DATA BLOCKPrice ₩1,647,000As of Fri 4 Sep 2026 close · +3.2%(first session on the Q2 share print — absorbed; foreigners #1 net buyer ₩527B; high ₩1,683K)Verdict GREEN 8.3Central case · 44% bear-regime mass · tripwires live
THE SIMPLE VERSION
Earnings: ₩97T → 360T → 816T → 1,102T (2025–28) — ×2.27 in 2027 on bits +24% and weighted price +79%, then +35% in 2028. We price next year’s profit at 9.5–10.6× (peers 46–58×) → ₩5.5M → 8.0 → 8.4 → 7.9 → 8.5 at year-ends 2026–30. Twelve months out: average across every scenario ₩3.92M, most-likely ₩5.12M, worst 1-in-20 ₩2.15M. The street’s ₩3.0–3.2M targets are the same multiple on a 2027 plateau — that is the whole disagreement. Everything that has happened since the engine last ran is queued in the Re-run Register for 27 Oct; numbers move on actuals, not headlines.
Built from the ground up: chips sold × price per chip = revenue, for each product line. Then subtract the three real costs (making the chips, wearing out the factories, running the company) to get profit, profit per share, and a fair share price. Move any slider — everything recalculates instantly.
How to read this page in 20 seconds: pick a scenario → look at the green box (fair value vs today's price ₩1.647M (4 Sep close)) → open any product line to change its assumptions → watch the waterfall and charts update. Every number here is derived, not typed in. Calibrated to real results: 2025 actuals (revenue ₩97T, profit ₩43T), Q1 2026 actual (₩52.6T revenue, 72% margin), Q2 2026 actual (₩79.32T revenue, ₩60.54T OP, 76% margin — a ~5% miss vs the ₩83.0T consensus this line previously carried; reported EPS ₩131.5K includes the ~₩63.3T Kioxia one-off, so clean EPS is the calibration metric).
The recalibration audit — why the number barely moves, and where the spring is
Three weeks of verified bullish news, and the headline number moved little. The honest answer has two halves: what the evidence has now earned, and what the engine refuses to pre-pay. Both are listed, with the payoff schedule for the discipline.
What moved today — earned
P(Scenario D): 20% → 22%. The corroboration rule blocked Dwarkesh (same structure as existing evidence), but Chain #6 is a new instrument — an independent demand-side ledger whose three cross-checks landed. Downside branches: 13% → 10.5% combined. Two of the falsifier pathways now carry named, first-evidence mitigants (Baker's contract-roll + Goldman's margin-inflection, with Anthropic's 38→70% margin as the print), and the CXMT pathway was decremented by cost-structure data (prices 5-10% under the leaders at >30% higher cost = no price war). Net: fan median ₩3.72M, P(>today) 94%, P(>2×) 81%, p5 lifts to ₩1.59M, blended E[V] ₩3.92M.
What deliberately did NOT move — and why
The multiple, 9-10× — and it is the whole ballgame. Three structural legs against the old ceiling are now identified (ceiling-free LTA convexity · ~10 LTAs de-cyclicalising revenue · full-stack co-architecture), but a re-rating must print before the engine pays for it — in the market's willingness to hold clean earnings above 10×, in a returns announcement, in index flows. Also held: LTA damping (waits for the Q3 pass-through print >55%), Storage-Next (telemetry until samples ship, end-2026), and the D-branch top (greed guard). The arithmetic of stickiness: the headline is base-case EPS × multiple. Every verified story of the past three weeks strengthens the tails and the odds; almost none of it is allowed to touch the multiple until the market itself does.
The spring — a payoff schedule, not a prediction
If the multiple legs print, the same ₩375K of 2027 earnings reprices mechanically: at 10.5× → ~₩3.94M · at 11× → ~₩4.13M · at 11.5-12× → ₩4.3-4.5M — before any 2028 roll-forward. The dated triggers that could release it: the shareholder-returns plan (promised by Dec 31) · the Nasdaq-100 reconstitution window (December) · the Q3 print (late Oct) with its pass-through tripwire · a weekly close above ₩1,744,000. The engine pays for printed facts on delivery, not for narratives on announcement. That refusal is exactly why it was trustworthy through the crash — and why its jumps, when they come, are earned rather than imagined.
The outcome fan — 200,000 futures instead of one number
Transfer from the strategy-tester methodology: a single E[V] is one path; a bootstrap gives the distribution. Each draw picks a regime (2.3% severe failure — trimmed Aug 11 as the Wuxi 1a upgrade retired the China-fab disruption scenario — 7% falsifier-fires bear, else the base/Scenario-D world), then samples every uncertain input from a range instead of a point. The fan below is what the engine actually believes, tails included.
ROBUSTNESS & SIZING (SI-CTRL deep pass, Aug 12):2,000-configuration sweep: the doubling conclusion (median >2× today) survives 100% of configuration space — including a hostile sweep with deepened tails. The "worst 1-in-20 above today" line survives 37% under hostile tails: it holds on our calibrated tails (Wuxi de-risked, floors) but is calibration-dependent, and is hereby labeled so. Kelly with a 10% model-wrong regime: full allocation = +163%/yr expected geometric, P(loss) 11.9% — the sizing is not the risk; the path is. Decision number = the lower tail, never the median.
Median 12-mo outcome
—
—
P(above ₩1.647M — current)
—
the fan includes the failure branches
P(better than 2× today)
—
the asymmetry the position is built on
Thesis robustness score
—
share of full slider-space draws where fair value > price — the "tight spread" test
Path limitation (from the ctrl-session vet): the fan prices terminal outcomes, not the traverse — and real drawdowns cluster, which iid resampling understates. The ₩1,246,000 intraday print is the class of path event the fan cannot see; it is handled by the zero-leverage rule, not by this chart. The decision number is the lower tail, never the median. Honesty note: the fan's mean sits below the headline E[V] ₩3.86M because the headline handles thesis failure by re-basing when a falsifier fires, while the fan prices those branches in from the start — 2.3% severe (glut + macro, ₩0.9-1.6M — Wuxi de-risk trimmed this tail Aug 11) and 7% bear (falsifier fires, ₩1.2-2.4M) — bear tails lifted Aug 10: contractual price floors and take-or-pay change what a bad world can look like. Both statements are true; they answer different questions. The headline says "if the thesis holds, what is it worth"; the fan says "across everything that can happen, where do we land."
Transfer report — what the strategy-tester methodology gave this engine
Walk-forward out-of-sample
≡ our prediction log. Every dated prediction (P1-P5, the Vahdat telemetry, the Kioxia-optics trap) is an out-of-sample test the engine was never tuned on. The log is the walk-forward.
The six-filter funnel
≡ our falsifier set, with one upgrade adopted: predictions must now beat their base rate, not just "happen" — filter #3's lesson ("a Sharpe that's too good means the asset did the work, not the strategy") translated as: a tape prediction that comes true in a rising market may just be beta.
Bootstrap stress test
≡ the outcome fan above — 200,000 reshuffled futures instead of the single path, with the 5th-percentile answer on public display.
Parameter sensitivity
≡ the robustness score — the conclusion sampled across the full slider space rather than at the defaults. A thesis that only works on one magic setting is curve-fit; this one clears in the high-80s% of the whole configuration grid.
The ctrl-session's decisive catch — vetted here, independently
Walk-forward makes each individual test honest; it does nothing about choosing among 9,000. The moment out-of-sample results select the survivors, that out-of-sample set has become an in-sample selection set. The arithmetic, recomputed and confirmed: pure noise alone passes ~450 of 9,000 configs at a nominal 5% bar; the standard error of an annualised Sharpe on ~5 years of stitched data is ≈0.45, so their survival floor of 0.5 sits 1.1 standard errors from zero — and the multiple-testing-corrected bar is ≈1.7-2.0 (Bonferroni 1.97; expected max of 9,000 noise trials 1.91). Their 0.5-2.5 acceptance band admits a large region indistinguishable from luck. Pre-committed now, before we ever run the sweep on the real universe: the floor scales with trial count, a third never-touched holdout segment adjudicates, and the trial count is always reported beside the survivors. Why our own fan is immune: it selects nothing from thousands — one thesis, one distribution over stated uncertainty. Different animal, same family name.
What does NOT transfer (stated, not hidden)
The 49-family × ~9,000-backtest sweep itself needs a price-data feed this site doesn't carry — it's a Claude Code project for the trading universe, not a thesis-site feature. And their min-30-trades rule cuts against us: our technical desk has zero scored predictions, so per their own filter it carries zero decision weight until the log fills. Written on the desk itself.
0 · The physical supply ledger — where the numbers come from
These predictions are anchored to factories, machines and signed plans — things with dates you can verify — not to mood. A fab takes 18-24 months from construction start to volume output, so 2027-28 supply is already fixed by what exists today.
Project
What it makes
Volume from
What it means
M15X (SK, Cheongju)
DRAM / HBM
Open — ramping now
Feeds the 2026-27 shortage; already in today's prices
Cheongju P&T7 (SK, ₩20T)
HBM packaging
Late 2027
Lifts SK's own HBM4E assembly ceiling — funded by the ADR raise
Yongin Y1 (SK)
DRAM (1st of 4 mega-fabs)
From 2027; cluster → 2033
First big new cleanroom — initial volumes only in 2027-28
Indiana (SK, $4B)
HBM packaging (US)
~2028
Assembly next to US customers (CHIPS-funded)
M17 (SK, ₩80T)
NAND
Construction 2027 → ops H1 2029
Lands exactly at the start of the danger window
National ₩816T (Samsung + SK)
4 new fabs
2029-31
Korea plans to DOUBLE memory capacity in 5 years — state-funded, so it gets built even if prices fall
CXMT (China)
Commodity DRAM
Scaling now (3%→8% share)
Pressures commodity prices from 2027; no HBM capability
Nittobo T-glass (JP: 3110)
Low-CTE glass cloth → AI substrates/PCBs
~90% world share; tripling capacity → relief mid-2027
Gates how many AI systems can be built — a floor under scarcity pricing until mid-2027
Mitsui Kinzoku foil (JP: 5706)
HVLP copper foil → CCL → AI boards
One of ~5 non-Chinese suppliers; Malaysia exports +91% YoY (May)
Same story one layer down — physical demand telemetry for our thesis
TSMC COUPE photonics
Co-packaged optics (light replaces copper between chips)
500 → 10K wafers/mo (Q2'26) → 15K (Q4'26) → ≥25K by 2028; yields still 20-50%
Widens the data pipe between chips from 2027 — a demand multiplier for memory, arriving slower than hype (MS: 23K CPO units in 2026 vs 200K once hoped)
TSMC CoWoS
AI-chip packaging
130K → ~200K wafers/mo by 2027
The pipe that pulls HBM demand — allocated years ahead
What the ledger proves — this is why the defaults aren't a mood: (1) 2026-28 tightness is physics. No greenfield fab can reach volume before H2 2028, and every HBM wafer eats ~3× the capacity of a commodity wafer — the shortage is structural, which is why Scenarios A and B share the same 2027-28 chip-growth numbers. (2) The genuine fork is 2029-31, when M17, Yongin 2-3, Samsung's national fabs and CXMT's IPO-funded lines all switch on together — now joined on the NAND flank by YMTC's $4.9B IPO (targets passing Samsung/SK in NAND by end-2027; #3 in bits, #5 in revenue) — now with ₩816T of state money behind doubling capacity. The materials layer (new): below the chip sits a second set of bottlenecks — T-glass cloth (Nittobo, ~90% share) and advanced copper foil (Mitsui) that every AI substrate and server board needs. These don't stop SK making chips; they cap how many systems the world can assemble per quarter. In a shortage that's margin-supportive (scarcity everywhere = pricing power everywhere, feeding the uncapped LTAs) while trimming the ceiling on unit growth. Their relief arrives mid-2027 — conveniently right when our 2027-28 volume ramp needs the pipe to widen. (3) Photonics is a demand lane, not a threat: TSMC's 30× optical-interconnect ramp (COUPE) un-blocks the wires between chips from 2027 — and faster wires expose the memory bottleneck harder. It also makes optically-pooled memory racks practical, which adds HBM and SOCAMM-class demand on top of local stacks. Three pipe-widening events now converge on 2027: T-glass relief, CoWoS→200K, COUPE ramp. Fact-weighted scenario odds: ~25% A · 50% B · 25% C. The LTA leg just hardened: SK's new contracts have floors but no price caps (unique among the big three — Micron caps at Q2-26 levels) plus 10-30% prepayments. Let the Jul 29 earnings-call LTA detail move the weighting further — not headlines. Demand-side anchors refreshed 27 Aug (NVDA Q2 FY27): FY28 guided +70% (~$691B) → Blackwell+Rubin envelope ~$1.2T; top-5 capex $800B → $1.3T in 2027 (backlog >$2T) — above the demand inputs used here; $/GW ladder $18→$25→$40 (Hopper→GB→Rubin), memory content est. 25–35% → FY28 ≈17GW ≈ $170–240B memory via NVIDIA alone; NVHBM extends custom HBM into Trainium "in partnership with memory suppliers"; GM 75.0% absorbed memory cost headwinds. And a fourth mega-buyer outside all of it: SpaceX ~10GW in 2027 (~$300–500B capex, SemiAnalysis), NVIDIA-exclusive, funded by inference economics of >$100B/GW/yr revenue vs ~$12B/GW/yr cost — the mechanism that keeps every demand input on this page conservative. Realised, not forecast (2 Sep): HBM export unit value $76.13 in July (+9.5% MoM, +47% Q1→Q2), DRAM +24.3%, with volumes falling and values holding — price and mix carrying revenue exactly as the chain assumes. Full ledger → Demand page.
1 · Pick a starting world
Each scenario pre-fills every slider. You can then adjust anything.
2 · The four product lines
Revenue for a year = last year's revenue × (1 + more chips sold) × (1 + price change). Tap a line to open its sliders. Showing assumptions for:
3 · Costs, tax, shares, multiple
The company-wide settings that turn revenue into profit per share and a fair price.
4 · From chips to share price — the waterfall
SI-CTRL v11.14 — second run, Aug 2 pm: four findings not previously on this site
Same engines, fresh pass over the post-print state. Rule for inclusion: only what changes a number, a falsifier, or a process. Framework IP: Sense2 Pty Ltd.
1 · E13 on the ceiling removal: SK's revenue line is acquiring convexity
Volume-locked, price-floating LTAs are, structurally, a portfolio of written supply commitments with retained price optionality — the customer owns certainty of bits, SK keeps the upside on price. Transfer test: this homomorphs onto the take-or-pay structures LNG producers wrote into the 2021-22 squeeze (guaranteed offtake, uncapped price) — verified at L3 on revenue mechanics. Consequence nobody has priced: the revenue distribution becomes right-skewed with a cushioned floor — downturns hit price but volumes are contracted, upturns pass through uncapped. A producer whose earnings distribution has a floor and an open top deserves a higher multiple than a symmetric-cyclical — this is the first structural (not narrative) argument against the 8-10× ceiling, and it strengthens quarter by quarter as LTA coverage rises toward 25-30% of revenue. Instrument added: track LTA share of revenue every quarter; each 5pts of coverage is roughly half a turn of justified multiple.
2 · E4 conservation: the glut risk didn't shrink — it moved onto the customers' books
Bear row 7 (the 2029-31 capacity wave) implicitly assumes new supply lands into spot demand. But ceiling-free LTAs plus customer prepayments mean a growing share of the new fabs' output is pre-sold at floating prices before the concrete cures. Conservation holds: the oversupply risk was not destroyed, it was transferred — from SK's income statement to the hyperscalers' procurement costs. Model action: bear row 7's severity is now conditioned on LTA coverage — if contracted share reaches ~30% by end-2027, the wave lands substantially on committed demand and the row de-rates a full grade. Same telemetry as finding 1; one number now governs two rows.
3 · Evaluator audit: one of our falsifiers measures success, not failure — re-specified
The "ADR premium <3%" falsifier was written when the premium was feral (30-40%) as a check against listing-hype collapse. But the thesis itself — KRX-obscurity mispricing closes as US access arrives — completes with the premium going to zero. As written, the instrument fires at the moment of victory: a completion detector wired as an alarm. Re-specified: premium <3% is bearish only if accompanied by collapsing ADR volume (access failing); on healthy volume it is the win condition. Unexpected class of error worth remembering: a thesis's own success can degrade its falsifiers — every falsifier now gets an annual "does winning trip this?" review.
4 · E12 consolidation: all five verified catches were the same failure, and it wasn't reasoning
Pattern across the five "unease" verifications (earnings date, stale prices, consensus mismatch, verdict-chip staleness, stale file lineage): every one was an ingestion-freshness failure; zero were logic failures. The organism's weakest organ is data currency, not inference. Process patch adopted: every load-bearing number on this site now carries an implicit last-verified date, anything older than 7 days is treated as unverified at decision time, and session starts re-stamp price, verdict, and calendar before anything else. The gap this closes is exactly the gap all five catches walked through.
One micro-prediction, logged for sport (E14 reflexivity): our tripwire (₩1.744M) sits on the same shelf as the momentum systems' 200-4H EMA — the line is now shared infrastructure across unrelated strategies. Clustered re-entry orders live just above it, so the →GREEN crossing, when it comes, should print as a gap or acceleration on elevated volume, not a drift. Falsifiable within weeks; costs nothing; teaches us whether the map has become part of the territory.
SI-CTRL v11.14 run — three engines, three findings the ordinary read misses
Structure algebra (E13) × gas lens (E14) × recursion, applied to the thesis after the print. Framework IP: Sense2 Pty Ltd.
E13 — the transfer that quietly underwrites this thesis, tested
The site has been running an untested analogy: SK's HBM position ≈ TSMC's node-leadership moat. Signature both sides. TSMC: carrier = process nodes; operations = shrink and allocate; distinguished element = the leading node (absorbing — it takes all high-margin demand); relation = customer lock-in through tape-out cost. SK: carrier = memory products; operations = stack-height increase and allocate; distinguished element = HBM4; relation = lock-in through qualification cost. Map φ: node → stack generation, tape-out → qualification, foundry share → HBM share.
Homomorphism test. Generic case passes: φ(shrink ∘ allocate) = φ(shrink) ∘' φ(allocate) — both preserve the ordering that leadership captures margin. Boundary case fails. TSMC's carrier is monotone: N-1 and N-2 nodes stay differentiated and profitably priced for years. SK's carrier splits — last generation's memory is not a mature node, it is a commodity whose price is set by marginal spot supply that SK does not control.
Obstruction report: φ is a homomorphism only on the HBM sublane. Depth L4 on HBM (verified: qualification cost creates the same switching hysteresis as tape-out cost), L2 on conventional DRAM (metaphor only — no structural claim licensed).
What this licenses and forbids. Licensed: modelling HBM share persistence and pricing power on foundry-leadership logic; expecting the HBM lane to hold a premium multiple. Forbidden: applying a foundry-style multiple to the whole company; assuming commodity DRAM inherits the moat. Actionable consequence — value the two lanes separately. And note what falls out for free: this is exactly why CXMT lands where it does. It attacks the lane where the homomorphism had already failed, and it cannot attack the lane where it holds. The obstruction predicted the threat's address before the threat arrived.
E14 — gas lens on the week that made no sense
Macrostate exists? Millions of holders, yes — but the validity gate is weak coupling, and that is precisely what fails here. Order parameter: leveraged-product flow per 1% move ÷ available liquidity. It changes kind at the boundary — below 1, price is set by opinion; above 1, price is set by mechanical rebalancing that has no opinion at all. Korea's single-stock leveraged complex pushed that ratio through 1, and the correlation between holders went to ~1 under stress. The gas condensed. That is the technical description of Jul 28-30, and the +55% intraday move in the 2× product on Jul 31 is the same mechanism unwinding.
Renormalisation (R2), and this is the finding. Coarse-grain the week: daily amplitudes were −14.65%, −9.6%, −5.6%, +29.95% — violence at every step. The weekly net is −5.4% (₩1,816,000 → ₩1,718,000). The cascade term exists at daily resolution and vanishes entirely under coarse-graining: same form as an ordinary week, flowing parameters. Emergence boundary located — the cascade is a daily-scale phenomenon with no weekly-scale counterpart. Falsifiable consequence: for a thesis whose inputs (fab schedules, contracts, capex) move on quarterly clocks, daily price carries near-zero signal and maximal noise, and the correct observation resolution is quarterly. Wrong if a daily move ever propagates into a quarterly fundamental — which is exactly what the falsifier list already watches for.
Ergodicity — the part that decides the ladder. Time-average ≠ ensemble-average when ruin absorbs. The intraday low of ₩1,246,000 was −43% from the stamp, printed on the day of the best operating quarter in company history. An unleveraged holder experienced that as a number on a screen; a leveraged one experienced it as liquidation. The absence of leverage is not caution — it is the mechanism by which this position survives its own volatility. At 100% concentration, with the order parameter above its phase boundary, marginal Kelly is negative, not zero. The ladder should be retired or converted to an explicit written no-trade, and this is now the structural reason rather than a preference.
E4 — conservation audit, and a scheduled trap nobody has priced
The ₩63.3T Kioxia gain is a transfer, not a creation: an asset left the balance sheet and cash arrived. No operating value was generated, so it cannot appear in any run-rate — hence the clean-EPS discipline above. But conservation has a second consequence that is predictable and dated: in Q2 2027, net profit will be compared against ₩93.92T. Even with operations growing strongly, the year-on-year net-profit line will read as a collapse, because the prior-year base contains a one-off worth two-thirds of it. A scheduled optical shock, twelve months out, entirely mechanical. Logged as a prediction with a review date: expect headlines reading "SK hynix profit falls sharply" in late July 2027 on operating results that are, in fact, up. Wrong if the company restates or pre-frames the comparison in its Q1 2027 materials — which, given how carefully this management pre-frames, is itself worth watching.
5 · The full path, 2025 → 2030
Solid history, then your assumptions. The dashed line on the price chart is today's price — the gap to the bars is your upside or downside.
Street cross-check (Jul 9): three independent answers to "what is this worth" — this engine's 2028 base case ₩5.84M (bits × prices × costs, 10× multiple) · SimplyWall.st's discounted-cash-flow ₩5.53M (their model, their assumptions) · 37 analysts' 12-month average target ₩3.21M. The first two were built completely separately and land within 6% of each other — that convergence is the anchor. The analyst number isn't wrong, it's answering a different question (12 months out, not 2028) with LOW agreement (spread >15%). Street forward EPS ₩355.65K vs engine 2027 ₩434K — engine ~20% above street, the size of the uncapped-LTA effect the street hasn't fully modelled yet. Debut-day validation: LSEG's analyst poll now puts 2026 revenue at ~$235B ≈ ₩335-355T — squarely on this engine's ₩335T base case. And the ADR adds a live cross-check: SKHY's first US close ($168.01) values the same share at ~₩2.53M — a 16% premium the American market is paying today over the Seoul price this page conservatively uses.
Recentered — Jul 18, 2026: oversubscribed-through-2030 is now the base case
The evidence stack got too tall to keep below the base: Kwak on the record ("demand above supply beyond 2030"), Nvidia's ≥$1T cumulative Blackwell+Rubin orders through 2027, TSMC's CoWoS and N3 sold out with lead times into 2027, ASML's 2027 book effectively full with large 2028 orders already placed, Gerstner positioned "durable through 2029," Baker's 30-40%-of-capex call, MS's memory wall, and Citrini's 28.7-exabyte 2030 supply gap even if every announced fab including China's comes online — and now the company itself, on the record Jul 22: DRAM constraints expected to persist to 2030, long-term contracts being signed to damp volatility, and the board accelerating P&T7 with ₩7.09T six days before earnings. Accordingly: the engine's base now runs on the printed-price extrapolator (₩357T '26 / ~₩543T '27), the prior hand-set lines (₩335T / ₩484T) become the conservative floor, and the 2029-31 oversupply wave moves fully to the risk case. The discipline stays: this recentering dies — mechanically, back to the floor — on any of: a first hyperscaler capex cut, the CoWoS supply-demand gap closing below ~5%, or two consecutive quarters of DRAM contract price declines.
Q2 print vs the model — scored honestly, Aug 2, 2026
The first hard calibration since the engine was built. Three assumptions were right, one was wrong in our favour, and one piece of the recentering was too hot. All five are recorded before any of them are used to justify anything.
LTA price damping
VALIDATED — almost exactly The extrapolator assumed 50% of volume reprices each quarter, with the rest damped by long-term agreements. Q2 printed contract DRAM up ~63% and realised ASP up ~30% — a pass-through of 48%. The single most uncertain number in the model, confirmed within two points by the company's own disclosure.
Net margin
TOO LOW — corrected upward The engine used 50% (2026) and 47% (2027). The actual operating margin printed 76%, which taxes down to a clean net margin near 58%. Raised to 57% / 53%. Consequence: 2026 clean EPS moves ~₩228K → ~₩268K, and the multiple at ₩1,718,000 falls to ~6.4×. The audit lowered four things a week ago; this print raises one, and both directions get logged.
Revenue trajectory
SLIGHTLY HOT H1 actual is ₩131.9T. With Q3 guided to +10% DRAM bits on still-rising prices, the full year lands near ₩345-355T — inside the extrapolator's ₩357T but at its lower edge, and the recentering should be read as the top of the range rather than the centre. The conservative floor (₩335T) is now formally too low, so the corridor tightens to ₩345-357T.
The HBM4 whisper
SCORED — TRUE, as timing Logged Jul 13 from a single NH analyst and carried as the one real fundamental question for sixteen days. It printed: the miss is attributed to HBM4 shipments below expectation, deferring revenue into later periods. The model does not lose the revenue; it moves it right. H2 now carries a ramp the company has explicitly committed to.
Capex / D&A
CONFIRMED AS RISK Capex raised to the high-₩40T range with M15X pulled forward and Yongin Phase 1 cleanroom in early 2027. The audit's D&A slider was added a week early and is now underwritten by the company's own guidance. The 2029-31 depreciation wave is not hypothetical — it is being funded now.
Net of everything: the model survives its first real test, with one upgrade and one trim. Clean FY26 EPS ~₩268K at ₩1,718,000 = 6.4×; the 2027 line moves to ~₩380K on the corrected margin, which at 9-10× brackets the ₩3.4-3.8M waypoint the conclusion already carries. No target was raised on the strength of a good quarter — the margin correction is offset by the revenue trim and the confirmed capex wave.
Bottoms-up: the Nvidia demand chain — how much SK Hynix likely gets
Ground-up, units × content × price. Anchors: Jensen at the MS roadshow — quarterly revenue "approaching $100B," Q2 guide ~$91B, ≥$1T cumulative Blackwell+Rubin orders through 2027; Rubin carries 288GB HBM4 (8 stacks), Rubin Ultra 512GB+ HBM4E; Goldman sees HBM4 pricing ~doubling into 2027 ($4-5/Gb). Slide the physical inputs; the chain does the rest.
Nvidia's share of world HBM demand (AMD, TPUs, Trainium, ASICs take the rest)
SK Hynix share of HBM (audited down Jul 25: Counterpoint 56.4% Q1-26, from 69% a year ago; Micron ~21% and rising; Samsung unveiled HBM4 Jul 20. Evidence points down, so the default follows it.)
Nvidia's HBM bill, 2027
—
units × GB × 8 × $/Gb
World HBM market implied
—
grossed up for non-Nvidia
SK Hynix gets
—
—
Capacity is the binder, and that's the point. The chain above is demand. SK's supply is gated by TSV/MR-MUF packaging (P&T7 arrives late-2027) and was sold out for 2026 before the year began — Kwak's on-record call is that 2027 is the worst supply year in the industry's history. When chain-demand exceeds SK's buildable bits, the surplus doesn't vanish — it shows up as price (the $/Gb slider), which is exactly Scenario B's mechanism. Reconciliation with the engine, stated honestly: at defaults this chain yields ~₩130T of 2027 HBM-stack-only revenue; the engine's HBM line runs hotter because it also carries what the chain excludes — SOCAMM2, custom base-die value, and the price upside if Goldman's $5/Gb prints (move the price slider to 5 and watch the gap close). Two independent paths, one corridor. Cross-check from the demand side: Morgan Stanley's Moore, after meeting Jensen — Nvidia's constraint list starts with "memory."
Assumptions audit — Jul 25, 2026: every load-bearing number, challenged
A model that only gets revised upward is a mood, not a model. This is the full audit; four fixes cut the numbers, and they have been applied to the sliders above rather than argued away.
Share count
FIXED — cuts EPS 2.4% Engine used 728M shares and ignored the 17.79M new shares listing Jul 29. Now 746M everywhere. Offsetting truth, not modelled: the ₩38T+ of net proceeds funds the fabs that produce the future earnings, so the dilution buys the growth.
SK's HBM share
FIXED — cuts chains ~7% Defaults ran 58-60% and had been nudged up. The printed evidence runs the other way: Counterpoint has SK at 56.4% in Q1-26, down from ~69% a year earlier; Micron ~21% and climbing; Samsung unveiled HBM4 on Jul 20. Defaults now 56%. Share loss in a market growing this fast still means rising absolute bits — but the model should not flatter the trend.
The D&A wave
FIXED — cuts 2027 EPS ~8% Margins were held at 50%/47% with no explicit depreciation from the record capex (P&T7 ₩19T + Korean facilities ₩45.5T + EUV ₩11.9T). A new slider subtracts incremental D&A directly (default ₩20T by 2027). This is the "earnings can peak while revenue still rises" mechanism, now visible instead of assumed.
Currency
FIXED — was invisible ₩1,500/USD was hardcoded in three places with no sensitivity. ~85% of revenue is dollar-linked, so a won rally to 1,300 cuts won revenue ~11% and the won share price with it. Now a slider. Korea's AI boom is itself a won-strengthening force — this risk grows as the thesis works.
Chain independence
OVERSTATED — corrected The Nvidia and TSMC chains were called independent. They are independent measurements (order book vs packaging lines) of a shared underlying driver — GPU volume. Real corroboration, weaker than claimed. The ASML chain (tool shipments) and the price extrapolator (contract tape) are the genuinely independent ones.
Terminal multiple
BIGGEST SINGLE LEVER 8-10× carries more of the fair value than any demand assumption. Sanity check now stated: each year-end waypoint equals roughly 9-10× the following year's EPS (end-2027 ₩3.9M ≈ 9.0× 2028E ₩410K). Internally coherent — but if the market ever pays cyclical-trough multiples of 4-5× on peak earnings, every target halves regardless of the demand being right.
Price elasticity
HELD LTA damping at 50% per quarter is conservative against a tape printing +90%/+63%/+20%. Long-term agreements heading to 25-30% of revenue support it. Sep flags, upgraded Aug 3 (verified): the Q3 forecast stack now reads TrendForce 13-18% · Samsung asking ≤20% (third straight double-digit quarter; ~340% cumulative if it lands) · UBS +32% Q3 and +18% Q4, with Goldman calling the imbalance the worst in 15 years. Default lifted +15% → +20% (Q4 +8% → +10%) — mid-range by rule, still well under UBS — and SK has reportedly removed price ceilings from its LTAs, so damping bias stays downward; tripwires: Q3 contract print >25% or pass-through >55% → revisit both.
Bit-supply ceiling
HELD, well-instrumented 45%/yr is funded by the ASML transmission card and SK's own ₩11.9T EUV order through 2030. Sub-linear tool-to-bit conversion (more EUV layers per wafer) remains the strongest structural floor under pricing.
What was missing entirely
ADDED as bear row 13 China exposure (Wuxi DRAM / Dalian NAND under US equipment-licensing rules), FX, and governance overhang (the chairman's ₩944B divorce asset ruling). None of these were anywhere in the model. They are now on the bear file with tripwires.
Net effect of the audit: 2027 EPS falls from ~₩350K to ~₩315K (dilution + D&A), and the chains trim ~7% on the share correction. The forward multiple at ₩1.83M moves 5.2× to ~5.8× — still less than a third of the peer average, which is why the conclusions survive the cut. An audit that never lowers anything isn't an audit. This one lowered four things and the thesis still clears.
Printed-price extrapolator — compounding what's actually on the tape
Honesty note (Jul 16): the bottoms-up chains used current price levels; this card compounds the printed percentage sequence forward. On the tape: DRAM contracts +90-95% Q1'26 (TrendForce), +63% Q2, MS/TrendForce Q3 forecast +13-18% — and MS notes memory prices are up ~6× in a year. Not every won reprices instantly: long-term agreements damp the realized blend, and SK shifts bits toward HBM. Both are sliders.
Q4'26 contract move (QoQ — after +90/+63/+15 printed)
2027 average QoQ drift (Citrini: ASPs elevated "for years")
Share of volume repricing each quarter (LTAs damp the rest)
HBM ASP growth 2027 (Goldman: contracts ~double)
Conventional DRAM bit growth (bits shifting to HBM)
HBM bit growth 2027
Audit add: incremental annual D&A by 2027 (the capex wave depreciating — P&T7 ₩19T + facilities ₩45.5T + EUV ₩11.9T)
Audit add: KRW/USD (engine assumed 1,500 — ~85% of revenue is dollar-linked)
2026 revenue, printed-price basis
—
—
2027 revenue, extrapolated
—
—
EPS and multiples at ₩1.647M
—
—
What compounding the tape shows: at the printed sequence with realistic LTA damping, the engine's hand-set lines (₩335T '26 / ₩484T '27) are conservative, not aggressive — the prints already run ahead of them. If the tape merely holds, the engine is under-clubbed; if the tape mean-reverts, the engine is roughly right. Falsifier: two consecutive quarters of QoQ contract declines before end-2027.
Bottoms-up #2: the TSMC packaging chain — same answer by a different road
TSMC's record Q2 (Jul 16: $39.6B, +36%, above its own guide; CoWoS + N3 sold out into 2027; capex $52-56B at the high end, US expansion accelerating) lets us count HBM demand from the packaging side: every CoWoS wafer exists to marry logic to HBM stacks. Wafers × packages × stacks × GB × price — no Nvidia assumptions anywhere.
AI packages per wafer (blended big-interposer mix)
HBM stacks per package (blended — GPUs 8, switches/small parts 0-2)
GB per stack, 2027 blend (HBM4 48-64GB, 12→16-hi — HBM4E now sampling, Jul 22)
HBM $/Gbit
SK share
HBM stacks the world packages, 2027
—
via TSMC's lines (plus small non-TSMC share)
World HBM market implied
—
—
SK Hynix gets
—
—
Bottoms-up #4: the ASML transmission — machines today, bits and demand tomorrow
ASML's Q2 (Jul 15): €9.3B sales and 54% margin above guidance, FY26 raised to €43-45B (from €36-40B at the start of the year), Q3 guided €11-12B. The lines that transmit to SK: memory-segment revenue set to grow 75% in 2026 (logic +25%, EUV +45%); ~65 low-NA EUV systems in 2026, +30% capacity for 2027, another +30% under investigation for 2028; 2027 is "close to all the orders we need" — effectively sold out a year out; CFO Dassen: DRAM demand is HBM+DDR volume and nodes using more EUV layers per wafer. Machines convert to revenue 12-18 months after order — this card is 2027-28 made visible today.
EUV systems shipped 2027 (65 × 1.3 = ~85)
Share of EUV going to memory makers (memory +75% says it's rising)
SK's share of memory-maker EUV intake (SK = lead EUV DRAM adopter)
EUV layers per DRAM wafer (1c/1d intensity — rising)
Share of new tools = truly incremental capacity (rest converts existing nodes)
SK EUV intake, 2027
—
scanners landing on SK floors
Incremental SK DRAM capacity enabled
—
—
What it feeds
the 45% supply slider
Scenario D's bit-growth ceiling is physically funded here
Three transmission lanes, one subtlety. Lane 1, demand: ASML→TSMC/Intel logic tools become AI accelerators that must attach HBM (logic +25%, EUV +45%). Lane 2, SK's supply: the +75% memory-tool surge is SK/Samsung/Micron's 2027-28 bit capacity being installed now. Lane 3, telemetry: 2027 effectively sold out and 2028 orders already large — the Vahdat prediction row's capex telemetry, printing a year early. The subtlety Dassen handed us: rising EUV layers per wafer means tool spend converts to bits sub-linearly — a chunk of that +75% buys intensity, not volume. Cost-per-bit rises industry-wide, which is a structural floor under pricing: the supply wave of 2027-28 arrives pre-throttled by its own lithography bill. Discipline, enforced by physics and Veldhoven's order book.
Bottoms-up #6 — the token ledger: demand counted in the industry's newest unit
The unit migration completed its arc this quarter: chips → dollars → gigawatts → tokens. Tokens are the demand quantum under every other chain — and for memory they are literal read-events: every generated token requires the model's active weights and its growing context to be read out of HBM. Decode is bandwidth-bound; token growth is HBM demand growth. Verified anchors: ~370T tokens/day today (~135Q/yr) — already 2.4× Dell's revised 2028 forecast, a forecast Dell had already revised up 57-fold; Goldman (May 2026): 47Q/month by 2028 (~565Q/yr) and ~120Q/month by 2030 — ×24 on agentic adoption; Google alone at 3.2Q/month, 330× in two years, 7× YoY, Gemini at 900M users.
Tokens per day today (trillions) — I/O Fund estimate 370T, floor 300T
Peak serving rate, tokens/s per GPU (frontier-optimized)
3000
Real-world utilization of that rate (SLAs, small batches, idle, redundancy)
25%
Fleet share on training / fine-tuning (not serving)
35%
Efficiency gain per year (Goldman: cost/token falling 60-70%/yr ≈ ×1.6-1.7 throughput)
1.65×
HBM content per GPU, 2028 vs today (Rubin 288GB → Rubin Ultra class)
2.0×
Serving fleet implied today
—
cross-check: installed base ~10-15M H100-class ✓
Implied AI revenue at ~$1.2/M tokens
—
cross-check: lab + cloud AI run-rates ~$150-250B ✓
Tokens 2028
—
cross-check: Goldman 565Q/yr ✓
Serving HBM bits, 2028 vs today
—
after the efficiency sponge · before training, ASICs and build-ahead stack on top
What this chain is for — and honestly not for. It anchors the demand slope from the consumption side with three independent cross-checks that all land (fleet ≈ installed base, implied revenue ≈ lab run-rates, 2028 tokens ≈ Goldman). It is deliberately not an absolute HBM-TAM claim — training clusters, ASIC accelerators and build-ahead inventory sit on top of serving bits, and the absolute sizing lives in chains #1-3. The finding it contributes: even after conceding Goldman's own 60-70%/yr cost-per-token collapse — the efficiency sponge at full strength — serving memory demand still ~2.4×'s by 2028, because tokens compound faster than efficiency and each package carries more HBM. Three multipliers stack: token count (Goldman ×24 to 2030, agents "the large majority"), tokens-per-task (agentic chains vs single answers), and context/KV growth (agent memory read on every step). The Jevons receipt is already printed: Google's cost per token fell ~60-70%/yr and its token volume rose 7× in the same year — spending rose anyway. And Goldman's Schneider closes our bear file's biggest loop unprompted: "If you raise gross margins, you raise operating cash flow, and that gives you more headroom to spend" — the margin-inflection mechanism, now a second named source on the row-12 mitigant.
The elasticity bridge — from token demand to share price in five steps
The economics that connect Chain #6 to a valuation. Memory in shortage is price-inelastic: it is a small share of a system's cost but the system cannot ship without it, so when demand outruns supply, price — not volume — clears the market. The bridge below computes the equilibrium price rise from the demand-supply gap, then walks it through revenue, profit and the multiple. Every step carries a printed cross-check.
DRAM bit demand growth %/yr (AI-led; SK guides mid-20s, Chain #6 pushes higher)
28%
DRAM bit supply growth %/yr (wafer + tool locked; HBM eats 3× wafers)
18%
Demand price-elasticity |ε| (lower = more desperate buyers; shortage era ≈ 0.3-0.7)
0.50
HBM demand-supply gap %/yr (Chain #6 serving ×2.4 + training stack vs capex-locked supply)
13%
Realised pass-through after LTA damping (validated 48% at Q2; ceilings coming off)
The convergence worth noticing: this bridge is the third independent route to the same destination. The printed-price extrapolator (compounding actual contracts) says ~₩3.6-3.7M at 12 months; the outcome fan (200K sampled futures) says median ₩3.70M; and now pure supply-demand economics — a demand gap divided by an elasticity — lands ₩3.5-3.6M at 12 months and ₩4.1-4.6M at 24, brushing the 2028 anchor of ₩5.54M from below. Three methods, three data diets, one neighbourhood. Honesty box: the elasticities are estimates (hence sliders, range 0.2-0.9) — and note the insider bound logged Aug 7: Musk claims demand +200%/yr against supply +20%/yr, far above our 28/18 defaults; the bridge is a coherence check on the other instruments, not a fourth E[V]; and the same arithmetic runs in reverse — when Yongin and M15X supply arrives 2029-31, the gap term can flip sign and elasticity cuts price down as fast as it cut up. That reversal is bear row 7, conditioned on the LTA coverage telemetry — and from Aug 7 it couples with bear row 14: etched-weight silicon (AMD-Taalas) capping the stable-segment price term is most dangerous exactly when the supply wave lands. Neither is a 12-month variable; both are on dated checkpoints.
Audit catch, Aug 2: the $/Gb slider was ~2× rich — printed contracts now anchor it
Scrutiny pass finding. The demand chains carried a default of $4.50 per gigabit — inherited from early spot-market chatter. Two printed anchors now exist and they agree with each other, not with the default: reported HBM4 per-stack contract pricing in the mid-$500s (12-hi, 36GB → ≈ $1.9/Gb, and >50% above HBM3E's mid-$300s), and BofA's 2026 HBM market estimate of $54.6B, which back-solves to ≈$2/Gb at plausible unit volumes. Action taken: chain defaults cut to $2.50/Gb (a blend allowing for HBM4E/16-hi premium SKUs), slider floors lowered to $1.50. What this does and does not change: the chains are cross-checks and floors — their implied SK-HBM revenue tightens toward the gigawatt ledger, and the whole family now sits consistently below the extrapolator. The extrapolator — which never used $/Gb, only SK's own printed revenue and printed price moves — is the validated primary line (48% pass-through confirmed at the Q2 print) and is unaffected. The audit rule held: when a printed number disagrees with an assumed one, the printed number wins, even when it cuts our own chart down.
Bottoms-up #5: the gigawatt ledger — because the buildout is now announced in watts, not chips
Every deal this week was denominated in gigawatts: SK Telecom 2GW, Naver 55MW→200MW→1GW, Japan 140MW, SK Group's stated 20GW ambition. Power is the honest unit — it is the one thing that cannot be exaggerated on a press release, because it must be physically delivered. This chain converts announced watts into HBM demand, and it is deliberately a floor: it counts new-build capacity only, ignoring refresh cycles in existing datacentres.
New AI datacentre capacity energized per year, 2027-28 (GW)
The announced ledger — what is actually on the books (Jul 24-27)
SK Telecom / NVIDIA
2 GW
Vera Rubin DSX AI factory, first online 2027. LOI. ~$10-11B of SK memory content by our rack build-up.
NAVER / NVIDIA / Brookfield
55MW → 200MW → 1GW
GAK Sejong. 55MW in 1H-2027, 200MW by 2028, gigawatt ambition long-term. $10B stack: Brookfield up to $9B (nonbinding term sheet), NVIDIA $1B equity. ~100,000 GPUs at 200MW — the calibration datapoint above. Vera Rubin + Blackwell.
NVIDIA Japan
140 MW
Vera Rubin AI factory expected 2028, alongside the physical-AI coalition (Toyota, Fujitsu, Fanuc, Kawasaki Heavy, Kioxia).
SK Group ambition
20 GW
The datacentre plan MaaS is tied to. Aspiration, not contract — excluded from the slider default.
Korea total, US partnerships
~$950B
Including Samsung-Broadcom ~$200B for memory and 2nm. Korea is being wired as an AI production base, and SK sits inside most of it.
Read it as a floor, and note what the floor does to the argument. At defaults this ledger produces roughly half the HBM demand of chains 1-2 — as it should, since it excludes the refresh cycle, the non-Nvidia ASIC fleet outside these projects, and every unannounced hyperscaler build. The useful conclusion is not the number but the consistency: a completely different unit of measurement (watts) lands inside the same corridor as order books and packaging lines. The top-down bound, for symmetry: Huang now says the semiconductor industry "probably needs to be somewhere between five to 10 times larger than it is… way too small today." Against a ~$1T industry, that is $5-10T by the late 2030s. Memory at 25-30% of an AI-weighted mix and SK at ~30% of memory would put SK's revenue an order of magnitude above today's. Stated for scale, deliberately not wired into any forecast: it is a CEO's decade-out directional claim, and the engine runs on dated, verifiable inputs.
Per-customer anchor: the Alphabet line — triangulated, error bars stated
Alphabet's Q2 gave the model its first per-customer anchor set: 2026 capex ~$195-205B, $811B of disclosed future purchase commitments, and Bloomberg's datapoint that Alphabet is a bit over 7% of SK's revenue today. The triangulation: 7% of the ₩357T 2026 line ≈ ₩25-27T (~$17-18B) per year from one customer — roughly 8-9% of Alphabet's capex landing at SK alone, implying total-memory intensity of ~15-18% of hyperscaler capex, exactly the band the chains assume. Against the $811B commitments stack, SK's slice at ~15% memory share × ~55-60% SK share runs ~$65-75B cumulative (~₩100-110T) over the multi-year horizon — a range, not a number, because 10-Qs disclose by category not counterparty, and this card says so. Two structural upgrades ride along: long-term contracts heading to 25-30% of SK revenue (a quarter of the income statement de-cyclicalized — the volatility-damping shift, quantified), and the leverage fact that outranks the arithmetic: hyperscalers offered to fund SK's fabs and were refused — urgency converted into terms. Anchor customers, contractual and diversified; single-customer risk falling even as hyperscaler mix rises.
Bottoms-up #3: the Physical-AI adder — robots pull memory twice
Jensen's Tokyo week made robots a countable lane. Onboard: each humanoid-class robot carries Jetson-Thor-style unified LPDDR (128GB reference). Cloud: every fleet needs continuous training + simulation — straight HBM. Honest scale-check built in: the onboard lane is tiny in 2028; the training pull and the 2030s curve are the story.
Smart money — what the three loudest holders actually said, and how much they can prove
Brad Gerstner Altimeter
Sold the entire Microsoft position to fund NVIDIA and SK hynix. ~80% of the firm's capital now sits in memory, logic and compute. His words: "You have to make choices in this market. We only have so much capital." He has said the rotation already paid off, naming SK hynix specifically. Highest-conviction signal of the three, because it was funded by selling a great business, not spare cash.
Gavin Baker Atreides
Pitched SK hynix at Sohn Montreal 2025 as "3× upside riding the AI inference wave with a dominant HBM supplier" — then his late-2025 13F showed him adding SK hynix, SanDisk and Micron, all of which have delivered triple-digit gains in 2026. Latest public reasoning on the All-In podcast gives four reasons the memory crunch continues, starting with consumer hardware lacking the memory to run models locally. Motto: "follow the money." Note the risk management: he holds concentrated AI bets while hedging market beta with QQQ puts.
Leopold Aschenbrenner Situational Awareness
SK hynix was already ~6.5% of the disclosed book before the ADR, and he then cornerstoned the offering alongside Baillie Gifford (a holder since 2000) and Coatue — ~$5B, about a quarter of the deal. Q1-26 filings show the same AGI-infrastructure shape: Nebius, Bloom Energy, SanDisk.
What none of them can show yet
No post-crash disclosure exists, and none is possible: the Aug 14 13Fs show June 30 books, so the cornerstone purchases and any crash-week buying stay invisible until Nov 14. Anyone claiming to know what they did during the −15.4% week is guessing. What is on the record is the position taken before it: a quarter of a $26.5B deal, and a fund that sold Microsoft to be here.
The $500B SK-NVIDIA partnership — decomposed, not swallowed
Announced Jul 24-25 at the San Francisco AI Summit and confirmed on both companies' own newsrooms: a "$500-billion-plus comprehensive partnership," signed as Letters of Intent, spanning AI factory construction and AI memory supply. Headline numbers this size deserve arithmetic, so here is the only piece that can actually be computed.
What is concrete
SK Telecom builds a 2-gigawatt Vera Rubin DSX AI factory, first facility online 2027, powered by SK hynix HBM4. NVIDIA and SK hynix enter a long-term partnership to secure and co-develop next-generation AI memory.
The 2GW memory math
2GW facility ÷ ~1.25 PUE ≈ 1.6GW IT load ÷ ~180kW per VR NVL72 rack ≈ 8,900 racks × 72 GPUs = ~640,000 Rubin packages × 288GB HBM4 = ~184 petabytes of HBM. At $4.50/Gb that is ~$6.6B of HBM, plus system DRAM and eSSD ≈ $10-11B of SK hynix memory content in this one factory.
So where is $500B?
Mostly not memory. One 2GW factory is roughly $40B of NVIDIA silicon, $10B of SK memory, and $10-20B of construction, land, cooling and power — call it $60-70B all-in. $500B therefore implies on the order of 15GW of eventual build, or a decade-long horizon, or both. Note the direction of much of it: SK buys from NVIDIA. It is "business together," not revenue to SK hynix.
Why it still matters a great deal
Three things the headline obscures. (1) Demand security: a long-term co-development agreement with the customer that is ~50% of world HBM demand converts spot exposure into contracted volume — the same structure taking LTAs to 25-30% of revenue. (2) Vertical integration: SK Group now owns memory and an AI factory, which is the missing venue for Memory-as-a-Service. (3) Sovereign anchor: Korea as an AI-native nation, with an AI Frontier Lab, a KAIST Korean LLM, and NVIDIA's earlier pledge of 260,000 GPUs to Korea over five years.
The discipline
LOI, not contract Letters of intent are non-binding. Model treatment: zero incremental revenue booked. It raises confidence in the 2027-28 demand floor and in the MaaS option; it does not move a revenue line until definitive agreements and volumes are disclosed. Tripwire: if no definitive agreement is signed within 12 months, treat the number as marketing.
Memory-as-a-Service — the option that just found its venue
Chey to Bloomberg TV on debut day: "We could be memory servicers, memory as a service." He has signalled this direction since 2022, and calls it "an idea at this point" — no pricing, no scope, no timeline.
The mechanism: today every server carries its own memory, much of it idle. CXL lets memory be disaggregated from individual servers into a shared pool any server can draw from on demand — so SK would supply memory capacity plus the software to operate it, tailored per customer, the way AWS leases compute. Reported as having synergy with SK's 20GW data centre plan.
Why the $500B deal changes this from talk to plan: MaaS needs a building. SK Telecom's 2GW AI factory is that building — SK Group can run pooled memory inside its own facility, on its own silicon, with NVIDIA compute, and meter it. That is the first plausible venue for a pilot.
Why it matters to the share price more than to revenue: memory manufacturers are valued at 8-10× because earnings are assumed cyclical. Recurring metered service revenue is valued at 20-40×. If MaaS ever gets a named customer and a price, the argument stops being about the size of earnings and starts being about the multiple — the single largest lever in the audit above. Model treatment: 0% of fair value. Free option, tracked. Prediction on the log: named pilot with pricing by end-2027, confidence 5.5/10.
Scenario D — The Vahdat Curve: AGI → ASI compute singularity
Google's infrastructure chief, on a slide to staff: "Now we must double every 6 months…. the next 1000x in 4-5 years." Run that doubling two more years and you get 10,000×. This card turns the claim into a machine with three organs: a demand rocket (AI usage compounding, accelerating after AGI via agents + recursive self-improvement), an efficiency sponge (better chips + better algorithms absorb part of every doubling — Google's Ironwood is 30× more power-efficient than its first TPU, and Vahdat's stated goal is 1000× capability at "the same cost, same power"), and a factory ceiling (HBM bits can only grow as fast as wafers, TSV and packaging allow). Whatever the sponge can't absorb and the factories can't build doesn't disappear — it becomes price. Memory pricing is violently superelastic in shortage: 2017-18 DRAM doubled on single-digit undersupply.
AGI year (after this, agents + RSI accelerate the curve)
Post-AGI acceleration (× the growth rate — the ASI kicker)
Efficiency sponge (×/year absorbed by silicon + algorithms — Google needs ~4; newest instrument: Rubin LUT-B 3.125-bit weights)
Industry HBM bit supply growth (%/yr — wafers/TSV/packaging physics)
Inference share on specialized/frozen-class ASICs by 2030 (each carries ~70% of GPU-class memory per token)
Price ceiling (× today's $/Gb before buyers balk / substitute)
Compute demand by 2030
—
vs 2026
HBM bits wanted vs buildable, 2030
—
—
SK HBM revenue 2030
—
—
Implied fair value (Scenario D)
—
group EPS × 8-10×
Read the machine honestly. Under the plain Vahdat defaults the story is not "SK grows 1000×" — it's rationing: most demanded bits go unserved, factories run flat-out, and price does the rationing work. That is Fortune's "stage two of AI, where the bottlenecks are physical" — power, cooling, build time — which is this site's thesis wearing Google's badge. Note the built-in tension: Vahdat's "same cost, same power" goal is literally a bet against memory pricing — Google wants the efficiency sponge to eat the curve. SK's upside case is the sponge failing to keep pace while Jevons converts every efficiency win into even more demand — and Dwarkesh's Jul-30 inequality (lab revenue ~10×/yr vs compute supply ~3×/yr → prices must rise; spot +40% since Feb, Google paying 2× spot) is the same structure printed at the compute-rental layer, one level above ours (the 1000× is Jevons in action — and Kimi K3 just staged the live demo: open-weight frontier model, maker's GPUs exhausted in 48 hours, ~1.1TB of HBM per serving instance; Baker's Jul 17 margin-waterfall thread is the argument in one screenshot). Physical telemetry that must appear if this curve is real: hyperscaler capex through $700B/yr by 2027 (McKinsey: $5.3T of AI datacenter capex by 2030) · tens of GW newly energized per year · ASML EUV book · Ironwood-class TPU volumes (each carries 192GB of HBM — Google's curve is an SK order) · P&T7/M15X on schedule. What kills it: efficiency >4×/yr (demand-in-bits flattens), AGI arriving late into the 2029-31 capacity wave (glut — bear row 7), power/capital ceilings, or stacked-LPDDR/HBF stealing the inference tier (bear row 11).
Scenario D → the master prediction
The singularity scenario doesn't replace the base case — it gets a probability and joins the weighted forecast. The market won't wait for 2030: if the Vahdat curve becomes consensus (visible capex acceleration + 2027 HBM contracts pricing up), the re-rate happens in the forward multiple within months. Default 12% → 15% (Jul 16) → 18% (Jul 18): ASML added a full 2027 order book + large 2028 orders. Revalued Jul 23: D-consensus target lifted ₩4.5M → ₩4.7M — the street’s non-singularity targets already print ₩3.9M (Daishin) and ₩4.3M (Hanwha, raised from ₩1.63M), so the D-case must sit above them. Blended 12-mo E[V]: ₩3.80M. Jul 24: Alphabet's raise moves four-company capex to ~$725B — the $700B telemetry bar cleared a year early by guides alone — P(D) 18%→20%, base +₩0.05M: E[V] ₩3.86M. Aug 3: September hikes verified (UBS +32% Q3 / +18% Q4) and propagated — Q3 default +20%, Q4 +10% → base ₩3.70M, blended ₩3.90M. Aug 4 recalibration: P(D) 20→22% (Chain #6 = new independent instrument) + downside branches 13→10.5% (CXMT cost data, two named row-12 mitigants) → blended ₩3.92M; fan: median ₩3.72M · P(>today) 94% · P(>2×) 81% · p5 ₩1.730M. Superseded 12 Aug by fan v9.2 (dashboard): median ₩5.12M · ≥2× at 66% post-recal · p5 ₩2.15M — the Aug-4 record stands above as the audit trail; the two runs merge at the 27 Oct re-run. External calibration (2 Sep): HSBC’s regime model assigns ~46% to adverse AI narratives (positioning capitulation 26%, China competition 20%) — in line with this engine’s 44–45% bear mass; and the street’s ₩3.0–3.2M targets are our multiple on a 2027 plateau, not a lower multiple on our path. Aug 7: capex guides revised $725→$732.5B (H1 actual $301B) + Goldman $7.6T through 2031 + SpaceX 10-GW Nvidia-exclusive (≈₩15T single-buyer HBM math) — logged; P(D) held at 22%, guide delta inside noise. TSMC (record Q2, CoWoS sold out into 2027, capex at record high-end) and ASML (second guidance raise, EUV ramping) printed the exact telemetry the prediction row demanded — early. 28 Aug decision — Patel trim reversed as a misapplication: his warning targets high-multiple names ("everything at 2-3×" is the endpoint — our entry is 2.9× forward), his exact scenario already carried 10% weight as the engine’s largest bear branch (double-count), and his self-flagged "highly speculative" rate guess has faded (10Y 4.74→4.66%) while the street re-rates the other way (Needham $220 at 6× CY28 EPS). Multiples restored; blended E[V] ₩3.92M stands unqualified; year-end path site-wide: ₩5.5 → 8.0 → 8.4 → 7.9 → 8.5M. Full engine re-run on Q3 actuals (Oct 27, estimated — date not yet confirmed on SK hynix IR).
RE-RUN REGISTER · inputs queued for the Oct 27 engine run · as of 2 Sep 2026
Pushes the chain and path UP (not yet in the numbers)
Demand envelope $1T → ~$1.2T (NVDA Q2, FY28 guided +70%) · top-5 capex $1.3T 2027 · SpaceX ~10GW outside it · AWS +2M GPUs · Korea sovereign 8.4→18.4GW on Rubin/HBM4. Scarcity window stamped by principals: buyer through 2028 · CEO through end-2030 · Micron 50% gap · Goldman shortfall widening 5.0% → 5.9% next year, HBM market +108%. Realised pricing: HBM export unit value $76.13 (Jul, +9.5% MoM; +47% Q1→Q2) · DRAM +24.3% · volumes down, values up · spot HBM ~5× contract (upside register if formulas track spot). Contracts: typically five years (Q2 call, primary), deposits, no ceiling (TrendForce) · NVIDIA multi-year DRAM+HBM reported. Attach/custom legs: SOCAMM2 shipping · NVHBM → Trainium, MediaTek channel opened · Jalapeño, HBC sockets. Not modelled at all: physical-AI / robot-fleet training infrastructure (Musk G20). Technical state (5 Sep): price above MA20 (₩1.61M) and MA200 (₩1.31M), below MA50 (₩1.84M); 20/50 bearish cross unresolved; regime gate ₩1.744M; support ₩1.558M; realised vol 74%. Street in EPS terms: consensus FY27 EPS ~₩326K vs this engine’s ₩583K — same multiple, half the earnings. Physical layer (5 Sep): TSMC tool requirement 1.0× → 1.9× vs its Dec plan, capex to $60–64B; SK wafers +~18% 2026–27 with HBM at ~22% of wafers for ~9% of bits — value grows, supply stays disciplined. OEM confirmation (4 Sep): Dell $60.9B AI-server orders in a quarter, $95B backlog, constraint "DRAM, DRAM, DRAM, then NAND"; Broadcom AI revenue +221%, FY27 >$100B of custom XPUs — every one an HBM socket.
Pushes DOWN or re-weights branches
L2 share — trigger touched 3 Sep: Counterpoint Q2 SK 50% of HBM revenue (58% Q1, 64% Q2-25), Samsung 33% (21%), Micron 18%; gap 37 → 17 points; Samsung share-loss branch re-weight 8.5% → ~10% queued here; Rubin HBM4 provisional split SK mid-50% / Samsung mid-20% / Micron ~20% (Goldman: two-thirds) — Samsung branch 8.5%, re-weight upward if SK prints <50% of HBM revenue in any quarter. In a shortage, share bites in 2029, not 2027. China: CXMT 8% of DRAM (Q1), HBM3E pilot reported, $8.6B IPO — 7% branch unchanged; HBM3E substitution now a monitored tripwire. Pricing nuance: LTA base levels sit below spot (UBS; Patel estimate) — affects the upside register, not the base chain. Q3 HBM ASP is the arbiter. Rates: Fed September hike odds ~59% (Warsh); KOSPI −4% on 2 Sep on rates and oil; re-trim trigger is a 10Y close >4.90% — not hit (≈4.66%). Content per GPU (FY28 input): the site now shows Rubin Ultra at 192–384GB/GPU depending on stack height (8-high HBM4 at 192GB per SemiAnalysis 1 Sep; 12-high 288GB; 16-high 384GB) and NVIDIA is reportedly testing cut-down Rubin Ultra configs — the FY28 content assumption is re-derived at the run, with the offset question (72 → 576 GPUs per rack) alongside. Supply-driven cuts raise price, not our volume. Substitution (2028+): d-Matrix 3D-DRAM-on-logic in working silicon (32GB/card, supplier undisclosed) — tripwires set: commercial qualification, a named hyperscaler, a disclosed DRAM source. Full register on the Demand page. FX: KRW/USD ~1,346 vs a ~1,400 planning basis — a ~4% headwind to every won-denominated figure at constant chip prices (Nomura trimmed Q3 OP ~10% on it); a wash on dollar intrinsic value for foreign holders. W6 Solidigm: pre-IPO raise undecided per 6-K (4 Sep), next by 4 Dec. Tariff rule: "build in America or pay" has no rate or date; if only wafer fabs are exempt, Indiana packaging may not count — unpriced; 24 Sep Trump–Xi. The whole chain sits above the street’s TAM path: FY27 ₩816T ≈ $600B needs ~45–47% of Bernstein’s $1.3T consensus memory TAM (SK’s blended share is ~40% — so price and share must both deliver), and FY28 ₩1,102T ≈ $820B needs the TAM near $2.0T. FY27 is the first test (Oct 27, then the January print). Confidence flag (5 Sep): the +119% HBM-price leg inside the +79% is now the most scrutinised assumption — Micron reportedly doubling HBM wafers to ~100K/month and Samsung at 33% mean a larger share of the scarcity rent may accrue to the other two; the L2 branch re-weight (8.5% → ~10%) already sits in this register.
Armed reviews: 2029 pause (₩7.9M at 8.8×) — trigger #1 LTA tenors fired and primary-confirmed; #2 supply-model crossover pushed right (Goldman’s widening shortfall is a partial); #3 second supplier guiding through 2030 — open. Two of three would move the 2029 multiple at the re-run. Everything above is logged, dated, and held for actuals — the chain, fan and path re-derive together on Oct 27.
P(Scenario D becomes market consensus by mid-2027)
12-month price if D goes consensus (fwd multiple re-rate + 2027 EPS visibility)
Blended 12-mo E[V]
—
—
Base-case component
₩3.65M
recentered Jul 18; +₩0.05M Jul 24 (Alphabet capex propagated through the chains)