60% of Nvidia's HBM; that memory is ~10% of everything Nvidia sells). Behind Nvidia: Google (TPU chips), Amazon (Trainium), Microsoft/OpenAI, Meta, and Broadcom-built custom chips — together about 40% of HBM demand. New this month: SpaceX (~₩15-19T of memory inside its Nvidia build). Roughly: Nvidia ~55-60% of SK's HBM sales · the giants' own chips ~30-35% · everyone else ~10% — and the customer list is getting longer, not shorter.
4 · What SK will earn — sales and profit
Sales: ~₩360T this year → ~₩816T in 2027 → ~₩1,102T in 2028. Clean profit per share (one-off gains stripped out, new-factory depreciation honestly charged): ~₩270K → ₩583K → ₩753K. The reason margins hold above 70% on the AI half: HBM is sold out through 2027 on contracts that now have no price ceilings — when spot prices rise, SK's contracts rise with them.
5 · What the shares are likely worth
Today: ₩1.647M = 2.8× next year's profit — a crisis price attached to a boom business. At the ordinary 9-10× that memory makers earn in good years: ~₩3.5-3.7M within 12 months — and three completely different methods land on that same number (compounding the actual printed price rises; 200,000 simulated futures; textbook supply-demand economics). ~₩4.1-4.6M within 24 months. If December delivers the buyback plan and index inclusion, the multiple lifts and up to ₩4.5M arrives sooner. And one upgrade is already queued: if the October results confirm the new no-ceiling contracts are capturing 60-75% of price rises (vs 48% last quarter), the 12-month value moves up to ~₩3.7-4.0M — the engine upgrades the day that number prints. The analyst consensus target is ₩3.32M — the street converging toward these numbers from below. From today's price, the flood-weighted simulation says: ~97% of futures end higher, the median lands at ₩5.12M (+211% from ₩1,647K; doubling probability recalibrated to 66%) — and the worst 1-in-20 outcome improved again to ₩2.15M (Wuxi de-risk), comfortably above today's price.
5b · The price ladder — what happens if the rises keep coming
Above assumes price rises slow down, as analysts model. If Musk's arithmetic holds (demand +200% vs supply +20%), the +30% quarters simply keep repeating. Each path, with +18%/yr volume and no-ceiling contracts working:
Superseded — the current forecast is the table at the top of this page and the bottom-up build on the flow model.
Honesty note: Rung 3 is not the base case — at those levels customers' memory bills triple and pushback intensifies, and UBS already models deceleration. But pieces of Rung 3 are printing right now, this cycle has precedent (memory roughly doubled in the past year), and the point of the ladder is that you don't have to guess: each quarterly contract print tells you which rung you are standing on — and the engine steps up the day it prints, never before.
5c · The year-by-year path — likeliest price at each year-end
Markets price next year's profit at year-end. Recalibrated after a review caught double-conservatism (cautious EPS × sub-Micron multiples, while the evidence prints both prices AND volumes rising — ASP +19.9%/qtr broker consensus, HBM4 floor +47%, Yongin pulled forward into <70% fulfilment — plus the structural re-rating already underway: ADR live, floors signed, index in December). This path blends Rung-1/Rung-2 earnings with a multiple converging past Micron's 9× toward Samsung's 12×. Fallback rule: if October's capture prints <55%, every year drops back one rung.Read the time axis carefully to avoid an optical illusion: these are waypoints on one rising curve — end-2026 is only ~4.5 months away, so an end-year waypoint naturally sits below the 12-month fan median of ₩5.12M (which lands in Aug-2027, between the first two rows). Nothing here is a downgrade of anything above; the simulation's every number rose again in the Aug-12 v9.2 recalibration:
Micron trades
~9×
same trade, US listing — SK passes this on ADR flows + index alone
Samsung trades
~12×
the ceiling of the convergence path — reached only as floors + buyback compound
SK path assumed
9.5× → 10.6× → 10.2×
never above Samsung, never below Micron by 2027 — the middle of its own peer set
End 2026
~₩5.5M (₩4.6-6.1M)
prices 2027E ₩583K at ~9.5× — the flood repricing lands in Q3/Q4 prints, returns paid, index in
End 2027
~₩8.0M (₩6.7-8.9M)
prices 2028E ₩753K at ~10.6× — two years of floors delivered, buyback running, Micron-parity passed en route to Samsung's 12×
End 2028
~₩8.4M (₩7.0-9.4M)
prices 2029E ₩828K at ~10.2× — the robotics wave begins (Micron CEO: multi-decade cycle), buyers at record FCF, floors hold the multiple
End 2029
~₩7.9M (₩5.8-9.0M)
prices 2030E ₩894K at ~8.8× — the wave lands, but into three demand curves that are starting (agents · robots · autonomous fleets), so the multiple pauses rather than breaks — met head-on by the physical-AI demand wave (absorption both sides) — history prices the trough here, softened by floors + robotics demand + flush buyers. The plan: repositioned before this, on the checkpoints
End 2030
~₩8.5M (₩6.2-9.6M)
prices 2031E ₩948K at ~9.0× — the higher-trough resolution if the price floors and contracts hold — the SK Securities case
Shape to remember: up hard into 2027-28 · breathe through the wave · resolve higher. 2026-27 are evidence-dense; 2029-30 are sketches with wide bands, refreshed each print. Plan: ride the tight years, watch the dated checkpoints, reposition before the year history says to fear.
6 · What could break it — honestly, and what we found when we checked
An AI spending stop? Watched constantly — every tripwire quiet; spending is accelerating, and the giants say demand still exceeds what they can build. China catching up? Checked: their chips cost ~30% more to make and sell only ~5-10% cheaper — they cannot start a price war, and they cannot touch HBM before ~2030. New chips that skip memory? Checked (Google's Frozen, AMD's Taalas): they serve only small, unchanging models — the fastest-growing part of AI, long-running assistants that remember, physically cannot exist without memory chips. The real risk is 2029-31, when all the new factories arrive at once — five dated checkpoints watch it, and Goldman's numbers say demand is still growing right through that window.
7 · What happens next — the dates
Aug 26: Nvidia results (guided $91B — your biggest customer's report card). Late Oct: SK's own Q3 — watch how much of the price rises reach the accounts. December: Nasdaq-100 inclusion window + the full shareholder-returns plan (the first slice is already running at ~1.7× its scheduled pace). Any day: a close above ₩1.744M flips the dashboard back to green mechanically.
8 · What history says — every shortage like this one, and what the stocks did
This situation has happened before, in memory itself and in other essential materials. The pattern is remarkably consistent:
Memory, 2016-18 (the direct ancestor)
prices +130% → SK Hynix ~4×, Micron ~6×
then −50-60% when new factories arrived
Lithium, 2020-22
material ~10× → producers ~5-6×
then −70% bust
Uranium, 2003-07
material 13× → Cameco ~10×
then multi-year bust
Shipping, 2020-21
freight 5-10× → ZIM ~8× + big dividends
then −70%
Memory, 2023-26 (this cycle, live)
server memory ~2× in a year → Micron +900% off its low
…in progress
📈 SHORTAGES PAY 3-6×
Driven by profits exploding, not excitement — multiples stay low the whole way up.
🧊 EVEN SKEPTICS SAY ₩2.7-3.6M
History's harshest rule (6-8× peak profit) still lands ~2× above today. No re-rating needed to win.
⚠️ THE BUST IS REAL — AND DATED
−50-70% when supply lands. Ours: 2029-31. Plan is to be repositioned before it, not ride through.
🏰 SOME ESCAPE THE CYCLE
ASML · TSMC · Nvidia turned shortage into moat via locked contracts — and re-rated permanently. SK is attempting exactly that migration.
This is the summary. The working machinery — the simulations, the sliders, the audit trail, every source — lives on the signals and engine pages, unchanged.
GREEN 8.3/10
Central case · 45% bear mass · three tripwires live
🧱 The three pillars — and what would change my mind
The bet in one line: the world has ordered more AI memory through 2030 than can physically be built — and SK Hynix makes more of it than anyone.
Three things that must stay true
1. Customers keep spending
TRUE Four hyperscalers guide $732.5B for 2026 (H1 actual: $301B); Goldman sees $7.6T cumulative through 2031 — and hyperscaler results are what ended the crash
2. Supply stays short
TRUE Suppliers meet only 75-80% of DRAM demand now, ~60% expected in 2027 — and Q2 printed it: DRAM prices +30% QoQ, NAND +mid-50%, 76% operating margin
3. SK keeps its lead
WATCH Still #1 at 56.4% of HBM, but down from ~69% a year ago — and now CXMT is listed, funded and being tested by Apple in commodity DRAM. The one metric trending against us.
What would change my mind
Any hyperscaler cutting its spending plan · two quarters of falling memory contract prices · the packaging shortage closing below ~5% · the share price below ₩1.744M and staying there. Any one of these fires and every number on this site re-bases automatically, without argument.
One company, two businesses — why we value the lanes separately
The structural test on the engine page proved the "TSMC-grade moat" analogy holds only for HBM — old memory becomes a commodity, old chip nodes don't. So the model treats SK as two lanes sharing one factory. 2027 revenue, as the engine sees it:
AI/server lane (HBM-led) ~₩266T
Conventional DRAM ~₩162T
NAND ~₩119T
The moat lane (HBM + server DRAM + eSSD — relabelled for precision when the Chain #6 cross-checks showed pure HBM is ~₩60-75T of it). Customer qualification locks buyers in the way tape-outs lock foundry customers in — verified structurally. LTAs, ~10 customers, 2027 pricing "progressing smoothly." CXMT cannot touch this lane before 2030.
The spot lane. Priced by marginal supply. Booming now (+30% QoQ) — but this is where CXMT's capacity lands from 2028, and where PC/phone demand destruction bites first.
NAND. 321-layer to ~50% of capacity by year-end; eSSD doubled in Q2 — and this lane has an AI roadmap: AI-N P samples end-2026, 100M-IOPS production end-2027, HBF with SanDisk 2027. The qualification moat starts reaching here too.
Half the company owns a fortress; the other half rides a boom. When you hear a bear case, ask which lane it attacks — most attack the spot lane and get reported as attacks on the fortress.
The conclusion — where this is going, in plain words
Everything on this site — four independent demand chains, the printed price tape, TSMC and ASML's sold-out order books, and the singularity scenario — boils down to one sentence (the company says constraints run to 2030): the world has ordered more AI memory through 2030 than can physically be built, and SK Hynix makes the most of it. Price rises to ration what's scarce; earnings follow price; the share follows earnings. The model's year-end path from today's ₩1.647M — with the street now arriving underneath it (Daishin ₩3.9M, Hanwha ₩4.3M bracket the 2027-28 waypoints):
End-2026
~₩5.5M (range 4.6-6.1M)
+~218%. The Q3 and October prints confirm the price surge is in the numbers; the crash-week fear washes out; multiple recovers toward 7× forward.
End-2027
~₩8.0M (range 6.0-8.0M)
+~328%. Kwak's "worst supply year in history" plays out: HBM4 contracts near-double, earnings roughly ₩350K/share, the structural story goes consensus.
End-2028
~₩8.4M (range 3.5-5.5M)
+~146%. Robots and agents extend the curve; earnings ~₩894K/share; the market starts watching the 2029-31 factory wave, capping the multiple below 10×.
End-2029
~₩7.9M (range 3.8-7M)
+~200%. The fork year — widest error bars on the page. If demand stays oversubscribed (the base case), price converges on today's fair value (DCF ₩5.5-5.8M): the model closes its own loop. If the factory wave lands into softening demand, 2029 marks the cycle top instead.
What kills this path, mechanically: a first hyperscaler capex cut · the packaging supply-demand gap closing below ~5% · two straight quarters of DRAM contract price declines · a weekly close back below ₩1.744M after clearing it. Any of these fires → predictions re-base to the conservative floor, on the record. Scenario D (the Vahdat curve going consensus, 18% probability) sits above every number here.
The debut worked; Monday flushed. SKHY closed day one at $168.01 (+12.8%), then Seoul answered with −15.4% — the worst day in company history — on zero fundamental disclosure: profit-taking, benchmark confusion, 2.5% new supply, and leveraged ETFs doing $9B of forced selling per 1% move. KRX ₩1.845M · ADR ~$152 · premium blew out to ~25%. One real watch item: NH says the Q2 HBM4 shipment step-up "does not appear to have materialised" — since verified. Kwak's call stands: 2027 = worst supply year in history. Engine recentred — oversubscribed-through-2030 is now base (Kwak, Nvidia $1T book, TSMC/ASML sold out, Citrini 28.7EB gap); 12-mo expected value ₩5.34M blended.
ADR / KRX
₩1.647M
4 Sep close, +3.2% — the first session on the Q2 share print, absorbed: foreigners #1 net buyer ₩527B, high ₩1,683K; tested support now ₩1,558K (3 Sep low). The ₩1.744M tripwire was tested 28 Aug (high ₩1,788K) and not held — ex-div close ₩1,653K; 31 Aug opened −4.4% on Warsh hawkishness + CXMT’s revenue print and closed +1.3% (₩1,674K), the buyback the breakwater. Prior week: intraday high ₩1,773,000 cleared ₩1.744M; the close did not. Buyback running ~1.7× schedule · company answered with ₩38-54T expansion + Q3 returns the same day · tripwire ₩1.744M
Entry
₩986K
Frankfurt GDR
Unrealised
+44%
vs entry (close) · was +122% at the peak
2028 fair value
₩5.5-5.8M
base case, 10×
Upside to FV
+202-219%
engine ₩5.84M ≈ SWS DCF ₩5.53M
Next event
Aug 26
Nvidia results ($91B guided) — then 2027 pricing settles in Q4. Old: Q2 print / HBM4-whisper test
Nvidia results — read the balance sheet firstThe number to read first is not the $91B — it is inventory and prepayments to suppliers. That line is where "stockpiling memory at any price" appears before it shows up anywhere else. Then: Rubin shipment guidance (KeyBanc models 1.9M units for the year, +250K next quarter) · HBM4 availability · any change to Rubin Ultra's stack count — a move from 16-high to 8-high would cut our FY28 memory content · and whether rising memory costs are denting demand.
Late Aug
Buyback paceRunning ~₩1.1T/day against a ₩0.65T even pace. If it holds, the ₩40T programme finishes in early October rather than November.
Sept
Shareholder-return frameworkSamsung has voted ₩90-110T for the year; proportionately SK lands near ₩40-55T. Watch for the announcement window.
Late Oct
Q3 results — the decisive printRevenue above ₩104T with HBM mix holding confirms the central case. Below ₩98T and the forecast steps down automatically. October HBM capture below 55% is the second tripwire.
Q4
2027 pricing settlesEdgewater: calendar-2027 memory pricing unlikely to be finalised until late 2026. The HBM4 ladder at $32-40/GB is indicative until then.
Dec
Nasdaq-100 reconstitution windowIndex inclusion would force passive buying into a float already shrinking through the buyback.
Feb 2027
Yongin first cleanroomCleanroom only — equipment installs from Q2-2027, and Y1 reaches 360K wafers/month by H1-2030. Not a 2027 revenue event.
Q1 2027
TrendForce share printDRAM revenue share below 34% trims the FY27 forecast — the third live tripwire.
2028
Competitor supply arrivesMicron's Boise ramps primarily in 2028 after qualification; M17 cleanroom Dec-2028; Yongin Y2 June-2029. This is when the wave starts.
Standing discipline: physical signals outrank narratives · inaction is a position (log it) · felt unease = mandatory verification · marginal Kelly on a maxed position ≈ zero · state risks once, lead with the confident case.