SK Hynix — The Thesis

The print: revenue $96.2B, +106% YoY — YoY growth accelerating for a fourth consecutive quarter at $400B scale; Q3 guided $108B (first $100B quarter) with zero China DC compute assumed. The guide that breaks cadence: FY28 at +70% (~$691B vs $570B consensus) — analysts read it as ~$200B on top of the $1T Blackwell+Rubin envelope: call it ~$1.2T through CY2027. The capex wall, from the best-sighted desk on earth: cloud backlog >$2T; top-5 capex ~$800B in 2026 → $1.3T in 2027 — 18% above consensus, and consistent with the ~$1.2T-memory-market conservation audit already on this page. The pricing-power receipt: GM printed 75.0% (+2.6pts) while absorbing "memory cost headwinds", Q3 guided 74.0% — a ~100bp bite, under the 200bp tripwire — as server prices reportedly rise ~15%. Our ASPs are a named line in the buyer P&L, passed through, absorbed. And the status line that starts our revenue clock: Vera Rubin production shipments commenced early August — HBM4 recognition is live. The $/GW ladder, now a modelling primitive: Hopper $18B → GB $25B → Rubin $40B per gigawatt; at est. 25–35% memory content, FY28 at ~17GW is ~$170–240B of memory through NVIDIA systems alone. Full teardown with charts → Demand page. Our tape around it: ₩1.730M (21 Aug) → risk-off week to ₩1.500M close (26 Aug) → ₩1.688M close +12.5% (27 Aug) on the print — then the ₩1.744M tripwire was tested on 28 Aug and not held — intraday high ₩1,788K, ex-div close ₩1,653K. Monday 31 Aug opened −4.4% on Warsh’s Jackson Hole hawkishness (September hike odds 35%→59%, SOX −3.5%) plus CXMT’s H1 revenue print, and closed +1.3% at ₩1,674K — foreigners and institutions net sellers, "other corporations" the buyer: the ₩40T buyback acting as breakwater. Tuesday 1 Sep ₩1,693K; Wednesday 2 Sep ~−4.7% as the KOSPI fell 4% on rates and oil, overwhelming buyback support; Thursday 3 Sep ₩1,596K (−1.05%) against a +1.6% index — company-specific weakness ahead of the Q2 share print, which landed at 6pm after the close; 4 Sep, the first session on it, closed ₩1,647K, +3.2% (high ₩1,683K, +5.45% intraday) with foreigners the #1 net buyer of the day at ₩527B and institutions +₩159B — the print was absorbed, not sold. The 3 Sep intraday low of ₩1,558K is now the tested support. Regime flip not confirmed; ₩1.744M remains the gate, ₩1.58M (31 Aug low) the tested support above the ₩1.500M capitulation. The ADR tells the same story: SKHY has carved a classic IPO base — a tightening wedge from the $194.80 July high through the ~$130 late-July low, volume drying into the apex near $162; a break above ~$172 is the base breakout and rhymes with the ₩1.744M gate. The drawdown was positioning; the reversal was the customer numbers — foreign investors returned as net buyers of Korean chip stocks Thursday (Seoul Economic Daily) with the KOSPI pushing toward a record. The buyer put our scarcity claim on the record: NVIDIA warns memory shortages could persist through 2028. Street confirmation, lagging as always: Needham PT to $220 (6× CY28 EPS; ~₩500T FCF 2025–27, ~₩250T returned modelled), JPMorgan ≥$130B shareholder returns through 2027.

The structural item: NVIDIA productised custom high-bandwidth memory as NVHBM and is extending it via NVLink Fusion to Trainium — Annapurna gets faster, more power-efficient memory "in partnership with memory suppliers." The hyperscaler-ASIC lane — the layer the three-supplier / Micron-in world was supposed to commoditise — just standardised on an NVIDIA memory spec that we manufacture. Feynman "Die-Stacking Custom HBM" (2028 roadmap) pulled forward to now, and licensed outward beyond NVIDIA GPUs. This is the counterweight to T3: co-designed sockets decay slower than qualified-commodity sockets. The volume line: AWS commits 2M additional GPUs in 2027–28 (Blackwell Ultra / Rubin / Rubin Ultra) on top of 1M+ from GTC — order-of ~16M incremental HBM stacks from one customer (est., mix-dependent) — plus 100K GPUs for US federal IL6+ AI factories, Vera CPUs on AWS (LPDDR/SOCAMM attach), and Amazon Robotics on Jetson/Isaac. Huang: demand is "running ahead of every forecast." One press release, five demand lanes. Bear row 11 note: the Q2 deck names LPX a "new growth driver" — SRAM decode is permanent; sizing held −2–5%.

The claim, from the best build-tracking desk in the industry: SpaceX aims 6–8GW incremental in 2027 (potential 10GW+) at ~$50B/GW — $300–500B of capex from one new buyer, on par with AWS or Google, and not inside Kress's top-5 $1.3T. SemiAnalysis believes it (~2GW by end-2026 → ~10GW by end-2027; Southaven 27→69 turbines in five months; MiniHard ~500MW in ~5 months). The line that matters for us: Elon declared NVIDIA-exclusive — TPU/AMD evaluations abandoned, likely against NVIDIA vendor financing. NVIDIA-exclusive = HBM4-at-~70%-ours exclusive. At our $/GW ladder (memory ≈ 25–35% of $40B/GW), a 2027 SpaceX build is ~$100–140B of memory demand from a customer that did not exist in the model. Alongside it: Microsoft's awakening — 10GW of signed binding contracts YTD (~$300B) after last year's pause. And the economics that keep it all funded: inference revenue >$100B/GW/year vs ~$12B/GW/year rental cost — GPU payback under a year — which is why 2027 memory sold out in August. New entrants arriving after the sell-out bid up what remains: the ASP risk is to the upside.

What was reversed: the 20 Aug −6% multiple haircut (9.5→9.0 · 10.6→10.0 · 10.2→9.6 · 8.8→8.3 · 9.0→8.5), taken on the rate-regime argument with the 10Y at its 4.74% nineteen-year high. Why — three legs, in order of weight: (1) Category error, caught by Nancy. Patel’s warning, verbatim on the card below: memory does great, memory stocks shouldn’t 10× again, because in that world "everything should trade at 2 or 3× multiples." That is a market-wide de-rating aimed at what trades at 30-58× — and his crash-endpoint of 2-3× is our entry multiple (2.9× forward). You cannot compress a stock to a level it already trades at; the compression he describes falls on the peers at 46-58×, and our 9.5-10.6× exit multiples already sit at roughly one-fifth of theirs. The card even said so at the time: "our end-2027 is 4.6× today’s price, not ten — not the thing he is warning against" — and then trimmed anyway. (2) Double-counting. His exact paradox was already in the engine as the multiple-compression scenario — 10% weight, outcomes ₩2.6-3.9M, the largest single bear contribution. Haircutting the central multiples on top counted the same risk twice. (3) The premise was self-flagged and then faded. He called his ~250bp rate estimate "highly speculative"; the 10Y has since eased 4.74% → ~4.66%, while the buyer itself warns shortages persist through 2028. Restored: multiples 9.5 / 10.6 / 10.2 / 8.8 / 9.0 on next-year clean EPS → year-end path ₩5.5M → 8.0 → 8.4 → 7.9 → 8.5. Earnings chain untouched in both directions. The honest verdict on process: the 20 Aug trim double-counted a self-flagged, market-wide mechanism onto the one stock already trading at its endpoint — the bear branch was the right place for it, and it was already there. Both entries stand in the audit trail; the lesson (risks live in branches once, not in branches and the central case) is logged.

Street (lagging confirmation, weighted as such): Zacks Rank moved #3 Hold → #2 Buy; its current-year consensus estimate rose +20.3% in one month with no negative revisions — the estimate-revision trend is the one street metric with empirical signal. ADR mean PT $245 (+48%), range $200–320; KRX consensus ₩3.16–3.22M (+87–92%, 37–39 analysts, zero sells); Hanwha ₩3.15M with the line "₩40T is the start, not a one-off". Insiders (DART): no executive open-market purchases in the window; the only filing is 82 treasury shares to independent directors (immaterial). The company itself is the insider: ₩40T / 24.07M shares (3.3%) bought on-market via SK Securities, 20 Aug → 19 Nov window, NH estimating ~₩645B a day, IBK noting the daily cap is 42% of average volume; shareholder return raised from "within 50%" to "over 50% of cumulative 2025–27 FCF" with fixed and special dividends alongside — and the board minuted the reason: intrinsic value "not sufficiently reflected in the current price". On 31 Aug that programme visibly absorbed a −4.4% open. Also this week: SK is considering Intel Foundry for HBM4E base dies (Herald Economy — not finalised; base-die diversification beyond TSMC, with the EMIB packaging tie-in shown at Hot Chips); Temasek disclosed investment in SK and Samsung mid-August; Counterpoint Q2 NAND share SK 22% vs Samsung 25%, SK gaining on Samsung’s supply constraints, Solidigm growing strongly — the NAND flank answer to the YMTC card below.

What was said: by video link to the G20, Musk predicted "well over a billion" humanoids within ten years, each ~5× a human’s productivity, calling it conservative and worth "serious money"; AI adding 20–30% to global GDP. How this site treats it: a long-duration upside scenario, not a demand input — Musk’s unit forecasts have a track record (Optimus years behind schedule; China ran 85% of humanoid installations last year, 140+ companies), and the sober anchors are Morgan Stanley’s ~8M US units by 2040 and Macquarie’s $3T market by 2050. The part that matters for us arrives first and is highest-margin: robots are trained, simulated and updated in data centres — multimodal video, world models, reinforcement learning — the most memory-bandwidth-bound compute there is. The HBM and server-DRAM infrastructure to create robot fleets lands years before the LPDDR/NAND inside each robot, and it lands on our product. Scale check for the eventual second leg: at an illustrative 32–128GB DRAM and 0.5–2TB flash per unit, a billion units is tens of exabytes of DRAM and 0.5–2 zettabytes of NAND across a deployment cycle — extraordinary if it happens, and none of it is in the chain. The layered read: 2026–28 the HBM cycle (unchanged); 2028–31 physical-AI training/inference capex as a possible incremental catalyst; beyond that, unit volumes broadening the commodity TAM. Roadmap notes folded alongside: HBM4E samples 2H 2026 → full production 2027; 16-Hi HBM4 qualification under way, hybrid bonding targeted at 20-Hi+ (not ready for HBM4E — the 775µm ceiling); SK sees HBM demand exceeding its own capacity for three years; a co-packaged-optics memory-fabric roadmap published in Nature Electronics — data movement, not compute, as the bottleneck this company sits on.

1 · The third supplier moves (AMBER, L2): ETNews reports Micron adding up to ~60K HBM wafer starts a month, taking it to ~100K/month by end-2026 from 40–50K last year — and Micron is already in volume production of 36GB 12-high HBM4 for Vera Rubin. Read: more scarcity rent shared three ways rather than two; the pressure lands on the +119% HBM-price leg inside our +79% weighted-price assumption for 2027 and on share, not on aggregate demand. The offset is real: wafers moved into HBM leave conventional DRAM, which tightens the commodity price we also sell. Hard tripwire added: Micron itself confirming ~100K/month with a large HBM4 mix → lower the 2027 SK share/ASP sensitivity even with HBM sold out. 2 · Google confirms the memory wall while showing the best compression yet (net slightly GREEN): Google Cloud’s supply-chain lead says high-performance memory has risen to more than 75% of an AI server’s hardware bill of materials, and that multimodal, MoE and agent workloads are shifting AI from compute-bound to memory-bound. Its TPU 8i — after aggressive optimisation and ~80% better performance per dollar — still carries 288GB of HBM plus 384MB of SRAM. The counterweight: TurboQuant, a 3–4-bit KV-cache representation cutting KV memory ~5–6×; SK hynix’s own guest article says compression is approaching practical limits, after which architectures need more HBM, HBF, CXL and near-memory. A genuine less-memory-per-task instance — the decode/KV substitution scenario gets a 5–6× compression stress test at the Oct 27 run — and the strongest Jevons datapoint yet: efficiency and memory capacity rising together, on the same chip. 3 · Independent voice on scarcity (GREEN): TechInsights’ Dan Kim (FT): DRAM/NAND pricing up more than 200% YoY, the market exceptionally tight, meaningful greenfield capacity not before ~2028 — the first non-supplier, non-buyer principal in the bracket. Micron can ramp HBM inside existing fabs; that redistributes scarce wafers, it does not end the shortage. Model action: keep ₩816T / ₩583K central; reduce confidence — not the number — in the +119% HBM-price component; the L2 branch re-weight already queued.

Tape: the ADR closed $177, +8.1% on heavy volume after Korea had finished at ₩1,647K (+3.2%) — a memory rotation (Micron +5% intraday, the chip index only ~+3%) that SK led. Korea trades alone Monday (US Labor Day), so a higher Seoul open is catch-up, not new news. The ADR now sits ~45% above the Korea-implied price ($122 at ₩1,346/$) — the gap widened because the US ran without Korea. Solidigm (W6): SK filed a 6-K saying there is no decision on the rumoured ~₩5T ($3.7B) pre-IPO raise; next update when decided or by 4 December. No extra cash this quarter; the rumour is parked, not killed. Nomura: kept ₩4.7M Buy but cut its Q3 operating-profit estimate ~10% because the won strengthened (Asiae reports ₩8.6T → ₩7.7T; SK’s quarterly run-rate is ~₩80T, so read the cut as ~10%, units aside). The FX point deserves its own line: SK is paid in dollars and reports in won; at ₩1,346/$ versus a ~₩1,400 planning basis, every won-denominated figure in this model carries a ~4% headwind at constant chip prices. For a foreign holder the same move raises the home-currency value of a won-priced share — roughly a wash on dollar intrinsic value, a real haircut on reported KRW EPS. Logged in the re-run register. Voices: Gavin Baker — hyperscalers are building behind-the-meter power and selling surplus back at spikes (the "grid is full, demand is fake" worry weakens); Beth Kindig — AMD + Cisco + Saudi HUMAIN plan up to 250MW from 2027, and NVIDIA’s own compute-rental business could outgrow the big clouds within quarters; Tom’s Hardware — discrete GPU sales at a four-year high despite soaring memory prices; The Information — high-end Macs for local AI are short, with buyers switching to NVIDIA (Macs eat DRAM, NVIDIA eats HBM — both ours).

Technicals (000660, 4 Sep close ₩1,647K): MA20 ₩1.61M (price above, turning up) · MA50 ₩1.84M (price 11% below, falling) · MA200 ₩1.31M (price 26% above; no close below it through the entire run) · RSI 50 · 20-day realised vol 74%. The −49% correction from ₩2.92M (22 Jun) to ₩1.50M (26 Aug) is a base above the 200-day, not a broken trend; the 20/50 bearish cross of 24 Jul is unresolved and the reversal completes on a close above the 50-day — converging with the ₩1.744M regime gate. Break of ₩1.558M → MA200 ₩1.31M becomes the test. Options (SKHY, 18 Sep monthly): 116K calls / 92K puts; 45K calls at $200 — a squeeze trigger above ~$190 and a lid below; max pain $157.5 vs spot $176; weeklies put-heavy with 23.6K deep $105 puts into 25 Sep — tail hedges in place, i.e. holders are protected and less likely to be forced sellers. ADR premium ~45% over Seoul. Street sweep: S&P Global raised SK hynix to A− from BBB+ with a positive outlook, citing AI-driven performance and the capital-return plan; new ADR coverage — Cantor Overweight $300, Stifel Buy $240, RBC and Wolfe Outperform $200–240, Needham $220, William Blair modelling FCF more than doubling by 2028; ADR consensus $245–248 (14 analysts, +52%); KRX consensus ₩3.21M from 38 analysts — 37 buy, 0 sell, 1 hold, range ₩1.54–5.3M; Simply Wall St raised fair value to ₩3.41M from ₩3.13M on tighter supply. The plateau, in EPS terms: consensus next-year EPS ~₩326K against our ₩583K — same multiple, half the earnings; that gap is the entire bet. Insiders: nothing new on DART; the buyback bought Friday. Footnote for the disruption register: TetraMem and SK hynix completed a joint analog in-memory computing project (journal cover) — SK is inside the compute-in-memory research too.

TSMC (the machines, not the narrative): the equipment TSMC says it needs to serve 2026 demand went from 1.0× its December plan to 1.5× in Q1 and 1.9× by July — a ~90% upward revision in required tools, not in output; 2026 capex lifted from $52–56B to $60–64B, 70–80% of it leading-edge, 10–20% packaging and test; ~20 fabs and packaging plants under construction, and fab construction labour is now the reported constraint. Tool demand spikes long before revenue-bearing capacity arrives — this is the capacity tracker’s thesis (physical orders lead revenue by 6–18 months) printing at the foundry. SK hynix (the supplier): physical wafer capacity rises only ~90,000 wafers, ~+18% across 2026–27 (Mirae estimate); 2026 capex in the high-₩40T range because M15X was pulled forward and Yongin Phase 1’s cleanroom opens early 2027; HBM4 mass-production shipments began in Q2 with the fuller ramp in H2, and HBM4E samples are with a major customer for 2027 volume. The number that reconciles "only 18%" with a 76% margin: HBM consumes ~22% of SK’s DRAM wafer capacity in 2026 for ~9% of DRAM bits, rising in 2027 — so AI-memory value can double while wafers grow a fifth. Management expands in phases against confirmed customer demand rather than flooding the market: constructive for price and margin, and the reason memory stays the binding constraint on NVIDIA’s +70% (≈$670–690B FY28) longer than investors expect. The chain: cloud capex → NVIDIA systems → TSMC logic + packaging → HBM + substrates + networking + power; the slowest link sets shipped revenue, and the slowest link is ours. Neo-cloud GPU capacity is projected above 8GW by year-end from 3GW; Rubin is already ~a fifth of NVIDIA’s data-centre revenue. Korean policy tape (4–5 Sep): Washington is tightening semiconductor tariffs again and Vice-President Vance warned foreign firms that US investment is rewarded and its absence penalised — Korea’s trade minister says Korea "will not be treated worse than competitors" (the Indiana rule question stays open); the ₩800T Samsung/SK Honam investment plan has budget-line gaps to resolve, with the two companies proposing a ₩25T electricity prepayment to unblock power — power is the bottleneck behind the bottleneck; and the government’s ₩200T R&D programme explicitly backs HBM expansion. Union bonus disputes ruled not strike-eligible.

Dell, verified: a record $60.9B of AI-server orders in one quarter, $16.4B shipped, a $95B backlog, FY27 guidance raised $25B to $192B, first Vera Rubin racks shipped, 6,500+ AI customers. Asked what stops it shipping more, COO Jeff Clarke: the binding constraint remains "DRAM, DRAM, DRAM, followed by NAND, NAND, NAND", with patchy CPUs and disk drives behind. Enterprises want more cores, more DRAM and more storage; configurations "cost more than they did last quarter, and the quarter before" — and large customers are now ordering further in advance and co-planning needs years out, which Clarke called a new phenomenon. Why it matters more than an analyst note: this is the OEM layer — the customer of our customer — saying orders run far ahead of shipments because supply is the limit, not appetite. It is difficult to reconcile with any near-term "cheap tokens collapse hardware demand" thesis. Broadcom (per watch note): Q3 AI-semiconductor revenue $16.7B (+221% YoY, +54% QoQ), Q4 guided to $21.7B, FY27 AI revenue above $100B across Google, Anthropic, OpenAI and Meta custom XPUs — every one of which is an HBM socket. Custom accelerators are expanding memory demand, not routing around it: the Jevons argument, with the invoices attached.

What was shown: a TSMC 4nm compute die bonded directly onto custom DRAM; early silicon reportedly ~100–105 TB/s of memory bandwidth at ~0.37 pJ/bit on the vertical interface, against ~2.4 pJ/bit cited for an HBM4 interface; ISCA work projects ~4.7× throughput per card versus an HBM design. The limits that keep it off the tripwire list for now: 32GB per card against 192–288GB on HBM4-class accelerators; end-to-end claims are modelled, not independently benchmarked; the DRAM supplier is undisclosed; volume and yields unknown. How it fits the disruption register: this is near-memory / 3D-stacked DRAM — a form-factor shift within DRAM, not a reduction in DRAM demand. If it works at scale, the bits are still made by three companies; what moves is where the packaging and bandwidth premium is captured. Risk one notch up for 2028+; neutral for 2027.

The print: SK hynix kept first place at 50% of Q2 HBM revenue — down from 58% in Q1 and 64% a year ago — while Samsung went 21% → 33% (the only supplier to gain) and Micron slipped to 18%. The gap narrowed from 37 points to 17 in a single quarter. Most of Samsung’s jump is HBM3E catch-up; Counterpoint says HBM4 is still a small slice but expects Samsung’s share to expand further as HBM4 shipments become meaningful in H2. In total DRAM, Samsung led at 38% with SK at 25% (down 14 points YoY) — the commodity price surge paid Samsung’s mix more than ours. How it lands on the model: the armed trigger was "below 50% in any quarter"; it printed at 50% with the trend against us — a half-step. The Samsung share-loss scenario re-weight (8.5% → ~10%) is queued for the Oct 27 run (branch weights re-derive together, never alone), and the verdict steps 8.4 → 8.3 now, because a verdict is a judgment, not an engine output. What it does and doesn’t change: volume in a shortage is unchanged — SK sells every stack it makes — but the monitor’s point stands: buyers now have a real second source, and a second source can nibble pricing power even while the shortage lasts. That pricing channel is the near-term L2 mechanism; the volume channel stays a 2029 question. The print landed at 6pm Seoul on 3 Sep, after a session in which SK had already lagged (−1.1% vs KOSPI +1.6%); 4 Sep, the first trade on it, closed +3.2% with foreigners the top net buyer — absorbed. What decides whether 50% is a floor or a slide: the H2 mix of HBM4 (12-high, Rubin) — not Samsung’s Q2 catch-up on HBM3E. Watch Rubin allocation at the Q3 print.

The hard numbers: Korea’s average HBM export price per unit reached $76.13 at end-July, +9.5% MoM and the first time above $70, after rising ~47% from $40.94 in Q1 to $60.22 in Q2 as HBM4 shipments began; July HBM export value fell 20.4% while volume fell 27.3% — price and mix are doing the work. Conventional DRAM export price +24.3% to $22.90 in July (+36.6% since May); TRASS DRAM unit value $92,183/kg, +401% YoY in 1–20 August; Micron’s FQ3 DRAM ASP +low-60% QoQ. Goldman now sees the DRAM shortfall widening from 5.0% this year to 5.9% next, persisting through 2028, with the HBM market +108% next year; TrendForce has HBM at 30% of big-3 wafer input by end-2027 for 13% of bits. What it settles from this week’s debate: pass-through is real — DRAM ASP +30% in Q2 under the new LTAs and export prices still climbing in July. What it leaves open, honestly: the same wire reports spot HBM at ~5× contract levels — which says LTA base prices sit well below spot. Read both ways: if the no-ceiling formulas track spot with a lag, that gap is the upside register; if they don’t, it is UBS’s point in one number. The Q3 HBM ASP print is the arbiter — not plugged into the model (generation and mix shifts are embedded in the export series), but it raises confidence in the upper half of the 2027 range. Custom-XPU channel widens: NVIDIA and MediaTek announced NVLink Fusion + NVHBM as the foundation for hyperscalers and labs building their own accelerators — customers bring an XPU design, customise memory, packaging and connectivity, and MediaTek/NVIDIA supply the production ecosystem. That opens the second-adopter channel the NVHBM card armed; NVIDIA still has not named its NVHBM suppliers, so it is a trigger opened, not fired.

What moved: over the past month SK hynix fell 8.98% against Samsung +1.18% and KOSPI +0.49%; foreign investors sold ₩5.98T of SK hynix vs ₩0.79T of Samsung; LS Securities cut SK to ₩2.4M and raised Samsung to ₩450K, with the analyst’s own framing: "it does not mean growth in the HBM market is slowing — it is the competitive structure among suppliers normalising." Samsung unveiled next-gen HBM4 Monday with improving yields. The share record behind it: Counterpoint has SK at 50% of HBM revenue in Q2 2026 (58% in Q1, 64% a year earlier); NVIDIA’s provisional HBM4 allocation per Hankyung is SK mid-50%, Samsung mid-20%, Micron ~20% — against Goldman’s "about two-thirds" — and Korean analysts now model HBM4 at 50/30/20. Correction to our own front page: the at-a-glance tile said "~70% of NVIDIA’s Rubin HBM4"; the defensible range is 50–67%, and it now says so. What it means for the model: this is the Samsung share-loss scenario (8.5% weight) being actively repriced, not a new fact — and in a shortage, share is a 2029 problem, not a 2027 problem: SK sells every stack it can make regardless of share; share loss only bites when supply loosens and allocation becomes price competition. L1 (demand) is untouched by any of this; L2 is the second weak link after L4 (multiple), and the market is pricing it first.

Reported: CXMT has begun small-scale HBM3E production, with Alibaba’s T-Head and Cambricon testing it and possible commercial use in 2027; yields and specs undisclosed, sources still place CXMT 3–5 years behind and struggling with yield. Context: Counterpoint has CXMT at 8% of global DRAM in Q1 2026 (from 3% a year earlier), H1 revenue ~₩31T, and the July IPO raised ~$8.6B to fund expansion. Model treatment: Rubin/HBM4 share untouched — the risk is price pressure on older HBM generations and Chinese domestic sockets from 2027–28, and commodity DRAM thereafter. The China scenario stays at 7% weight; what changes is that Chinese HBM3E substitution becomes a monitored tripwire rather than a hypothetical. Flows diagram updated: CXMT "~0% HBM" → "HBM3E pilot (reported)".

The claim: SemiAnalysis estimates NVIDIA "has likely secured" favorable HBM pricing and large 2027 volume commitments through the SK Group relationship — its own words: "the final negotiation result could vary from our current estimate." UBS separately reads revised LTAs as possibly capping near-term pricing upside. The record: on the Q2 call SK hynix said the LTAs are "designed in various forms," typically five-year terms, with pricing structures negotiated "to respond to price volatility rather than a single method," plus purchase commitments and deposits for enforceability. TrendForce, citing industry sources, reports SK hynix removed the industry-standard price cap — spot increases flow fully into contract prices in a shortage — making it possibly the only major supplier without a cap, while Micron’s SCA caps existing products at Q2-FY26 market prices with a ~62%-margin floor. And the prints under those contracts: Q2 DRAM ASP +30% QoQ, NAND +mid-50%, gross margin 83%. Verdict: floors without ceilings is the on-record structure; "favorable pricing" is a self-flagged estimate, and UBS’s "cap" reads as negotiated base levels below spot peaks, not a contractual ceiling. Primary source outranks preview until the ASPs say otherwise. Magnitude re-weighted 6 → 3. The sovereign build, read correctly: Korea’s 8.4→18.4GW programme (SK Group 5GW + 10GW; 2GW of Rubin committed) is, for the supplier, home-market HBM4 demand, deeper NVIDIA co-development (custom HBM, SOCAMM lock-in), and privileged compute for memory R&D and simulation at a moment when compute itself is the scarcest input — an advantage competitors cannot buy. The one legitimate residual from Patel’s "cost" framing is capital ring-fencing: does SK hynix’s cash fund group AI-factory vehicles? The evidence today says no — ₩88T cash, 7% debt-to-equity, ₩40T buyback, returns raised to >50% of FCF, memory capex guided in the high-₩40T range, no disclosed equity or guarantees into DSX. Bonus, primary-sourced: the Q2 call’s "typically five years" is primary confirmation of review trigger #1 (LTA tenors into 2029–31).

BofA’s trajectory corroborates our FY26 almost exactly: Q3 operating profit ₩81T, Q4 above ₩90T, DRAM ASP +25% QoQ in Q3 and +9% in Q4, an annualised run-rate above ₩300T by year-end — that quarterly path sums to ~₩355–360T of 2026 revenue against our ₩360T. Where the street parts from us is 2027: BofA holds the ₩300T+ run-rate flat through 2028 on 80%+ DRAM margins; the PTs (BofA ₩3.0M / ADR $250, UBS $204, consensus ₩3.16–3.22M, Simply Wall St ₩3.18M) are that plateau at roughly our multiple; the one cut this week — LS Securities to ₩2.4M — is explicitly a share-structure call, not a demand call. So the bet, in one line: same ~10× — the street prices a 2027 plateau, we price growth with the customer (+70% guided). Support for our side: UBS models DRAM bit demand accelerating to +36% in 2027 (from +22%) and a 2027–31 ROE of 40% against the 17.7% the price implies; BofA sees Big Tech capex at $1T+ in both 2027 and 2028, SK >40% share of HBM and QLC eSSD through 2028, Samsung 30–40% of HBM, China under 10% of the 2028 TAM. Honest cross-check on our own chain (corrected 5 Sep — an earlier note here used the pre-Q2 ₩633T vintage): FY27 ₩816T ≈ $600B needs ~45–47% of Bernstein’s $1.3T consensus TAM against a ~40% blended share — price and share must deliver; FY28 ₩1,102T needs the TAM near $2.0T. The whole chain sits above the street’s path, which is exactly what "growth vs plateau" means. FY27 is tested on 27 Oct and at the January print. Calibration check: HSBC’s regime model puts "AI positioning capitulation" at 26% and "China competition" at 20% — ~46% adverse, against this site’s 44–45% bear mass. Exec quote graded (Park Joon-deok, Q2 call): acknowledges the rental-vs-build and efficient-model worries directly, then reframes them as utilisation and monetisation of installed infrastructure — a specific, testable claim ("additional supply requests continue to mount") rather than a deflection.

What was said: a major long-horizon allocator publicly frames SK hynix in the same breath as TSMC and ASML — the two most consensus-quality monopolies in semiconductors — as positions US export controls and capex scale have put beyond China’s reach. And the China assessment comes with a number: EUV restrictions leave CXMT one to two generations behind in advanced memory. How to weight it, per the standing rule: institutional narrative is lagging confirmation, never signal — but this is the specific kind of lagging confirmation that moves flows: the anomaly-gap thesis (bottleneck supplier priced like a commodity maker) closing in public, told by an allocator whose framing shifts other allocators. The CXMT line independently corroborates the EUV-wall pillar and the 7% China scenario weight — a named institutional voice now carries what this site derived from tool counts. The $1T club, with the arithmetic done honestly: SK crossed $1 trillion in market cap during 2026 (at the ₩2.99M highs); at the ₩1,647K close the cap is roughly $0.89T, regaining $1T above ~₩2.0M. Membership was earned this year; re-entry is ~18% away. TSMC context: $160B+ combined capex and US investment — the moat trio is spending like a moat trio. Counter-datapoint, logged the same day: CXMT reported H1 2026 revenue ¥150.3B (~₩31T) — the commodity price cycle is funding China’s DRAM champion fast even while it stays EUV-limited; the market flinched at Monday’s open and reversed by the close. Revenue scale is not process parity; the row stays AMBER at 7%.

The claim and the funding: at its IPO roadshow YMTC set a target of surpassing Samsung and SK hynix in NAND flash by end-2027; the parent filed for a $4.9B Shanghai IPO last week, with Q1 revenue up ~5× YoY to ~$7B (note: everyone’s NAND revenue is up in this price cycle — the 5× is partly the tide). The honest sizing, before the headline scares anyone: YMTC is #3 in NAND shipments at 14% but only #5 by revenue — a consumer-heavy, lower-margin mix with no meaningful enterprise eSSD presence, and it remains on the US entity list, equipment-constrained at the leading edge. "Surpass by end-2027" is plausible in bits with IPO-funded capacity; in revenue it requires an enterprise qualification story YMTC does not yet have. An IPO roadshow target is marketing until wafers say otherwise. What it actually touches in this model: (1) the NAND flank of the 2029-31 fork — YMTC’s IPO is the same state-capital mechanism as CXMT’s, now funding both memory types’ China capacity waves; commodity NAND pricing 2027+ is the first pressure point, not HBM/DRAM where the profit engine lives; (2) W6 timing — a YMTC-led NAND price war would compress Solidigm’s valuation window, mildly arguing for earlier monetisation; (3) the Kioxia-ties headline (single-source, on watch) reads naturally as the consolidation counter on exactly this flank. Magnitude 2-3/10: real, dated, NAND-specific — logged where the fork ledger can see it.

Two reports, one direction: Edgewater Research says NVIDIA has signed multi-year DRAM and HBM supply agreements with SK hynix and Micron (Samsung notably unnamed in that item), and memory contracting industry-wide is shifting from one-year deals to three-to-five-year commitments. Micron’s now-CEO-level framing on CNBC: AI demand keeps supply constrained beyond 2027, new fab capacity not on stream until 2028, customers demanding 50% more than deliverable. Why tenor is the tell: a 3-5-year commitment signed in 2026 runs into 2029-2031 — buyers are contracting straight through the very window the bear case reserves for oversupply. This is Kwak’s commodity→custom argument showing up in the paper itself. Trigger note: this is corroboration test #1 of the armed 2029-pause review (LTA tenors into 2029-30) — fired via secondary sourcing, 1 of 3. Primary confirmation point: the Oct 27 earnings call’s LTA disclosure. Logged, not acted — the review holds its date.

Hard numbers for the crowding-out math: HBM yields one-third the gigabytes per wafer of standard DDR (matching this site’s 3× wafer-consumption factor; Patel argues 4×) and now consumes 23% of total DRAM wafer output — the eviction of commodity supply, measured. And the ceiling is physical: Hot Chips 2026 material shows die-thinning, thermal density and TSV area pushing HBM stacking toward compounding yield and capacity limits with no quick fix — scarcity is not a planning choice, it is lithography and heat. Hard to make = priced like it. The fourth custom socket: Qualcomm unveiled High Bandwidth Compute (HBC) — a near-memory architecture claiming ~6× efficiency over current HBM, with Samsung and SK hynix confirmed as memory collaborators; first silicon back in the lab, commercial shipping targeted 2027, a server CPU in late 2028 with Meta as first customer, plus the Modular acquisition and a Hugging Face partnership aimed at NVIDIA’s software moat. The read, consistent with Jalapeño: every efficiency challenger builds with the memory makers, not around them — Rubin, Jalapeño, TPU, Trainium, now HBC: the accelerator war keeps multiplying memory buyers. Sized as a demand lane with a share question, not a threat: watch item — what memory product HBC actually consumes (custom stack? PIM-class?), and whether "6× efficiency" means less bandwidth per socket or more sockets per watt (Jalapeño resolved to the latter).

The quote, in full: "No one can know exactly when the shortage will end. However, there are currently no clear signs of a downturn. We expect the supply shortage to continue until the end of 2030." Three principals now bracket the scarcity window independently: the biggest buyer (NVIDIA: shortages could persist through 2028), the biggest HBM supplier (Kwak: through end-2030), and the third supplier (Micron’s Sadana: data-centre demand outstripping supply by 50%, meaningful new capacity unlikely before 2028 — capex doubling to ~$26B FY26, >$45B FY27). And the structural argument is ours, in his words: past cycles were commodity boom-bust — companies raced production. In the AI era memory becomes customised product co-developed with customers, demand more predictable, downturns less severe. That is the LTA-floors / no-price-caps / prepayment / custom-HBM pillar of this site, stated by the person who signs the contracts. What it does to the model — measured, not reactive: 2026-28 is already fully priced in the chain; the live question this statement opens is the 2029 pause (₩7.9M at 8.8×), which assumes the capacity wave bites on schedule. If the shortage genuinely runs through 2030, that multiple step-down is one to two years early. Not changed today — a CEO talks his book, and the path has moved twice in 24 hours — instead the 2029 review is formally armed for the Oct 27 re-run with corroboration tests: contract tenors extending into 2029-30, TrendForce supply models pushing the crossover right, Samsung/Micron capex phasing. Also logged, both ways: the workforce constraint — Kwak calls skilled labour "a critical issue" in Indiana (McKinsey: US short up to 150K semiconductor workers by 2030) — an execution variable for the US ramp and one more reason industry capacity cannot arrive fast. And a watch item: SK is "open to investing anywhere" with water, power, talent and subsidies — front-end US fab pressure exists; nothing finalised beyond Indiana.

What happened: shovels in the ground on the $4B, 133.5-acre West Lafayette complex — a next-generation HBM packaging line + AI-semiconductor R&D centre, ~7,000 direct and indirect jobs, cleanroom open Oct 2028, mass production — the first "Made in USA HBM", HBM4E — from Q3 2029 (Purdue partnership), framed by the CEO as the key US memory production base by 2030. Washington is co-funding our moat: up to $458M direct CHIPS money + a $500M federal loan — $958M, ~24% of stated cost. Why it matters: advanced packaging (TSV/MR-MUF) is the single tightest bottleneck in HBM — this localises it beside NVIDIA and the US customer base, hedges tariff and geopolitical risk on Korean-packaged parts, and hardens the ADR / Nasdaq-100 story with physical US presence. The honest timing read, both directions: it does nothing for today's crunch (supply discipline preserved through the sold-out window) — and it lands new back-end capacity squarely in the 2029–31 fork window. Packaging converts wafers, it doesn't mint bits, so it is not wafer oversupply — but it goes on the fork-year capacity ledger regardless. Also on the 28 Aug tape, logged not weighted: CEO says SK is weighing closer Kioxia ties (single source, Bloomberg headline — unconfirmed, watch); union members rejected the tentative wage deal (~50%+ against — minor operational friction in a record-profit year); Samsung announced $80B in dividends/buybacks after AI gains (peer capital-return pressure — supportive comp for our ₩40T programme).

The correction is right and it matters. On 1 October 2025 OpenAI signed letters of intent with both Samsung and SK hynix — confirmed on OpenAI's own site — "targeting 900,000 DRAM wafer starts per month at an accelerated capacity rollout." The scope covers DDR5, LPDDR, HBM and specialty DRAM, supplied as undiced wafers. SK's framing went further than supply, describing joint work on next-generation memory-computing architectures. So yesterday's read was too negative: OpenAI deliberately dual-sourced, and SK holds its own OpenAI relationship. Jalapeño's first HBM4 going to Samsung is a socket, not a defection. Corrected.

But the headline number has to be sized, and it does not survive it. 900,000 wafers a month is about 47% of all global DRAM capacity — KED describes it as "more than twice the current global production capacity." At 170,000 wafers per gigawatt it implies 64 GW of compute against a Stargate programme announced at 10 GW — 6.4× the stated project. No timeline was ever attached. It is an ultimate ambition, not a plan. The more usable figure is the analyst estimate of ₩100 trillion over four years across both suppliers — roughly ₩12.5T a year each, about 3.5% of SK's FY26 revenue. Real, and modest.

And here is the part the correction did not carry, which cuts the other way. The letters were non-binding — no obligation to buy a single wafer. And it is reported, citing Bloomberg, that roughly five months later the Stargate expansion that justified them was scrapped, with Oracle and OpenAI ending plans to expand the flagship site. If that holds, the 900,000 figure is a stale, non-binding ambition attached to a programme subsequently cut back. It is not a demand input and this page will not use it as one.

The Broadcom programme, sized the same way. OpenAI and Broadcom describe 10 GW of OpenAI-designed accelerators through 2029. At 170,000 wafers per gigawatt that is 1.7M wafers in total — spread across four years, about 425,000 a year, or 1.9% of annual world capacity. If SK wins half, roughly 0.9%. Meaningful; not a step change.

Net, and the tripwire is the right one. The strategic relationship is real and verified. The Jalapeño read is correctly softened. The headline number is unusable. The model input is unchanged, deliberately — because a letter of intent is not an order, and this one is attached to a programme that has already been trimmed once. What would move the model is a confirmed supplier allocation on production Jalapeño and its successor: Samsung-only would be evidence for the 8.5% share-loss scenario; a dual-source or SK win would be materially positive. That is a fact. The LOI is a hope with a signature on it.

Custom HBM — real, sold, and near. SK's customised HBM is backed by orders from Nvidia, Microsoft and Broadcom, with requests received from across the Magnificent Seven and design-ecosystem ties to Broadcom and Marvell. The vehicle is HBM4E, whose 12-layer 48GB samples shipped on 18 June 2026 at 16 Gbps, with customisation moving from experimental to mainstream during 2027. The market is forecast to grow from $18.2B in 2024 to $130B by 2033, and SK is described as set to lead it. On standard HBM4, SK is reported to hold two-thirds of Nvidia's orders — consistent with UBS's ~70% estimate — and supplies HBM3E into Microsoft's Maia 200.

PIM — no named customer, and 2028. LPDDR6-PIM is scheduled for 2028 and described as being for "specialised applications". That phrasing is doing work: it is a capability, not a product line with a buyer. High Bandwidth Flash is further out still — not before 2030, requiring all-new media and a specification agreed with SanDisk and the other NAND makers. Both belong in the thesis as free optionality and neither belongs in a revenue line. Our model carries them at zero, which is correct.

Why the distinction matters more than it sounds. A roadmap slide showing MRAM, PIM, HBF and 3D DRAM reads as a moat until you ask who has signed for it. Custom HBM passes that test — three named buyers and samples in customers' hands. PIM does not, yet. Treating them as one thing would have let a technology demonstration do the work of a revenue forecast.

And the contest is genuine, which is where our share scenario actually gets decided. Samsung has added 250 engineers to custom projects targeting Google, Meta and Nvidia, moved its custom logic die to a 2nm process — a first — and already supplies over 60% of Google's TPU HBM3E, with reports it beat expectations in Broadcom's HBM4 testing. Micron is reported as trailing because it kept its existing DRAM process to control cost, described as "a structural disadvantage in the custom HBM race." So the 8.5% share-loss scenario is not an abstraction — it is Google's TPU sockets and Broadcom's ASIC sockets, being contested right now. That is the thing to watch, rather than the architecture papers.

What OpenAI built. Jalapeño went from team hiring to tape-out in ~16 months, and SemiAnalysis — who benchmarked it in OpenAI's lab — call it "industry leading, beating every Nvidia, AMD and Google chip we have been able to test." It reaches 15.4 TB/s of memory bandwidth per package, besting everything shipping on HBM3E, and implies HBM4 running at 10 Gbps pin speeds against the 9.6 Gbps Nvidia gets from Rubin. It is "one of the relatively early adopters after Nvidia and AMD, even beating the established TPU and Trainium programs."

So the first-order read is bullish, and clearly so. Every serious new accelerator architecture now runs HBM4: Rubin at 288GB, AMD's MI455X at 432GB per GPU — half as much again as Rubin, with Helios racks carrying 31TB against Rubin NVL72's 20.7TB — and now OpenAI's own silicon. Competition between accelerator vendors pushes memory content up, not down. The custom-ASIC threat everyone worried about has arrived, and it turns out to be another HBM4 customer with the highest bandwidth per package in the industry.

Now the line that matters to us, and it is a share risk rather than a demand risk. "The HBM is likely provided by Samsung." With OpenAI and Anthropic expected to absorb roughly half of incremental compute by late 2027, HBM inside OpenAI's own silicon going to a competitor is volume SK does not get.

Sized before reacting. Jalapeño's stated next goal is 100 MW, with production ramping through 2027 and most output scheduled for Q4. At 170,000 DRAM wafers per gigawatt, 100MW is about 17,000 wafers — 0.075% of annual world capacity. Even scaled to 2 GW, far beyond anything on the schedule, it is under 2%. And "likely" is SemiAnalysis inferring, not OpenAI or Samsung disclosing. Strategically important; volumetrically small inside our window.

Filed correctly, which matters. This does not belong in the decode-substitution scenario — that covers architectures which reduce HBM, and Jalapeño increases it. It belongs in "HBM share lost to Samsung and Micron", already carrying 8.5% and already the second-largest bear case. This is evidence for a scenario we hold, not a new one, and one unconfirmed supplier attribution does not justify re-weighting it. The trigger that would: OpenAI or Samsung confirming the award, or Jalapeño volumes moving past 1 GW.

And one thing now confirmed from three directions. OpenAI: "limited by datacenter power, not by budget or floorspace — tokens per MW is paramount." Jensen at Computex: "If you have 1 gigawatt of power, then throughput per watt is revenue." Nvidia at Hot Chips: "The data center is power limited today." Power, not money, is the binding constraint on AI — which is exactly why memory that computes beats memory that only stores, and why SK's PIM, custom-base-die and computational-storage roadmap is worth more than it looks. A footnote in the same vein: OpenAI declined to split prefill from decode because "a fungible fleet shifts capacity while a fixed split strands hardware" — the identical argument SemiAnalysis made against Cerebras, now made by a buyer designing its own silicon.

First the physics, because it settles the literal version. Low Earth orbit is ~550km, so a round trip at light speed is 3.67 milliseconds before any processing. HBM4 access latency is around 100 nanoseconds. That makes an orbital hop roughly 37,000× slower than local memory — up to 245,000× against a best-case 15ns. Bandwidth is the second wall: one HBM4 stack moves about 2,000 GB/s while an optical inter-satellite link manages roughly 12.5 GB/s — 160× less, per stack, and a Rubin GPU carries eight. Transmitting a single rack's 74.7TB over one such link would take 1.7 hours. You cannot serve GPU working memory from another orbit, let alone from Earth. HBM has to sit beside the processor. That part is not an engineering problem awaiting a solution; it is arithmetic.

But the architecture that forces is the interesting answer. What can move between tiers is colder state, and the hierarchy becomes: Earth archive → laser link → orbital SSD → orbital DRAM and CXL pool → local HBM → GPU. Every one of those four layers is a product SK sells — HBM4/HBM4E, SOCAMM and server DRAM, CXL pooled memory, and Solidigm enterprise SSD. Instead of selling one component into each accelerator, SK could participate in every tier of an orbital memory hierarchy.

And the second-order argument is the strongest part of it. An orbital rack is power-limited before it is compute-limited — every watt demands solar panel and radiator mass, and mass is launch cost. Moving data is among computing's largest energy costs, and SK is building precisely the technologies that reduce it: custom HBM with the controller moved onto the base die (cutting interface power), LPDDR6-PIM in 2028, CXL memory with compute, and computational storage. In a power-constrained system, memory that computes is worth more than memory that only stores. SK's value per GPU could be higher in space than on Earth.

A correction to something this page said two days ago. We described DRAM as "among the most radiation-susceptible components there is." Reporting on Starlink's operating experience suggests processors and memory are reasonably radiation-tolerant in low Earth orbit, and that power electronics may be the more vulnerable part. The real requirement is fault tolerance — error correction, redundancy, memory scrubbing, bad-block isolation — rather than exotic radiation-hardened silicon. That is a lower barrier than we implied, and it also happens to mean more error-correction capacity per useful bit.

Model weight: zero, deliberately. First launch is Q4 2027, one satellite. Distributed memory pooling across a constellation is speculative architecture — the note conceding that CXL would not run over those distances is the honest part of it. Logged as named optionality that costs nothing to hold. The triggers that would change that: SpaceX naming HBM4/HBM4E suppliers, disclosing memory capacity per orbital rack, or describing PIM, CXL or near-memory architectures in the orbital design.

The physical anchor. A gigawatt of Vera Rubin compute consumes ~170,000 DRAM wafers against ~55,000 leading-edge logic wafers — 3.1× more memory silicon than logic silicon. This is not a demand forecast that can be revised down; it is a property of the machine. Worked against industry capacity of roughly 1.9M DRAM wafers a month, it says something stark: if every DRAM wafer on Earth went to AI, the world could build 134 GW a year and no more. At the $2 trillion of capex implied for 2028 — roughly 40 GW of new build — about 30% of global DRAM capacity is consumed by AI data centres. Consumer memory rising 90% in a quarter is not a squeeze. It is an eviction.

Two figures that appear to contradict, reconciled. Reporting that "AI consumes ~70% of global memory production" sits oddly beside a ~30% wafer calculation. Both are right: 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 precisely where SK's margin lives, and it is why the mix shift matters more than the volume.

The unit economics, which are the first hard answer to "does the buildout pay for itself". Anthropic is reported at up to $50M of revenue per megawatt against $10-15M per megawatt of compute cost — a return of 3.3× to 5× on compute. That is the engine allowing the labs to keep bidding aggressively for infrastructure, and it is the strongest available rebuttal to the bubble framing. Against it, SemiAnalysis estimates ~$11 trillion of capex across 2024-29, of which more than $5 trillion may need credit financing — the same figure sitting underneath the rate warning.

And an honesty note on our own adjustment. We trimmed multiples roughly 6% on Patel's argument that heavy borrowing lifts rates and compresses valuations. He explicitly calls his own ~250bp rate estimate "highly speculative." The mechanism is sound; the magnitude is his own guess, and he says so. A 6% trim on a directionally-right but self-flagged-uncertain argument is proportionate — we are not reversing it, but the record should show he hedged it, because he did. The more durable conclusion is the reframe: value this on realised earnings, cash flow and buybacks rather than on multiple expansion, which is already how the path here is built.

China, dated more precisely. Patel expects the Chinese capacity inflection around 2028 — millions of domestic accelerators and 5-10 GW of domestic AI compute — while still lagging Western systems materially on performance. That is at the edge of our forecast window rather than inside it, and it argues for monitoring CXMT yields, node generation and wafer starts rather than haircutting 2027 earnings today. Our China scenario stays at 7%.

The disagreement, stated straight. Susquehanna's channel checks put DRAM contract pricing above +50% quarter-on-quarter this quarter with NAND near +60%. TrendForce's survey, reiterated on 13 August, puts conventional DRAM at +13-18% and NAND at +10-15% — explicitly "a marked slowdown from the roughly 60% jumps recorded in the second quarter." Korean desks cited by Hankyung sit at about +20% for SK specifically. That is a three-way spread of 13% to 50% on the same quarter, and our model uses ~20%, between TrendForce and the Korean desks. We are not adopting the highest number because it is the most flattering. If Susquehanna is right, Q3 revenue lands far above our ~₩104T path requirement and the ₩98T tripwire looks absurdly conservative — but that is a reason to wait for the print, not to move the forecast toward the loudest estimate.

Now the part that runs against us, and it should have been caught earlier. Realised second-quarter ASPs: Samsung's DRAM rose mid-40% and NAND high-60% quarter-on-quarter. SK hynix's DRAM rose about 30% and NAND mid-50%. SK realised less price than Samsung. The explanation is consistent with our own thesis rather than against it — SK's higher HBM mix was contracted earlier, and mobile mix diluted the blend — but it is a caution against assuming the removed price cap shows up immediately. The uncapped structure applies to new and renewed agreements, so the benefit lands in 2027, not in a quarter already sold. Our FY27 assumption depends on that timing being right.

The volume datapoint is genuinely useful. KeyBanc's John Vinh now models roughly 1.9 million Nvidia Rubin units for the year and says explicitly that he raised the estimate because SK hynix HBM4 supply is ramping, with Rubin shipments rising about 250,000 units quarter-on-quarter next quarter. That is an analyst tying accelerator volume directly to SK's output — it supports the HBM4 socket-share and yield assumptions and argues against SK packaging becoming the constraint on Rubin. Separately, Hankyung reports SK's 2026 DRAM bit growth reaching the mid-20% range, above the previous 20% estimate, driven by 1c-based SOCAMM2. Both sit comfortably inside our inputs.

And a live Jevons reading, small but vivid. Gavin Baker reports his own firm's AI spending in August at roughly 100× its March level and still doubling monthly. One organisation is not an industry, and it is anecdotal. But it is direct evidence for the mechanism the whole agentic thesis rests on: efficiency gains get spent on more usage rather than banked as savings.

One watch item, unverified and flagged rather than modelled. A market-data aggregator reports Nvidia's Rubin Ultra HBM stack under review — potentially moving from 16-high HBM4E down to 12-high or 8-high. If it lands at 12-high, that is exactly the 48GB configuration we already model, so our number is the downgraded case. If it lands at 8-high, our FY28 HBM content assumption is too high. Single-sourced, so it is a watch item, not a revision — and it is the specific thing to listen for on tomorrow's Nvidia call. Also confirmed there: 2027 HBM supply is fully booked across all three vendors, and HBM4E targets 70% yield ahead of an H1-2027 launch.

1 · The correction, because it is ours. This page carried "$6B of fab capital produces one gigawatt a year, and one gigawatt produces $100B a year of revenue." That compressed three separate layers into one sentence and it is wrong. The transcript describes: $6B of industry-wide fab tooling produces the silicon for a gigawatt; that gigawatt then requires roughly $50B of datacentre capital to deploy; and the deployed gigawatt earns revenue at the service layer. The $6B does not produce the $100B. And the $100B itself looks high — the same interview gives unit economics of "$10-15M per megawatt costs, earns up to $50M", which annualises to about $50B per gigawatt. Corrected to a $50-100B range. It is also not memory capital — it is whole-industry wafer fab equipment. Both halves of the phrasing were wrong; both are fixed.

2 · Musk's exclusivity closes a door, and the inference is sharper than we had made it. Musk: "We think it's the best AI computer… we're exclusive to Nvidia." Nvidia's Rubin and Rubin Ultra run on HBM4 and HBM4E from the three qualified suppliers. Terafab targets MRAM, RRAM and 3D DRAM — not HBM — with operations "as early as 2028". Intel's XBM lands around 2030 and ZAM in 2029-30. So Musk's entire compute demand, terrestrial and orbital, flows through Nvidia and therefore through HBM from the big three for the whole Rubin generation. He cannot self-supply inside this window even owning a fab. What is left from that quarter is share risk, not substitution risk — a genuinely narrower thing.

3 · Is SK working on MRAM and RRAM? Yes — and stronger than "working on". TrendForce: "Samsung and SK hynix both hold leading patents and prototypes in next-generation memory technologies such as MRAM and ReRAM." The published roadmap runs to 2031: 2026-28 brings HBM4 16-Hi, HBM4E 8/12/16-Hi and a custom HBM4E that moves the memory controller onto the base die — co-developed with TSMC, freeing silicon area for compute and cutting interface power — plus LPDDR6, SOCAMM2, MRDIMM Gen2, CXL LPDDR6-PIM and 245TB-plus QLC enterprise SSDs. LPDDR6-PIM arrives 2028. Then 2029-31: HBM5, HBM5E, custom HBM5, GDDR7-next, DDR6, 400-plus-layer NAND and 3D DRAM. Terafab is aiming at precisely the technologies SK already holds leading patents in — and custom HBM with per-customer controller silicon deepens the moat rather than commoditising it. One more: UBS expects SK at roughly 70% of HBM4 for Rubin, above its 57% overall HBM share.

4 · And the market-size check, which had to be done. Annualising the printed second quarters: SK ~$229B, Micron ~$200B, Samsung memory ~$250B, plus Kioxia, SanDisk, CXMT and Nanya — roughly $800B of run-rate. Our model assumes a ~$890B memory market in 2026, so it reconciles. Worth stating plainly: WSTS forecast total semiconductors at ~$975B for 2026, which cannot contain an $800B memory market — that forecast predates the price explosion and is stale. We are modelling against observed run-rates rather than a superseded forecast, and that is the correct choice, but it should be visible rather than assumed.

What was announced. SpaceX and Nvidia have co-designed a space-optimised Vera Rubin NVL72 for launch in Q4 2027 — pulled forward from 2028 — with "significant scale" following in 2028. The first-generation Starmind AI1 satellite carries a 120kW compute payload peaking at 150kW on a craft with a 75-metre wingspan, in sun-synchronous orbit for continuous solar power and radiative cooling. Musk on the supplier question: "We think it's the best AI computer… we're exclusive to Nvidia." Alongside it, $100B committed to a Starbase site in Louisiana.

The line worth more than the space story. Musk: "We expect to actually deploy this on the ground as well as in orbit, because we think it's going to be a radical simplification of the normal NVL72 rack. It will cost less. It'll be more effective. If we're going to put it in space, why not put it on the ground?" The same design, minus solar arrays and radiators, goes into SpaceX's terrestrial data centres. So this is a new rack architecture with a terrestrial market, not only an orbital experiment — and every one of those racks carries the 74.7TB of DRAM the UBS teardown measured.

The number to file: JPMorgan's Doug Anmuth writes that "from 2029 on, we expect SpaceX to pursue orbital compute towards ~75GW by the end of 2031." At roughly 130kW per NVL72-equivalent, that implies hundreds of thousands of racks. It is also entirely outside our forecast window, which runs to 2029, and Musk's timelines are famously optimistic — the first satellite is one satellite. Logged as telemetry, not as a model input.

Three reads that do matter now. First, the honest one: radiation qualification is unsolved. Nvidia and SpaceX must still qualify Rubin GPUs and Vera CPUs for a multi-year low-orbit mission, and DRAM does suffer single-event upsets, though reporting on Starlink's operating experience suggests processors and memory are reasonably radiation-tolerant in low Earth orbit and that power electronics may be the more vulnerable part — a correction to what this page said two days ago. Solving it means more error correction and more redundant capacity per useful bit, but it is engineering risk before it is demand. Second, and more useful: SpaceXAI has acknowledged that orbital compute at its target scale "requires significantly more chips than it currently has access to." That is another large buyer stating plainly that it cannot get enough. Third, and the real signal — this exists because the ground is running out of room. Firms are willing to solve radiation, thermal rejection and launch economics rather than fight for land, power and water permits. Demand intensity, revealed by what people will endure to get compute.

One speculative connection, flagged as speculation. Terafab's own job postings name MRAM and RRAM — memory types that happen to be inherently radiation-tolerant — and Musk is now building orbital compute that needs exactly that. Whether those two facts are related is unknown and unstated by anyone. Worth watching; not worth modelling.

The argument, verbatim, because it deserves to be quoted rather than paraphrased: "It's an argument for why — I think memory is going to do great — memory stocks shouldn't 10× or whatever again. Because if we're in the market where there's that much demand for memory — which means AI's caused this drastic change in the economy — then everything should trade at 2 or 3× multiples and the stock market should crash."

He is making two separate claims and it matters not to blur them. The business does great. The stock may not, because the same conditions that create the memory boom — roughly $5 trillion of borrowing by 2029 — push interest rates up two to three points for everyone, and higher rates compress what investors will pay per dollar of profit across the whole market. Record earnings, compressed multiple.

Does this model survive it? Mostly, and here is the honest accounting. Our end-2027 figure is 4.6× today's price, not ten — so it is not the thing he is warning against. Our multiples were already 9.5-10.6× against peers at 46-58×. And the literal version of his claim is priced: at a 3× multiple, FY28 earnings of ₩753K give ₩2.26M. Our multiple-compression scenario — 10% weight, outcomes ₩2.6-3.9M — is precisely his paradox, and it is already the largest single bear contribution in the engine. The most credible analyst in the field has independently arrived at our biggest bear case. That is confirmation, not new information.

But one thing genuinely should change, and it has. His mechanism is not a tail event — it is his central expectation for the rate regime. If that is the world, the central multiples here should reflect it, not only the bear branch. Trimmed roughly 6% across the path: 9.5→9.0 · 10.6→10.0 · 10.2→9.6 · 8.8→8.3 · 9.0→8.5. End-2027 moves from ₩8.0M to ₩7.5M, end-2028 from ₩8.4M to ₩7.9M. Earnings untouched — this is a valuation adjustment, made because the argument is right, not because the news was bad.

And the other half of the interview is larger than the warning. "CapEx next year is like $2 trillion" — our 2026 demand input was $765B. Roughly 30% of Big Tech capex flows directly to memory. HBM consumes 4× the wafer area per gigabyte of commodity DRAM — this page has been using 3×, so the crowding-out of consumer supply is worse than modelled. $6B of industry-wide fab tooling produces the silicon for one gigawatt a year — which then needs roughly $50B of datacentre capital to deploy, and that deployed gigawatt earns somewhere in the range of $50-100B a year. Three separate layers, not one. Memory prices have already quadrupled and he expects them to double again. Every one of those supports or exceeds the demand and pricing inputs used here. Net: the earnings side stands or improves; the multiple comes down. Which is exactly what a man who says memory does great and memory stocks still shouldn't 10× is telling you to do.

What was published. On 20 August SK hynix, with researchers from the University of Virginia, MIT, Illinois Urbana-Champaign, Nanyang Technological University and Yonsei, published its co-packaged optics roadmap in Nature Electronics — the first public disclosure of its AI interconnect architecture at that level. The problem it names is precise: compute performance grows roughly 3× every two years while interconnect bandwidth grows only 1.4×. That gap is the "bandwidth wall".

The proposed answer matters more than the paper. An optics-centric architecture using a photonic interposer to connect processor resource pools directly to memory resource pools, so that multiple AI accelerators can share one large-capacity memory pool, breaking today's packaging limits. Targets: over 100 Tb/s per node, under 1 pJ/bit, chip-to-chip latency below 10 nanoseconds. Read that carefully — it is a design for using more memory, more efficiently shared, not less of it. Pooled memory across racks is an expansion of the addressable market, and it sits alongside the same architecture Marvell is building with its photonic fabric.

The strategic shift is stated openly by Seunghoon Hong, who leads SK's AI Infrastructure team: "The role of memory companies is evolving beyond simply supplying products into being a partner that helps build the competitiveness of a customer's entire system." A company that co-authors the interconnect standard is much harder to substitute than one that ships parts against a specification someone else wrote. Alongside it at FMS 2026: 16-high HBM4 and wafer-bonded 375-layer NAND.

And it explains a day this file had not captured. On publication day the shares rose 12.73% to ₩1,691,000, driven by the roadmap together with the ₩40T (~$28.6B) buyback. Friday's close of ₩1,730,000 sits 2.3% above that. So the sequence into our anchor is now complete rather than inferred.

Where this leaves the optical question. Three separate things are being conflated in the market and they should be kept apart. Optical interconnect to memory is happening now and expands the memory system — Marvell's photonic fabric targets pooled KV cache across racks, and its Celestial acquisition was described as freeing package edge space for more HBM. Photonic memory — light replacing DRAM cells — does not commercially exist and is a post-2030 watch item. And in between, SK is not defending against the optical transition; it is co-designing it.

What they built. AgentX 1.0 is the first fully open-source multi-turn agentic coding inference benchmark at one million tokens of context, replaying 393 anonymised Claude Code traces across MI355X, GB300 NVL72, GB200 NVL72, B300, B200, MI325, MI300X, H200 and RTX Pro servers. Their framing of why it was needed matters: "In the past most measured performance based on fixed sequence length prefill and decode workloads, but this is an inaccurate way to measure workloads. Reality is multi-turn, long context, high prefill reuse, with sub-agent bursts, KVCache offload, and numerous tool calls." And the timing marker: since the Claude Code inflection point in November 2025, agentic workloads "now dominate traffic for production inferencing", with OpenAI's enterprise agentic spending overtaking ChatGPT spending in April 2026. This is not a forecast about agents. It is a measurement of what is already running.

The sentence that matters most to this position: "Long context conversations stress the HBM capacity for KV cache and necessitate offloading KV tensors to different tiers of memory (DRAM, SSD)." Every tier in that sentence is a product SK sells.

And then they quantify it. A B300 node running 384 concurrent agentic traces with 3TB of DRAM attached achieved a 91% HBM cache hit rate, with an HBM KV working set of roughly 43 million tokens. Drop to a B200 at concurrency 196 and the HBM hit rate falls to 73%, with the system leaning on DRAM for nearly 20% of hits and the HBM working set halving to 22 million tokens. Then the design rule, stated plainly: DRAM offloading is a write-through cache — every prefix written to HBM is also written to DRAM — so it is "most effective when the amount of DRAM available for offloading is significantly bigger (a multiple of 1.5-3) than HBM KV cache capacity."

That rule is a structural multiplier on conventional DRAM, and it is new to this model. The agent economy does not choose between HBM and ordinary memory — it requires HBM and one-and-a-half to three times as much DRAM sitting beside it, plus SSD as a third tier. It explains, from the workload side, why Nvidia tripled the Vera CPU to 1.5TB of LPDDR and entered the LPDDR market at smartphone scale: not a cost dodge, but the physics of serving long-context agents. Our forecast carries the conventional DRAM lane at 44% of revenue growing 55% on price; this is the demand mechanism underneath it, now measured rather than assumed.

One more finding worth keeping. On the H200: "it cannot compete with the newer SKUs in high throughput scenarios due to lack of HBM." Not compute, not interconnect — HBM capacity is what separates GPU generations under agentic load. And the honest counterweight from the same document: Qwen3.5's GatedDeltaNet architecture has "theoretically constant state storage requirement instead of vanilla attention's linear storage requirements", meaning lower memory needs per model. Architectures that reduce KV growth are real and are being adopted — logged against the decode-substitution scenario, which stays at 7%.

The divergence, in returns. Through 17 August: SanDisk +653%, Seagate +262%, Micron +255%, Western Digital +211% year to date. SK Hynix's year-to-date return is slightly negative. Cramer named four memory and storage names as "indispensable" — Micron, SanDisk, Seagate, Western Digital — and the company holding 57% of the HBM market was not among them. His own words: "While I acknowledge that I am not early, I do not think I am late."

The trailing multiples are the anomaly in one line. Seagate 61.7× · SanDisk 41.5× · Western Digital 32.6× · SK hynix 19.6× · Micron 18.6×. Seagate makes spinning hard-disk platters and trades at more than three times SK's multiple. Western Digital makes storage and trades at 1.7×. 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.

Now the honesty note, because it matters more than the comparison. Simply Wall St publishes SK's P/E at 7.8×; 24/7 Wall St publishes 19.6×. Both are trailing, and the gap is almost certainly the Kioxia-related investment gain of roughly ₩63T that inflated last quarter's reported net profit. On reported earnings the multiple looks like 7.8×; on clean operating earnings it is closer to 19.6×. The 7.8× is the flattering number and should not be used. The forward figure carried here — 3.0× on 2027 earnings — is built on clean earnings throughout and is unaffected, but anyone quoting 7.8× as evidence is quoting a one-off gain.

Why the gap exists, mechanically. Reporting on the July pullback attributes it to profit-taking, concerns about efficient Chinese models, and "the collapse of AI-focused hedge fund Situational Awareness" — independent confirmation of the liquidation traced here. That fund's forced selling hit Korean names hardest, and 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. That is the divergence, and it is a flow explanation rather than a fundamental one — which is precisely why it should close.

Two supporting facts. The Roundhill Memory ETF holds SK at 24.2%, second only to Samsung at 25% — so index money already treats SK as a core memory holding even where commentary does not. And Bank of America projects Micron's fiscal-2030 earnings at $200-250 per share, nearly double Street consensus, on the argument that memory is structurally stronger. If that thesis is right for Micron at 18.6× trailing, it is right for the company with the higher margin and three times the HBM share.

The scale of it. Samsung posted consolidated revenue of ₩171.5T and operating profit of ₩89.5T — up 56% on the quarter and 1,814% on the year. The semiconductor division alone delivered ₩127.5T of revenue and ₩89.2T of operating profit. Everything else Samsung makes — phones, televisions, appliances, displays, networks — contributed roughly zero.

1 · The margin comparison, and it settles a claim made here two weeks ago. Samsung's chip division ran a 70.0% operating margin. SK hynix ran 76.3% in the same quarter. SK is 6.4 points higher on a third of the revenue base. Dylan Patel's assertion that SK is the industry's lowest-cost producer is no longer an analyst's view — both companies have now published the numbers that prove it, and that margin gap is what makes losing share cost SK less than it would cost anyone else.

2 · The detail that captures the whole thesis in one company. Samsung's Galaxy division posted an operating loss of ₩0.7T — its first ever — with the company attributing it directly to "elevated component cost pressures." That is the Apple decode happening inside a single set of accounts: Samsung's memory business is so profitable it 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. No clearer evidence of where the pricing power now sits.

3 · The valuation argument. Annualised, Samsung's chip division earns roughly ₩357T of operating profit against SK's ₩242T. Samsung's market capitalisation is roughly ₩2,600T against SK's ₩1,289T. On price to annualised operating profit that is Samsung 7.3× versus SK 5.3× — SK trades at a 27% discount to its own domestic rival, while carrying a higher margin and 57% of the HBM market against Samsung's 21%.

4 · And the re-rating SK has not had. Samsung's shares ran roughly 150% over the year, so a record quarter was already in the price — 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. That gap is the position.

5 · The item that matters for Tuesday. Samsung's wage agreement ties 10.5% of the semiconductor division's annual operating profit to employee bonuses — stripping it out would have pushed the quarter above ₩100T. SK's worker vote closes Tuesday 09:00 Seoul. On similar terms the cost to SK would be roughly ₩25T annualised, about ₩26K per share after tax. Worth noting that the 70% margin assumption used in this forecast is modelled on Samsung's post-bonus 70.0% — so a comparable settlement is already covered. 6 · For the supply row: Samsung has committed roughly ₩400T to a new semiconductor hub in south-western Korea, outside the established chipmaking corridor — utilities and infrastructure must be built from nothing, which argues for the later end of the 2029-31 wave rather than the earlier.

What SK presented. Lee Jae-sik — VP of SK hynix America and head of package engineering, previously at Nvidia, Qualcomm, Google and Meta — spoke on advanced HBM packaging, centred on the proprietary MR-MUF process: liquid underfill injected between stacked dies and cured, which is both more process-efficient and better at dissipating heat than the film-per-layer method Samsung uses. The product behind it matters more than the talk: SK has shipped 12-layer HBM4E samples ahead of schedule, using 1c DRAM for the first time (HBM4 used 1b) on a 32Gb core die. The specification step is substantial — per-stack capacity rises 36GB → 48GB (+33%), per-pin speed goes to 16 Gbps from roughly 11-13, and energy efficiency improves over 20%, all on the same 2,048 pins. Base die understood to be TSMC 3nm-class. Add the cooling layer disclosed this weekend — heat reduced by more than 30% — and hybrid bonding under exploration for the generation after.

That per-stack jump answers a question raised here last week. HBM capacity per GPU was flat from GB300 to Rubin at 288GB, which looked like content discipline. HBM4E ends it: 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.

What Samsung presented, and it is a serious long-term answer. Sangwook Han laid out a three-phase roadmap for the HBM base die: reclaim area, expand function, then zHBM — true 3D vertical integration of the accelerator and the memory stack, removing the 2.5D interposer entirely. Distributed I/Os shorten the internal data path and eliminate conventional 2D interfaces like the HBM PHY. Samsung's worked example: four zHBM stacks beside a 1,200W GPU deliver significantly more bandwidth while saving around 100W. It requires wafer-on-wafer and hybrid copper bonding plus a unified SoC-and-DRAM design flow. This is the most credible HBM successor anyone has shown — and note what it is: still DRAM, stacked differently. It is phase three of a roadmap, not a product. Samsung is playing architecture; SK is shipping silicon. Both matter, and the near-term revenue sits with what ships.

The supply number that matters most this weekend. Omdia via Chosunbiz: Samsung's conventional memory capacity grows 6% this year and just 1.3% next year, while SK adds roughly 10% a year. Separately, SK is scaling its 1c node from about 20,000 to 160-190,000 wafers a month by end-2026 — an eight-to-nine-fold ramp — while Samsung targets ~50% more HBM capacity with qualification, not wafers, as its binding constraint. The read is that the number-two producer is choosing mix over volume: pushing HBM, barely adding ordinary DRAM. Our own supply model assumes conventional DRAM capacity grows 12% in 2027. If Samsung adds 1.3% and SK adds 10%, that assumption is too generous — even allowing for CXMT and Micron. Logged as an asymmetry favouring the upside case rather than a revision, because it rests on a single analyst estimate reported through one outlet. It is the thing to confirm at the October print.

And the pricing item firmed up. The Information reports Nvidia AI chip prices rising about 17%, with Bloomberg putting servers above 15% — memory named as the cause. Micron's chief executive says there is "no end in sight" to memory catching demand. Both markets were shut when this landed.

1 · Etched is a customer, not a competitor. Sohu hardwires the transformer into silicon and claims 20× an H100 — but it is "a conventional reticle-size die packaged with HBM", carrying 144GB of HBM3E on TSMC N4P. Its throughput advantage comes from HBM bandwidth: roughly 4,800 GB/s against the H100's 3,350. One analysis calls it "a throughput machine, not a latency machine — the opposite of SRAM-based chips." Etched has raised about $1B with $1B of signed orders. Every Sohu shipped is 144GB of HBM sold.

2 · Taalas is the real HBM-free architecture — and the constraint is arithmetic. HC1 etches Llama 3.1 8B into the upper metal layers of an 815mm² N6 die: no HBM, no external memory, 250W, air-cooled, ~17,000 tokens/second. AMD bought it in August. But 8 billion parameters at 4-bit is about 4GB of weights. A frontier model at 1.6 trillion parameters is two hundred times larger — it does not fit in the metal layers of any die, at any node. The model is also frozen at fabrication, while frontier weights change every few months. Even the most enthusiastic write-up puts the odds of the roadmap working at 55-65%. Genuine threat for small, frozen, enormous-volume models. Not for the frontier.

3 · Photonic memory does not exist as a product. Q.ANT and its peers build photonic compute co-processors, and describe themselves that way — light performs matrix operations, it does not store state. Photonic interconnect is real and shipping, but it moves data between HBM systems faster; it does not replace them. There is no commercial optical DRAM.

4 · Intel's plans, with dates. XBM (patent filed Dec 2024, published Jul 2026) is an ultra-high-bandwidth memory using back-end-of-line transistors and serial UCIe links instead of a silicon interposer — attacking HBM's packaging cost, matching HBM4's footprint. ZAM, with SoftBank's SAIMEMORY, is a fusion-bonded nine-layer stack of largely conventional DRAM targeting roughly twice HBM4's bandwidth density. ZAM is slated for 2029-2030; XBM "at the start of the next decade." Business Post's read is that even if commercialised, ZAM would likely complement HBM in custom inference chips rather than replace it, because Nvidia controls 80%+ of accelerators with an ecosystem built on HBM. And the base rate is unkind: Intel has tried RDRAM (abandoned 2001), Optane (shut 2022), HMC and MCDRAM (never shipped), and sold its NAND business to SK hynix for $9B. Five attempts, five exits.

5 · What Terafab is actually building. Its job postings name the roadmap directly: "integration development for emerging memory technologies — MRAM, RRAM, or 3D DRAM — as Terafab expands memory roadmap", alongside sub-20nm DRAM capacitor work. So not HBM — magnetic and resistive memory plus 3D DRAM, with operations expected as early as 2028 and TrendForce reporting speculation that Intel could produce its next-generation memory there.

The pattern that resolves all of it. Every serious attempt to supersede HBM either buys HBM (Etched), works only for a narrow frozen case (Taalas), doesn't commercially exist (photonic memory), or is still DRAM wearing different packaging — ZAM is conventional DRAM fusion-bonded, XBM is a DRAM stack with a different interconnect, 3D DRAM is DRAM. The underlying requirement — enormous bandwidth to a large working set — is physics rather than fashion, and it keeps producing the same answer. The critical fact for this position is the calendar: 2029, 2030, "the start of the next decade." The forecast here runs to 2029. None of these arrive inside it. The honest counterweight: SK is not standing still either — iHBM thermal architecture, HBF with SanDisk, co-authoring the standards — but it also just lost the executive who built HBM to Intel's foundry business.

The news. Nvidia has told its largest customers that servers containing its AI chips will rise more than 15% in many cases, effective on systems shipping early next year, covering both Vera Rubin and Grace Blackwell, with the increase varying by generation and memory configuration. The stated cause is soaring memory cost.

The teardown that explains it, and the number is remarkable. UBS puts a Vera Rubin Superchip at roughly $39,000. Of that: HBM4 on the Rubin GPU costs $4,943 — 53.4% of the GPU alone — and SOCAMM2 on the Vera CPU costs $19,355, which is 96.4% of the entire CPU. Strip the memory out and the Vera CPU costs $704. Together, memory is 62% of the Superchip's cost, up from 53% in Grace Blackwell, with memory cost per Superchip rising 2.5× generation-on-generation against a total cost rise of 2.1×. A single Vera Rubin rack carries 74.7 terabytes of DRAM — roughly 4,500 smartphones' worth.

Now decode the price rise — and note that it inverts the Apple case. When Apple repriced products up 18-25%, memory was only 12-15% of the bill, so the implied memory increase was +125% to +250%. Nvidia is the mirror image: with memory at 62% of cost, a +15% server price rise implies memory up only ~24% if passed through in full (a +20% rise would imply ~32%). Same arithmetic, opposite leverage — and the Nvidia case is the more important of the two, because it demonstrates something the Apple case could not: when memory is 62% of your cost you cannot absorb an increase, you can only pass it on. Nvidia, the most profitable company in the supply chain, is choosing to raise prices rather than compress its own margin. That is pricing power flowing upstream, to the memory maker.

Four independent routes now bracket our own assumption. The model uses +55% for FY27 realised blended pricing. Against that: Counterpoint expects server memory prices to double by end-2026 and overall memory up 50% from current levels; Cantor has 2027 HBM4 contracted at $32-40/GB against ~$15-16 today, or +100-160%; the Nvidia pass-through implies a further ~24% on a base that has already moved 2.5×; and conventional DRAM contracts printed +90-95% then +58-63% quarter-on-quarter in the first half. Every route sits at or above +55%. The input remains conservative.

One new demand item worth its own line. Nvidia has pivoted from DDR5 to LPDDR (SOCAMM2) in AI servers to cut power. Counterpoint's read: that makes Nvidia "a customer on the scale of a major smartphone maker — a seismic shift for the supply chain which can't easily absorb this scale of demand." A buyer the size of Apple has just entered the LPDDR market from a standing start. And the macro marker: central bankers including the SNB's Petra Tschudin warned on Friday that AI-driven chip shortages could fuel near-term inflation — memory has become a variable that monetary policymakers now discuss by name.

The talks. Jensen Huang met Rebellions' chief executive at Santa Clara this week. Bloomberg reports early-stage discussions covering a technology licence, an equity investment, or an outright acquisition. Rebellions, founded 2020 in Bundang, designs neural processing units for data-centre inference. It has raised about $850M — investors include SK Hynix, Samsung, SK Telecom, Arm, Saudi Aramco's Wa'ed Ventures and the Korea National Growth Fund — most recently valued near $2.3B after a $400M pre-IPO round in March, with a Korea listing targeted for 2027 under JPMorgan. The Korean government holds a direct stake and treats the company as central to its "K-Nvidia" programme, so any acquisition draws scrutiny in Seoul as well as from US antitrust. It follows the Groq template from late 2025: a non-exclusive licence plus most of the engineering staff.

But the specification is the story. Rebellions' flagship Rebel100 is built on Samsung 4nm with a four-chiplet architecture — and carries 144GB of HBM3E. This is not an SRAM-only decode machine. The Korean national-champion inference chip, the one Nvidia may buy to strengthen its inference position, is itself a large HBM customer. That is the third independent confirmation in a week that the "engineer around HBM" thesis keeps failing on contact with silicon: Cerebras needs HBM systems paired beside it to hold context, more racks means more bits rather than fewer, and now the specialist NPU ships with 144GB of stacked memory on board. Every architecture that tries to route around high-bandwidth memory ends up carrying it.

What it does for the position, honestly sized. SK is an investor, so a deal above the $2.3B mark books a modest gain — immaterial against a company worth roughly $931B, and if regulators block it, the 2027 listing monetises the stake anyway. The strategic read is worth more: Nvidia keeps absorbing inference capability rather than being displaced by it — Groq in 2025, Rebellions now — and each absorption preserves the heterogeneous architecture with HBM at its centre. One footnote from the same filings: Nvidia's public equity portfolio has reached $63.4B, and its largest single holding is Intel — about 214.8M shares worth some $30B, or 47% of the book — with a $21B SpaceX stake confirmed for the first time. The company at the centre of this shortage is now also one of its ecosystem's largest shareholders.

The shortage, quantified by the competitor. "Data-centre customers are seeking about 50% more memory supply than Micron can currently commit to delivering." And more bluntly: "All our customers across our end markets will buy everything that we make." On the industry's character: "Today there is no AI without memory. The value of memory — that equation has totally changed", and "memory is no longer a component in a system, it's strategic infrastructure for the AI era." He expects the crunch to run beyond 2027. The contractual proof sits in the accounts: 16-plus multi-year take-or-pay agreements backed by roughly $22B of customer cash deposits and commitments — at Micron alone, against the $38B industry prepay figure carried here. Micron's fiscal Q3 printed $41.46B of revenue with a record $18.3B of free cash flow, and guides Q4 to $50B — up 21% in a single quarter.

Now the comparison that matters, because Micron's own share price sets it. Micron carries a market capitalisation near $1.3 trillion. SK Hynix, at Friday's close, carries roughly $931B. Yet SK's most recent quarter was $57.3B of revenue against Micron's $41.46B — 38% more — with 57% of the HBM market against Micron's ~21%, and the industry's lowest production cost. On price to annualised revenue: SK 4.1× versus Micron 7.8×. Micron trades at nearly twice SK's revenue multiple while selling a third of the HBM. Priced on Micron's own multiple, SK would be ₩3.34M today — 93% above Friday's close, before a single won of the 2027 earnings growth arrives. That is not a forecast; it is what one competitor's market price implies about the other, today.

Two calibration items that run against us, corrected rather than buried. First, Edgewater reports 2027 memory pricing is unlikely to be finalised until late 2026 — the calendar here had the HBM-2027 settlement at end-September, which is too early. The pricing confirmation now sits in Q4, not September, which pushes weight onto October's quarterly print as the nearer test. Second, Micron's Boise fab produces first chips "in the middle of next year" and ramps primarily in 2028 after qualification — meaningful competitor supply arrives roughly a year earlier than the 2029-31 framing used here. Alongside it, YMTC pulled its Wuhan Phase III NAND completion forward from 2027 to the second half of 2026, and Micron's capital spending steps from ~$28B this fiscal year to ~$47B next. The supply response is being funded harder and arriving sooner than modelled.

And the structural change worth marking. Nvidia has now signed multi-year supply agreements with both SK Hynix and Micron, covering conventional DRAM as well as HBM — securing capacity rather than price. Reporting adds that Samsung and SK have stopped signing one-year agreements altogether, preferring three-to-five-year commitments. An industry that used to reprice every twelve months now contracts for half a decade. Mehrotra's claim is that AI has broken the boom-bust cycle outright; the honest position is that it has lengthened it — which is precisely why the modelled path here plateaus in 2029 rather than collapsing.

The chain, in order. Leopold Aschenbrenner — 25, ex-OpenAI, no prior trading experience — launched Situational Awareness LP in July 2024 on a thesis that will sound familiar: that increasingly powerful AI models would trigger an unprecedented expansion in demand for computing infrastructure, including advanced chips, memory, data centres and energy. That is the thesis's thesis, almost word for word. The fund was up 439% net through 30 June and had grown to $45B. It ran leverage reported at roughly 4×. SK Hynix was one of its largest longs, held through an approximately $7B cornerstone indication at the 10 July US listing — and one account traces the whole unravelling to that listing.

Then July happened. The SOX fell 28.6% from its 22 June peak and the Morgan Stanley Momentum TMT index dropped 53.5%. The fund's roughly $16B public book lost 67% of its value in a single month. Goldman Sachs, JPMorgan and Bank of America issued margin calls. On 30 July the entire public equity portfolio — longs and shorts — went to Citadel in one block before the open. The fund shrank from $45B to about $10B. Six days earlier Aschenbrenner had written to investors calling July "some of the most attractive opportunities since early 2025" and asked for fresh capital. It never arrived. His letter afterwards: "We let you down."

So yes — forced selling was a major driver, and the dates line up exactly. SK Hynix bottomed at ₩1,320,000 on 30 July — the very day the book transferred — then rose about 30% the next session. That is the signature of a mechanical seller finishing, not of a fundamental repricing. Then came the second leg: Citadel spent August distributing what it had bought. Griffin's investor letter of 21 August discloses the firm has unwound more than 80% of the portfolio through over 100 block trades representing more than $4B of market value, including single-session records in ten separate names. That distribution is why rallies kept stalling.

Answering the question directly: no, not fully — about 20% remains. More than 80% is done, which is why Gavin Baker judges the forced selling "most likely over", and why our own note logged it as a market-wide flow observation rather than an SK call. The residual is a shrinking overhang, not a thesis problem. Worth separating clearly: the July collapse was this liquidation; the 18 August −9% was a different event entirely — a rates shock with the 30-year at a 19-year high and uniform damage across the whole complex. Two mechanical flow events and one macro event, none of them about SK's business.

And the lesson, which matters more than the flow. Aschenbrenner had the right thesis, on the right stock, at the right time — and was destroyed anyway, because 4× leverage turned a drawdown into a liquidation. He was up 439% and finished as a forced seller of everything at the lows, while the assets he was right about recovered without him. This is the ergodicity argument in its purest form, and it is the reason the thesis's discipline reads "the sizing is not the risk; the path is" and refuses to add at a maxed position. Same thesis, same stock, different survival — and the only variable was leverage.

First, what the move was and was not. Seoul closed ₩1,730,000, up 2.3%, with an intraday high of ₩1,773,000 and after-hours at ₩1,761,000. But the KOSPI rose 0.9% and Samsung rose 3.9% — Samsung more than SK. This is the Korean memory pair re-rating together, not a company-specific event. The US receipts moved only +0.2% to $163.41 while Nvidia fell 1.0% and Micron 0.8%, so Monday's Seoul open should look softer purely as catch-up to a US session Korea missed. Anchor updated across the site: ₩1,645,000 → ₩1,730,000. Targets are unchanged; the starting line moved, so every upside percentage compresses — the 12-month median reads ₩5.12M (+179%), and forward P/E is 3.1×.

The item that actually matters: the company is buying its own stock at nearly double the scheduled pace. KRX flow shows the "other companies" category taking roughly ₩1.04T on Thursday and ₩1.13T on Friday. Spread evenly, a ₩40T programme is about ₩0.65T a day; this is running near ₩1.1T — about 1.7× faster. At that rate the programme exhausts in roughly 36 trading days, finishing in early October rather than 19 November. Yuanta reads it as a floor under the shares, and the arithmetic supports that: ₩40T is 3.2% of the company being withdrawn from the float inside two months. The honest caveat: "other companies" is a KRX flow bucket, and the attribution of most of it to SK is analysis, not disclosure. Watch whether the daily print continues — that is the falsifiable part.

Samsung set the benchmark, and it corrects one of our own expectations downward. After Friday's close Samsung formally voted to return ₩90-110T this year, including ₩30T of cash this quarter. Samsung's market capitalisation is roughly twice SK's, so a proportionate SK return sits near ₩45-55T. The street sketch the thesis logged on 17 August put SK at ₩100T — that now looks roughly twice too high. With ₩40T already running, ₩40-55T is the realistic frame for the late-October announcement. Expectation corrected before it could disappoint. Samsung slipped after hours precisely because the size was hoped for and the remainder waits until January.

The level worth marking. Friday's high of ₩1,773,000 punched through ₩1,744,000; the close did not, falling ₩14,000 short — 0.8%. That is the second failure in four sessions, after Tuesday's ₩1,792,000 high closed at ₩1,662,000. A level tested twice and rejected twice is a real one, and it now sits 0.8% away. Fundamentals from the same weekend, all pointing one way: Korean DRAM export prices rose again in early August; Micron is reported to be filling only about half of data-centre orders — the same squeeze from the competitor's side; SemiAnalysis confirms HBM4's base die is now in real factory volume at Samsung, which validates the technology ramp while sharpening the share contest; and Gavin Baker reports a large Citadel book mostly unwound, with forced selling "most likely over" — a market-wide flow observation, not an SK call. Two open items: the worker vote runs Monday 6am to Tuesday 9am Seoul and is not signed, and the Japan-factory chatter was officially denied.

What happened. On 10 August Nvidia signed memorandums of understanding with Apollo, Blackstone, BlackRock, Brookfield, Goldman Sachs and KKR to build financing platforms mobilising more than $500B of third-party capital — using "institutional credit, insurance funds and private capital to underwrite GPUs and data centers" so that hyperscalers, frontier labs and enterprises can build without touching their own balance sheets. Nvidia's framing of its chips is the tell: "revenue-generating assets" that are "productive, long-lived, fungible and flexible." Huang put it plainly: "This is really the first time that technology chips have become an investable asset class." Then on 17 August the enabler landed — the SEC sided with Latham & Watkins that certain data-centre debt sits outside the securitisation rules requiring sponsors to retain risk. Those are the Dodd-Frank risk-retention rules written after 2008, and the friction they created is now gone for this asset class.

Why this dismantles the bear case we ourselves logged. Watch W4 — the credit and financing channel — assumed the funding pool was hyperscaler cash flow plus investment-grade corporate bonds. Large, but bounded, and visibly stretched: PIMCO had capex at 94% of operating cash flow, and corporate spreads were widening. Securitisation changes the denominator entirely. It reaches insurance general accounts, pension funds, private credit and the whole structured-credit buyer base — a pool measured in tens of trillions, not hundreds of billions. And the proof it is already working: Morgan Stanley reports data-centre ABS spreads sitting near year-to-date tights even while hyperscaler corporate spreads widened, with Huberty attributing the corporate sell-off to "supply technicals rather than a deterioration in data-centre fundamentals." The securitised channel is not transmitting the corporate credit stress. Issuance has gone $2.4B (2020) → $15.5B last year, on pace for a record, and JPMorgan projects $30-40B a year by 2027 — 7-10% of the combined ABS and CMBS markets. Separately the BlackRock/Microsoft/Nvidia partnership bought Aligned Data Centers for $40B. W4 is downgraded from live to monitored. The buildout is no longer capped by what buyers can afford out of earnings.

And now the part that must be said in the same breath. This is the 2008 machinery, rebuilt, and the people building it are not hiding it. The BIS warned in January (Bulletin 120) that AI companies are shifting from cash-flow to debt funding and that the opacity of private credit makes true systemic exposure impossible to assess, with total IT-related investment now around 5% of US GDP — above the dot-com peak. The risk is being distributed to pension funds, insurers and retail through chains that obscure the collateral's depreciation curve. That collateral is the weak point: CoreWeave's $7.5B GPU-backed facility carries roughly 11% interest with repayments starting exactly as chip resale values fell, SPV structures keep the debt off balance sheet, and 2027 brings the first major GPU lease-renewal cliff — if Rubin makes Hopper inference uneconomic, liquidation values collapse. Net for the thesis: near-term the funding runway is vastly longer than our model assumed, which strengthens the durability of the base case rather than its level. Late-cycle, this is a named systemic fragility landing in 2027-29 — the same window as our supply wave — and it is logged there rather than pretended away. No probability mass moves today: less near-term funding fragility, more late-cycle systemic risk, and those roughly offset.

The finding. "SK hynix is now adopting contracts that eliminate the industry-standard price cap, marking a departure from Micron's approach… SK hynix's new structure allows spot market price increases to be fully reflected in contract pricing when supply shortages push market prices higher." Micron, in its Strategic Customer Agreements, capped existing products at the Q2-2026 market price with a floor beneath. Two suppliers, the same shortage, and only one of them collects the upside.

The industry-standard structure has changed. Our elasticity card stated that ~65% of the book sat under contracts with ceilings, and that ceiling is precisely what held modelled realised pricing at +42% against +100% theoretical. For SK, the ceiling is gone. Corroboration: Cantor reports SK's 2027 HBM4 capacity sold out at a price ladder — Nvidia $32/GB, Broadcom $36, AMD $40 — against roughly $15-16/GB today, so +100% to +160%. Three buyers, three different prices is a seller pricing customer by customer, not a fixed contract. And the printed commodity moves we had not fully absorbed: TrendForce revised Q1-2026 conventional DRAM contract prices up to +90-95% quarter-on-quarter, then +58-63% in Q2; Counterpoint has global memory +80-90% QoQ in Q1. Compounded, that is roughly 3.1× on commodity DRAM in one half-year.

The counter-evidence, kept. Odaily reports the LTAs "enhance order visibility but limit short-term price increase gains", and SK's own President Song describes pricing that "varies by customer and product characteristics, but is designed to respond to price fluctuations", with deposits enforcing compliance. So this is not pure spot exposure — it is a mechanism that tracks the market with a lag. The correction is therefore real but bounded: FY27's realised-price input moves +42% → +55%, not to the theoretical +100%.

Cars, EVs and robots — verified, and the Hormuz channel is real. Per-vehicle memory runs 90GB (2025) → 278GB (2026), L4 autonomous 300GB+, and next-generation EV and autonomous-ready platforms potentially 2TB per vehicle. Critically, autos run on DDR4 and LPDDR4 — the most supply-starved segments — and car-grade DDR4 is already +70% year-on-year. On the oil linkage: JD Power confirms higher oil prices support EV adoption, and while overall light-vehicle sales decline marginally in 2026, EVs grow 7.8% globally to a 25.7% share, with Europe +23%. So the Hormuz spike that knocked the share price on Tuesday is, on a longer horizon, a demand tailwind for the highest-memory-content vehicles. Omdia frames the whole picture: AI demand drives a 94.1% surge in the 2026 semiconductor forecast, with automotive and industrial explicitly "competing with AI demand for packaging and memory ICs."

The car story, sized rather than admired. Micron puts the average vehicle's combined DRAM and NAND requirement rising from 90GB to 278GB — a 3.1× jump in content per car — as architectures shift to centralised computers. Ford's CFO told investors DRAM increases add roughly $1B to Ford's 2026 cost sheet; GM and Volkswagen have flagged the same. Average US new-car prices sit near $50,000 and analysts see $60,000, with Telemetry putting the shortage's direct contribution at about $2,000 per vehicle. Priced out: 90M vehicles at roughly $350 of memory each is about $31B of annual demand, of which the content tripling contributes some $21B incremental — around 2.4% of the memory market. Our flows model already carries this lane at $39B, so the magnitude was right; what was missing was the hard content number, now on the record. It sits inside the existing +24% bit-growth assumption and is not a reason to raise the chain. What it does add is a demand quality argument: a car ships with its memory or it does not ship, and Ford is absorbing $1B rather than pausing production. That is inelasticity demonstrated, not asserted.

The line that matters more, and it is the second independent report in two days: "All three major memory manufacturers — Samsung, SK hynix and Micron — have sold out their entire production capacity for 2027." If that holds, FY27 revenue becomes largely contracted price × contracted volume rather than a forecast. The practical consequence is not a higher number — it is a narrower band around the number we already publish, and it is the datum that would settle the open tension between our two models (the bottom-up flow model implies +32% realised pricing for 2027, the revenue chain uses +42%). October's Q3 print is where that resolves. Alongside: DDR5 kits up as much as 485% year-on-year in the US — confirming the ~500% figure we logged as unverified chatter two days ago — and Micron's quarterly profits running roughly 15× year-on-year with its stock up 242% in 2026. Cramer cites $93.9B of multiyear supply agreements across eight clients, against the $38B prepay ledger the thesis has carried; logged with the source caveat that it is a broadcast figure, not a filing.

The honest negative, and it deserves its own paragraph. The same reporting raises the bullwhip effect: shortages make buyers over-order out of fear, which inflates apparent demand, and the last chip shortage ended in exactly that glut. This is our inventory-correction regime, already carrying 6% of probability mass. What is new is not the risk but its arrival as a mainstream narrative — when the New York Times and Jim Cramer are both discussing whether the demand is real, that becomes a sentiment risk independent of the physical facts, and it is the kind of story that moves a multiple before it moves an order book. Watched, not dismissed.

The launch is real. CS-4 keeps the same 5nm WSE-3 wafer but doubles clock speed and power delivery, doubling memory bandwidth and therefore roughly doubling tokens per second per user — ~4,000 tok/s/user on frontier models against ~2,000 on CS-3 and a realistic ~100 for Blackwell. Three wafers per rack now instead of two, off-wafer I/O up from 1.2 to 2.4 Tb/s, network latency down to 2-3 microseconds, and a modular "backpack" design that should hold bill-of-materials roughly flat. For customers that is close to double the token revenue at similar total cost — a genuinely strong product, and the roadmap promises 2× a year to 20× throughput by 2027.

But the headline number is bandwidth, not capacity, and that distinction is the whole story. The marketed 43 PB/s — "2,000× Rubin's memory bandwidth" — describes how fast data moves on the wafer. What did not change is 44GB of SRAM per wafer, unchanged from CS-3 because it is fixed by the number of bit cells on the silicon. SemiAnalysis calls this out plainly as "the main drawback… low memory capacity per wafer is one of the key tradeoffs inherent with Cerebras's architecture." A frontier model does not fit in 44GB. Their own arithmetic: running a 1.6-trillion-parameter model at 1M context needs about 20 wafers, and at a realistic 256 concurrent requests about 40 wafers — over $20M of capital and 1MW of power before you get a single forward pass. They expect customers like OpenAI to cut context from 1M down to 256K purely to avoid holding KV cache on-wafer.

So Cerebras does not remove memory demand. It relocates it — into HBM. The article states the design intent directly: the new field-upgradeable I/O module exists so that "disaggregated inference setups will go a long way to help overcome the memory capacity constraints of the CS-4 by pairing it with HBM-based systems" — explicitly "in the same way that Nvidia is positioning the Groq LPUs." In every partnership named — AMD, AWS Trainium — Cerebras serves as the decode chip only, because its rooflines are wrong for compute-bound prefill. This is the identical architecture we logged when Nvidia's own LPX rack pinned attention and KV cache to Rubin's HBM and offloaded only frozen feed-forward weights to SRAM. Two independent designs, same conclusion: SRAM does the fast small thing, HBM holds the model and the context.

And the agent economy makes this worse for SRAM, not better. The piece spells out the mechanism: KV cache scales with concurrent users × context length — precisely the two variables agents multiply. Persistent per-user context across billions of agents is a capacity problem, and capacity is exactly what wafer-scale cannot buy. Their own observation that "agentic cache hits drive prefill costs down while decode costs remained the same" confirms the split rather than eliminating the memory.

The honest risk, kept. CS-4 is shipping, OpenAI is evaluating purpose-built inference systems, and if decode migrates to SRAM at scale then HBM content per decode token falls — that is our decode-substitution regime, already carrying 7% of probability mass. Two things in this article argue for holding that weight rather than raising it: performance per watt is "at best a slight improvement" at 125-135kW per rack, and SemiAnalysis flags the structural flaw in all disaggregation — the prefill-to-decode ratio is fixed the day the purchase order is signed, while a GPU fleet can be reallocated as workloads shift. They note workloads have already shifted twice. Net: no model change. The bear case for HBM does not run through Cerebras; Cerebras is one of HBM's new customers.

The two facts, verified. On 27 July a state-backed Shanghai maker (Yuliangsheng, later named as Aishengna) began mass-producing China's first homegrown immersion DUV lithography machines — roughly five units in 2026, twenty in 2027 — with first deliveries to SMIC, Hua Hong and CXMT. The same day, CXMT listed on Shanghai's STAR Market, raising $8.6B (Asia's largest IPO this year, second-largest in A-share history) and closing day one up 466% at a market capitalisation near $487B — instantly the most valuable company on the A-share market, ~$523B a week later. Q1 revenue was RMB50.8B, up over 700% year-on-year, on a 2025 global DRAM share of 7.67%, with proceeds earmarked "mainly for mass-producing memory wafers." Tom's Hardware reports CXMT is close to matching Micron's memory capacity in 2026.

Why this matters here, and it is uncomfortable. The thesis published a chart arguing new supply cannot arrive early because ASML builds only 48-80 EUV machines a year. That argument is sound for leading-edge logic — but DRAM at 1a/1b nodes is built largely with DUV immersion and multi-patterning, and design-technology co-optimisation reaches feature sizes that would otherwise demand EUV. The China supply channel is therefore not gated by ASML's order book the way we implied. Worse for the argument: domestic tooling neutralises US House Resolution 8170, which would otherwise cut SMIC, Hua Hong, CXMT, Huawei and YMTC off from ASML by statute. That is an honest weakening of one of our own published claims,.

The counterweights, sized equally honestly. Five tools this year and twenty next is small — a single leading-edge fab needs dozens. Reuters reports the machines still trail ASML on performance and reliability and require further validation. Critical components remain Japanese, an unresolved chokepoint. And the decisive one: CXMT sells conventional DRAM, not HBM — no TSV stacking at scale, no advanced-packaging qualification, no Nvidia or hyperscaler approval. Our profit engine is not directly contested. The correct transmission is narrower and slower than the headline: China floods commodity DRAM, and SK's non-HBM lane compresses.

What actually happened, and it has nothing to do with SK Hynix. The 30-year Treasury yield hit a fresh 19-year high near 5.33%, the 10-year reached 4.75% — its highest in 19 months — and Japan's 10-year touched a three-decade high. Simultaneously, US-Iran talks over the Strait of Hormuz deadlocked, pushing WTI above $85 and Brent above $91. The SOX fell 5.4%, shedding roughly $680B of market value in a session. The damage was uniform across the complex: SanDisk −9%, Seagate −9%, Western Digital −7%, Marvell −8%, the Roundhill Memory ETF −7.9%, Micron −5 to −7%, SKHY −9.2%, Frankfurt HY9H.F −8.5%. One desk summarised it exactly: "uniform sector declines confirming a macro rate reset rather than any company-specific news."

The mechanism, stated plainly by the wires: high government bond yields "lower the present value of future tech profits and increase corporate borrowing costs." That is multiple compression — the tape sold the multiple, not the earnings. Memory fell hardest precisely because it had run furthest (Micron +255% year-to-date, SanDisk +653%): in a rate reset, the biggest winners get marked down the most. Nothing about supply, demand or pricing changed on Tuesday.

This is the scenario our own macro work flagged, arriving on schedule. The liquidity analysis integrated last week named two triggers: long-end Treasury yields and oil. Both fired at once. Watch W4 — the credit and financing channel — is now live rather than theoretical, and the question it poses is the right one: does the AI buildout's debt funding survive a genuinely higher cost of capital? Nvidia's $25B at 3.4× oversubscribed says yes for the giants; the second tier is where to look for cracks.

Same-day counter-evidence, all of it fundamental. Beth Kindig reports customers naming DRAM as the number-one constraint — ahead of power, land and logic wafers. Daniel Newman makes the argument that directly answers our HBM-content-per-token risk: less HBM per GPU means more racks, not fewer bits, and the "engineer around HBM" thesis has not scaled. And the item that matters most, now confirmed by Digitimes: SK's second-largest customer spent ₩17.19T in H1 against the largest at ₩17.61T — nearly identical, and consistent with Yonhap's ~₩17T Nvidia print. SK is no longer a one-customer story, which cuts directly against the customer-concentration risk its own SEC prospectus flags.

Where Korea stands this morning. Tuesday's KRX session closed at ₩1,662,000, up 1.03%, after a violent range: open ₩1,736,000, high ₩1,792,000, low ₩1,640,000. The high cleared the ₩1,744,000 gate but the close did not — so Model A's rule has still not fired, which is exactly why it is a close rule. Korea has not yet priced the overnight −9%: a full catch-down implies roughly ₩1.51M, while the Frankfurt line implies about ₩1.59M. Two local items remain unpriced: labour talks ended 18 August without a printed settlement, and Gyeonggi province has demanded tighter wastewater rules at Yongin with reports of no night work at the cluster — construction friction, not a confirmed delay, but it belongs on the Yongin watch. No tripwire has fired. A price move is not a thesis change.

What was said. Asked about Apple sourcing memory from CXMT and YMTC, the Commerce Secretary said: "The Trump administration is not in favor of that. There have to be other solutions to the memory issue, but it's not great American companies using Chinese memory." Asked whether he had put it to Apple directly, one word: "Plainly." Apple had petitioned Commerce on 27 June for clearance on CXMT and, critically, for an assurance it would not be added to the Entity List. That request has now been answered in public, and the answer is no. Micron and senators from states hosting new fabs have been lobbying for exactly this outcome.

Why it matters to us. Apple is a top-tier memory buyer trying to escape the very price rises this thesis is built on. If it had been allowed to buy Chinese DRAM at scale, that volume leaves the incumbents' order book at the precise moment 2027 contracts are being priced. Blocked, Apple's demand stays with SK, Samsung and Micron — and the political headwind also slows CXMT's Western expansion, which is the China threat we quantified last week at "under three years" of technology gap. The ADR rose 3.04% to $171.38 (high $178.43) on Monday with Micron up 4.1%, while Korea was closed for the holiday.

Now the part the tape is not pricing, and it is the part that matters. Lutnick expressed a preference, not a rule. Commodity memory requires no shared design information, so no export licence is needed and no legal mechanism exists to stop the purchase. As one analysis put it precisely: he has a strong opinion and a podium; what he lacks is a lever. The WSJ's own second sentence reads "Apple may do so anyway." Apple's COO has already said memory involves "not a lot of customization" — the technical argument for substitution, made by the buyer. And the dam already leaks: HP and Acer are shipping CXMT memory today in products sold outside the US, and nobody stopped them. The only development that would convert opinion into rule is CXMT being added to the Entity List — YMTC already is — and that is now the item to watch, not the speech.

Positioning arithmetic for this morning. Korea reopens at ₩1,730,000 having missed the move. A matching +3% opens around ₩1.695M — Friday's high, and still ~2.8% below the ₩1,744,000 gate that Model A tracks. Even a full catch-up does not clear it today. Against that sits an unpriced local item: labour-management final talks are today, and a settlement or a breakdown will move the open more than Washington will.

Logged but unverified this run, and flagged rather than absorbed: consumer DDR5 reportedly up ~500% year-on-year with 128GB kits at $3,399, and a claim that hyperscalers have prepaid almost all 2027 DRAM — that last one would be enormous if confirmed, since it would mean next year's revenue is largely contracted already. Also noted: SK reportedly resuming Dalian/Solidigm investment for +50% output by 2027 on enterprise SSD demand, and a calendar item worth keeping — Micron's Sumit Sadana speaking in November on "memory per FLOP" and on the shift from old-style LTAs to Strategic Customer Agreements. "Memory per FLOP" is precisely the HBM-content-per-token risk already on the thesis.

Three genuinely new items, and the third is the most useful thing on this card.

1 · Consumer memory is up ~90% in a single quarter. HBM stacks consume so much DRAM capacity that they are starving the consumer lane outright — which is why iPads and MacBooks repriced. Two duration quotes land alongside: AMD's Lisa Su at Computex says DDR5 will not normalise until 2028, and a Lenovo executive at ISC says prices will not return to pre-surge levels for at least five years. Our Apple decode inferred +125-250% on the memory line; the printed quarterly move is now ~90% on its own.

2 · Micron's new capacity is already pre-sold — a real mitigant on the 2029 wave. Micron's US commitment is $250B through 2035 (the $50B Boise and $100B Clay projects sit inside it, resolving the figure discrepancy: both numbers are right at different scopes), targeting 40% of DRAM output on US soil. But the detail that matters: Micron has locked that future capacity under non-cancelable customer agreements through 2030. Supply arriving pre-committed cannot flood a spot market the way 2018's did. This does not reverse the bear regimes added last week — one data point should not undo a structural fix — but it is logged on row 7 as the fourth mitigant beside contract floors, buyer cash flow and the robotics wave.

3 · SK is the cheapest producer — the moat that survives losing share. Patel's assessment: SK Hynix has the best margins in the industry, cheapest to manufacture, even in places where Samsung edges ahead technically on HBM4. This is materially better news than a share statistic, because it changes what share loss costs. Our tripwire T3 assumes falling share compresses earnings; a low-cost producer losing share to a higher-cost rival compresses far less, and survives a price war the rival cannot start. Logged as a structural offset to T3 — not a reason to disarm it.

Patel's bear half, stated straight, because it is now our own base case too: Micron and Samsung's capacity lands ~2028-2030 and "we cannot avoid a cycle." That is precisely what last week's rebuild priced — six bear regimes at 38.5% of probability mass and a 2029 plateau. The most credible sceptic in the field and this model now agree on the shape. His bull half: if AI demand holds for roughly a decade, memory makers are among the cheapest stocks in the entire AI trade, with ten years of earnings ahead.

Verified detail from the visit: Yongin's first fab has six cleanrooms stacked across several floors rather than sprawling single-storey like TSMC's and Intel's Arizona sites. The US plant is West Lafayette, Indiana — $4B, packaging only, construction concluding 2028, not front-end. SK also restructured Solidigm (the $9B Intel NAND purchase from 2020) and launched a $10B "AI Company" in the US in January, with no front-end manufacturing planned. And the China number to keep: CNBC reports China's memory gap with Samsung and SK has narrowed to under three years — the CXMT threat, quantified, and the wildcard sitting beneath everyone.

The correction first, because it is ours. This site has carried "₩100T net cash arriving early" as a support for the buyback leg. The 17 August filings wrap prints cash ₩88.0T, debt ₩18.6T, net cash ₩69.4T. Daishin's Ryu ties a ₩100T return to an AGM net-cash goal he describes as already beaten — that does not match the printed balance sheet, which sits roughly 30% below the marker. The tailwind row has been corrected to the printed figure. The returns case survives on cash generation and management's stated intent, not on a threshold that has been crossed.

What is genuinely new and positive: Korean media have pulled the shareholder-return framework forward to possibly August — a month ahead of the JPMorgan end-September window — with a street sketch of Samsung ₩200T and SK ₩100T a year. None of this is IR-confirmed, and the distinction matters: a street sketch is a hope with a byline.

Solidigm, newly on the board. Chatter values SK's NAND subsidiary near ₩50T, with up to ₩10T raisable pre-IPO. Toss flags dilution; Mirae reads a limited hit plus roughly $15B of US firepower for the American fab ambition. SK confirmed nothing on 5 August. Watch item: W6: a Solidigm monetisation would fund returns and US expansion without new debt — materially relevant to the credit-channel risk already on the thesis.

Flows, both directions. The national pension fund net bought ₩1.53T of 000660 between 1 July and 14 August while net selling ₩0.96T of SK Square — buying weakness in the operating company. Against that, margin credit stands at ₩4.81T, up 20.9% from end-July: leverage rebuilding into a rally is a fragility, not a confirmation, and it is the same instrument that amplified the July drawdown.

The share-drift item, and it belongs with T3. SemiAnalysis previously had Micron's Nvidia HBM4 share at zero, splitting the platform 70/30 between SK and Samsung. Nvidia certified all three suppliers for Vera Rubin on 5 June, and the monitor reports Dylan Patel now has Micron in. Three qualified suppliers is a different pricing world from two — this compounds the Samsung 80%-yield signal already on file and tightens tripwire T3. Alongside it, Beth Kindig notes AWS's 2026 GPU is GB300 and Meta runs Blackwell/Rubin plus AMD Helios this year, with custom MTIA arriving 2027: this year's HBM remains Nvidia-rack heavy, but the buyer set diversifies from next year.

Two items this run could not verify and they are flagged rather than absorbed: a "Friday Nvidia guarantee cut" and the precise wording of the Micron-in note. Both are sourced to the monitor rather than to primary documents. Prices confirm our anchor exactly: 000660 ₩1,730,000, NXT ₩1,667,000, SKHY $166.33, USD/KRW 1,418.50.

The sequence, precisely. Yongin's first fab (Y1) opens its first cleanroom in February 2027 — pulled forward from May, as reported. Equipment installation begins Q2 2027, and Y1 is built across six cleanroom phases, each adding floorspace and tools. The capacity figure is the one that matters: Y1 alone reaches 360,000 wafers a month — by the first half of 2030, en route to SK doubling DRAM wafer capacity by 2030-31. On 7 August the board approved a further ₩54.3T: Y2 in Yongin (₩35.2T, first cleanroom June 2029) and M17 in Cheongju (₩19.1T, December 2028), inside a ₩1,100T long-term plan with ₩600T for Yongin alone. H1-2026 capex was already ₩18.3T, up 73%.

For 2027 turnover: less than the headline implies. A cleanroom precedes equipment, which precedes qualification, which precedes volume. The supply actually reaching 2027 revenue comes from M15X in Cheongju, already deploying wafers, and 1c node conversions at M14 and M16 — both already inside our +24% bit-growth input. Yongin is a 2029-2031 supply event, not a 2027 revenue event.

For 2027 profit: first-order, it subtracts. Depreciation and start-up costs begin when assets enter service, ahead of meaningful output — the classic new-fab drag, and the risk the thesis has flagged as the main 2027-29 earnings compression. Traced explicitly rather than assumed: D&A stepping roughly ₩31T → ₩47T → ₩68T across 2026-28 costs about ₩16K then ₩21K per share after tax — 3-4% of EPS a year. Our chain already applies a 7.6% haircut in FY27 and 7.3% in FY28 versus flat-margin scaling: roughly twice the cover required. The drag is in the numbers; no revision needed either way.

The bullish read is about conviction, not capacity. A company does not commit $38B to plants that cannot ship a chip before December 2028 unless it expects the shortage to outlast them. That is SK betting its own balance sheet on the thesis's calendar — worth more than any analyst target. The counterweight, on the bear row where it belongs: 360,000 wafers a month from Y1 by 2030, with Y2 and M17 behind it, is exactly the supply wave the model prices as the 2029 plateau. Yongin does not lift 2027. It confirms 2029.

What launches. Two contracts — Silicon Data H100 Rental Index Futures and B200 Rental Index Futures — each representing one month's rent of a single GPU, priced off the first daily benchmarks of on-demand rental rates. They will list on NYMEX, the oil exchange, under CME's Global Head of Energy and Environmental Products, and CME's chief executive calls compute "the new oil of the 21st century." The institutional classification matters as much as the product: compute is being formally filed alongside crude and power, not alongside software.

Does it touch SK's revenue? No — it trades GPU-hours, two layers above memory. But four second-order effects are real, and the first one is the most valuable thing on this card.

1 · It repairs the exact risk we added yesterday. A datacenter takes two to three years to build, and until now its developer had no way to lock in what that compute would earn — so lenders priced the uncertainty into the debt. With a forward curve, revenue can be hedged, projects become bankable, and the cost of capital for the marginal datacenter falls. That marginal, debt-funded build is precisely the fragility behind Watch W4. Compute futures are financing infrastructure for the buildout — more projects clear, more memory gets bought. Alongside it, CNBC reports Nvidia is working with major asset managers on a vehicle that could channel up to $500B into AI infrastructure. If these markets gain real liquidity, W4 gets downgraded.

2 · From October 5 we get a free, daily, market-implied read on our own thesis. A forward curve in contango — future compute dearer than spot — is the market pricing continued scarcity, which means more buildout and more memory. Backwardation would be the first market-priced warning that relief is coming. That curve is now on the monitoring list, and it prints months before any earnings report.

3 · Futures relocate volatility, they do not remove it — from operating businesses to speculators. Mildly good for SK: customers who can hedge can plan capex with more confidence.

4 · Will it stabilise memory prices? Almost certainly not — and the reason is the best news here. Futures markets have been tried for DRAM before and failed, alongside onions, uranium and bandwidth. They fail for two reasons: concentration and non-interchangeability. Memory has both in abundance — a three-firm oligopoly selling product that is speed-binned, density-specific, customer-qualified and, for HBM, co-designed with the buyer. The very reason DRAM futures failed is the reason SK has pricing power. A commodity you can standardise and exchange-trade is a commodity whose producers are price-takers; memory resisted standardisation, and that resistance is the moat. Note how hard even compute is proving: the same study found H100 performance varying up to 34.5% across providers running identical workloads. If GPU-hours are that unruly, HBM stacks are a different universe.

The risk to name, dated far out. If compute futures work, financial engineering moves down the stack and someone will try memory futures again. Standardised, liquid memory contracts would erode bilateral pricing power at the margin. Not near-term — HBM is co-designed and qualified per customer — but watch item: W5. And one capex datum worth carrying forward: AI capital spending reached $765B in 2026, passing oil and gas at $681B for the first time, and is projected to nearly double by 2031. Our model's capex input moves from $732B to $765B. AI is now the largest capital-formation project on earth — and memory is its scarcest input.

What actually happened. Nvidia — a company generating $48.6B of free cash flow in a single quarter — sold $25B of bonds on June 15, its first since 2021, drawing $85B of orders and upsizing from $20B, with tranches running to 2056. It was not a funding rescue; it was a company locking three decades of cheap capital. Around it: Alphabet issued $31.5B including a 100-year bond — plus an $80-85B equity raise, the largest single equity raise in market history (Berkshire took $10B). Amazon $54B in March and more since; Meta $30B then $25B. Six tech names have issued roughly $244B through mid-July. Morgan Stanley projects ~$570B of AI-linked debt issuance in 2026, double 2025, against a 2020-24 hyperscaler average of just $28B a year. The bull read: our Goldman-FCF signal showed these buyers had spent all their cash — PIMCO now puts AI capex at ~94% of their operating cash flow — and the answer was not to stop, it was to borrow, and the market took it three times oversubscribed. A 30-year Nvidia bond and a 100-year Alphabet bond are institutional investors underwriting AI infrastructure demand out to 2056 in hard paper. Capex estimates rose through the year — $650B → $725B → $732B for 2026, past $1T for 2027. The demand runway just got a second fuel tank.

The bear read, and it is the best-formed bear mechanism on the thesis after Samsung's share drift. The marginal datacenter is now debt-funded — and memory demand is set by the marginal build, not the average one. That gives our entire demand chain a credit-cycle dependency it did not previously have. Spreads on hyperscaler paper are already widening as supply surges; AI debt is on track to become the largest investment-grade sector, which concentrates the exposure inside index funds and pensions. Meanwhile the US Treasury must refinance roughly $12 trillion over the next twelve months into shrinking structural demand — foreign central banks have been swapping Treasuries for gold, and official gold holdings recently passed official Treasury holdings for the first time in decades. Sovereign supply and AI corporate supply are now competing for the same marginal dollar. This risk does not require anyone to be wrong about AI. It only requires credit to tighten.

Owner challenge, and the correction it forces: "isn't the fix just paying higher interest?" For the giants — yes, and the arithmetic is not close. +300bp on $100B of debt costs about $3B a year, roughly 2% of Microsoft's operating income. A datacenter that Jassy says breaks even inside three years does not stop for 6% money instead of 5%, and Nvidia's deal was 3.4× oversubscribed and upsized — this market is nowhere near closed. So the risk is not a general "credit cycle" and the thesis should not have implied one. It lives in exactly three narrower places: (1) Windows, not rates. Credit markets rarely get expensive; they briefly get shut — 2008, March 2020, September 2022 all saw investment-grade issuance freeze for weeks, and projects slip a quarter. (2) The second tier. Not Microsoft — CoreWeave, Oracle (negative free cash flow and rising leverage), the neoclouds, smaller labs and sovereign projects. They borrow at spread rather than at AAA, and they are a real slice of marginal demand, which is what sets the memory price. (3) Hurdle rates. If the risk-free rate moves 4.5% → 6%, every project is discounted harder: the best datacenter still clears, the worst one gets deferred — and the worst one is the marginal buyer. Net: the risk narrows from "credit cycle" to "credit accident." The 0.3-point tail widening stands, because it prices the freeze scenario rather than a rate move — but the framing is corrected here rather than left to imply that dearer money alone could stop this.

Every load-bearing claim on this site, now said by the man who owns the company. On demand: "Customer demand has doubled this year, and everyone is requesting nearly twice as many chips." On timing: "Next year will be the worst year for memory shortages" — because "expanding production requires long lead times, at least four to five years" and "no memory companies are yet ready to scale up." On the market's character: "It's somewhat like a war — everyone wants more memory chips, and without them, they cannot produce AI computing systems." On price: asked directly about memory up 40-50%, he named it "chip inflation" — "chips are too expensive, forcing even Apple to raise prices… I don't have a short-term solution." That is our Apple decode confirmed by the seller.

The agent multiplier, from the source. "In the past, the memory business depended on population size — each person bought one phone per year. In the AI era, you'll have ten different agents doing tasks for you, and each agent will require more memory." Chain #6 is no longer our inference; it is the chairman's business model. And on the physical wave: "The AI agent, physical AI robot — that needs a lot of memory chips."

The cycle-shape sentence, and why it lands today. "I'm not saying there are no cycles, but the shape of the cycle must change — it will become longer. This is not the memory business we were familiar with in the past; this is a turning point, a structural shift." That is precisely the correction we made yesterday on the owner's audit — flattening 2029 from a dip into a plateau by lifting the multiple from 8.3× to 8.8×. The chairman's own framing arrived one day later. Reinforcing it: "Nvidia wants their own custom chips and Google wants their own customized HBM, so it's not just a commodity — it actually changes the memory chip's status" — the same commodity-premise death Intel's CEO announced this week, now from the seller's side.

The capex number, correctly read. $720B to triple capacity by 2034 — the largest buildout in memory history — works out to roughly 14.7% annual bit growth, below the ~20% the industry has historically added and far below AI demand growth. Chey says it plainly: even doubling capacity within five years "isn't enough for most customers." The biggest capital programme ever announced in this industry still does not out-run its own demand. Alongside: 10 long-term agreements signed, LTAs renewed every year (an annual repricing mechanism), fab JVs under consideration, and a Memory-as-a-Service model under review — recurring revenue, which is the structural bridge to the utility-style multiple our renormalization argument predicts. US expansion is live but deliberate: a $3.9B advanced packaging plant in Indiana under construction, US memory-fab sites under evaluation, with the honest obstacle named — relocating 600-700 supplier companies. And the artifact of the year: Chey showed CNBC a wafer inscribed by Jensen Huang — "Please make more."

Two facts to keep beside the enthusiasm. Counterpoint puts SK at 58% of HBM in Q1, with Samsung and Micron both at 21% — consistent with share drift, and with our re-armed T3 tripwire. And the supply response is real: Micron is spending $50B on two Boise fabs (first wafers 2027) and $100B on a Clay, NY campus of up to four fabs — that is row 7's 2029-31 wave, funded and named.

Part 1 — the tools are being pushed back, not pulled forward. Samsung wanted High-NA EUV in volume at 2nm and 1.4nm and has decided the technology "is not yet mature enough", deferring it to A10 and beyond (~2030) despite two installed systems and over ₩1T spent. TSMC said the same in April: no current plans, and its A13/A12 nodes in 2029 will not need it. That leaves Intel as the only company running High-NA in production (selected layers of Panther Lake on 18A). Both leaders will instead squeeze more from existing 0.33-NA tools using multi-patterning — more exposure passes per wafer, so effective capacity grows slower than the tool count implies. Every roadmap for new supply just got harder, and ASML's own €60B-by-2030 target faces headwinds. The bottleneck is not being relieved; it is being deferred.

Part 2 — the arithmetic that dates row 7. A leading-edge fab at 100K wafer starts/month needs roughly 20 EUV scanners. Terafab at a billion chips a year needs 50-100; fully ramped, published estimates run to 400. ASML shipped 48 EUV systems in all of 2025, targets 60 in 2026 and 80 in 2027 — for the entire planet. Even at the 2027 rate, 400 machines is five years of the world's total output going to one customer. This is the hardest number yet behind our 2029-31 supply window: not an opinion about competition, but a machine-count constraint on a single-supplier tool with a multi-year lead time. New entrants cannot arrive early because the instruments to build them do not exist yet.

Part 3 — what Terafab actually is, and the memory connection. Announced March 22, 2026 in Austin: a joint venture of SpaceX, Tesla and Intel (Intel joined April) near Giga Texas, building on Intel's 14A process, ~100M sq ft, targeting one terawatt of compute hardware a year. Investment figures reported inconsistently — $16.8B initial commitment in some accounts, $55B in others, up to $119-122B if fully built — the range is the honest state of the record. Its S-1 describes a closed loop spanning mask design, logic and memory fabrication, and advanced packaging. Musk did a fireside with ASML CEO Fouquet on June 9-10; Fouquet had already called the project a "serious endeavor" and ASML says it will collaborate. No public EUV order or delivery schedule for Terafab has been reported — which is why ASML's order book stays on the thesis as the public tripwire.

Part 4 — the FEL wildcard, correctly placed. In August Musk endorsed free-electron-laser lithography ("FEL FTW") — a fundamentally different EUV light source using a particle accelerator instead of ASML's tin-droplet plasma: higher power, no tin debris, tunable wavelength, potentially lower running cost. Crucially it targets the light source, not the scanner — a plug-in upgrade path rather than a rival machine. xLight, backed by former Intel CEO Pat Gelsinger, is pursuing the same route for 2nm and beyond. Nothing is in production; no timeline is confirmed. This is the single most credible way our 2029-31 supply window could arrive early, and it is now armed as Watch W3: a FEL production milestone. Until one exists, the machine-count arithmetic above governs. The conclusion this does NOT license. The machine-count wall blocks newcomers; it does not cancel the 2029-31 wave, which never depended on newcomers. That wave comes from incumbents already inside ASML's queue (Yongin from Feb-2027, M15X, M17, Samsung P4/P5, Micron, plus Korea's $585B programme), from node shrinks that add 15-25% more bits per wafer with no new cleanroom — historically the main cause of gluts — from wafer re-allocation (HBM consumes ~3× the wafers of DDR5, so any slowdown in HBM growth floods conventional DRAM within months), and from depreciation climbing from ~₩31T in 2026 toward ~₩120T by 2030, which compresses margins even at flat prices. Row 7 stands. Footnote on the Intel diaspora: Gelsinger backs the lithography bypass, Tan hired SK hynix's former CEO for packaging, and Musk builds on Intel's node — the same alumni network is attacking both bottlenecks at once. Worth watching; not yet worth pricing.

Why the quote outranks the news. This site's entire anomaly is one belief: memory is a commodity, so price it at a commodity multiple — which is how a company growing profits 70% trades at PEG 0.05 against an industry at 1.08. Tan is not a commentator; he ran Cadence, sits at the centre of the industry, and spent a career telling people not to invest in memory. His public reversal — "now it has become different… there is a lot of new technology coming out" — is the belief itself dying at the highest level. Multiples move when the category changes in people's heads, and that reclassification is now happening in public. It is the same mechanism our renormalization argument predicts: contracted, capex-heavy, oligopoly supply gets re-rated toward utility-grade comparables, not spot-commodity ones.

Three corroborations riding along: Tan separately says there will be "no relief" on memory supply until at least 2028 — the third CEO on the thesis to name that window, after SK's Kwak (worst-ever 2027, demand > capacity past 2030), Micron's Mehrotra (no line of sight) and OVH's Klaba (squeeze until 2028). Reporting alongside notes future supply is largely sold out for two years — the LTA lockup, confirmed from the buyer side. And Intel's own $20B equity raise — its first since 1971 — is earmarked for foundry manufacturing, not memory: no new memory wafers are being funded here.

The bear half, stated plainly. Tan hired Seok-Hee Lee, SK hynix's former CEO, as EVP of Intel Foundry — running advanced packaging, system integration and back-end manufacturing, which is precisely where SK's process moat lives. That is the most senior talent transfer to a potential rival yet, and it follows the thesis's IP-leak entry: the knowledge is the asset, and it walks. Technically, Intel's XBM patent eliminates the silicon interposer that makes HBM expensive, replacing a 1,024-bit parallel bus with serial UCIe links and back-end-of-line DRAM cells; a second project, ZAM, is co-developed with SoftBank's SAIMEMORY. If HBM's packaging difficulty is the moat, these attack the moat directly.

What the dates do to that threat. XBM is targeted post-2030; ZAM 2029 — both land beyond this thesis's window, and both are patents rather than products. TrendForce's own read: SK and Samsung have spent years cutting interposer cost via chiplets, UCIe and fan-out packaging themselves, while platform compatibility and the Nvidia-optimised software ecosystem are major adoption barriers — the same barriers that killed Optane. And the base rate deserves its say: Intel was founded as a memory company in 1968, exited in the 1980s, then abandoned both NAND and Optane. Net: near-term, a rival conceding our thesis in public; long-term, a third named entrant on the 2029-31 supply row beside Terafab, CXMT and Korea's $1T programme — logged there, not here. Watch W2 armed: an actual Intel memory capex commitment (a fab, a number, a date) — patents are intent, capital is threat.

What the reporting says: Samsung's HBM4 yield has climbed to ~80%, up from under 60% when mass production began in February — beating its own end-2026 target months early, on progress in the TC-NCF bonding process (its historic weak point) and 1c-nm DRAM yields already above 80%. Samsung is targeting ~38% HBM share by year-end. And the line that matters most to us: TrendForce's June forecast raised Samsung's 2026 HBM4 share sharply while trimming SK hynix's, citing "qualification delays and capacity shifts." Separately, SK hynix labour negotiations add an execution variable. The same report notes general-purpose DRAM price growth is expected to moderate.

How this sits against the 70% Nvidia lock. Both can be true and probably are: SK holding ~70% of Nvidia's Rubin HBM4 allocation (six outlets, UBS concurring) while its overall HBM4 share drifts toward 60% as Samsung qualifies elsewhere — Nvidia has said openly it prefers multi-sourcing. But the direction is unambiguous and it runs against us: a second qualified supplier is the mechanism by which pricing power decays, and Samsung reaching 38% is the fastest plausible path to it. This is the single most credible bear development since the ₩2.8/Gb rumour, and unlike that rumour it is sourced to yield data rather than chat.

What changes, concretely: tripwire T3 is re-armed tighter — TrendForce Q1-2027 DRAM revenue share below 34% trims FY27 to ₩620T (end-27 → ~₩5.7M), and the conservation audit is corrected upward in honesty: FY27's ₩633T requires SK at roughly 40% of a ~$1.2T market, not the 36% first calculated — a more demanding assumption than stated yesterday, and now stated correctly. Added to the watch list: Samsung's year-end share print (38% target hit or missed) and the labour settlement. What does not change: contract floors, the prepaid Nvidia structure through 2030, and the physical fact that neither supplier can add wafers before 2029.

The invoice, not the forecast. OVHcloud — Europe's largest independent cloud, a buyer with no long-term agreements who "must order monthly without a guaranteed price" — published what it actually pays: RAM up 6× in the year to June 2026, heading to 9× by September and 12× next year. Alongside: NVMe +7×, hard drives +3.5×, CPUs/boards/NICs +15-20%. The pass-through: new-generation server rentals +51% on average, +87% for the 2026 gaming tier, +49% Advance, +40% Scale, +59% High Grade, +28% on 2024-era kit — effective Sept 1 for new orders, Oct 1 on renewals. Klaba expects the squeeze to persist until 2028.

Why this is the most important price datum on the site. Apple's repricing let us infer memory up 125-250%. OVH doesn't require inference — it prints the purchase price. And the trajectory of OVH's own guidance is the real story: in March they estimated 5-10% and capped increases at 9-11%; five months later it is 87%. The flood outran a professional buyer's own forecast by roughly tenfold in five months — which is exactly what our "stop being cautious" recalibration was built to respect. Independent confirmation in the same reporting: TrendForce has DDR4 +158% and DDR5 modules +307% since September 2025; Samsung has already lifted some memory prices ~60%. The Register's headline word for it — "RAMpocalypse" — is the same term BlackRock's Tony Kim coined; the vocabulary of the trade has converged on our thesis.

What it does and does not mean for SK's revenue. It does not mean SK's realized prices rise 12× — OVH pays the marginal spot price as a small, uncontracted buyer, while SK's blend is contract-weighted with floors and ceilings. What it does mean, in order of force: (1) the spot-to-contract gap is enormous, and contracts reset toward spot on renewal — this is the pipeline that fills 2027; (2) conventional DRAM, the lane our model treats most cautiously, is the one exploding, because HBM is eating the wafers that used to make it; (3) hyperscalers are crowding out everyone else — enterprises can't buy hardware, so they rent, which pushes yet more demand into the datacenters that buy SK's memory; (4) Jassy's framing that AWS could become a trillion-dollar-revenue business with servers breaking even in under three years is the payback math that keeps the buildout funded. The honest counter-entry: this is also the first hard evidence for our consumer-lane step 4 — demand destruction at the margin. OVH customers are publicly threatening to leave; 87% is the price level where small buyers exit. Logged on the bear row, watched, not dismissed.

Handling first: this specific claim could not be verified — no indexed post was found, so it is logged as an unverified secondhand claim and scored on substance only, not treated as evidence. The substance splits cleanly in two. On positioning, the call has real force: SanDisk is up 505% YTD, Micron 223%, memory is now consensus, and crowded consensus trades mean-revert violently — which is precisely what the last three weeks have looked like. On fundamentals, the tape says the opposite: Micron's Q3 FY2026 printed revenue $41.5B, +346% YoY, 85% GAAP gross margin, guiding ~$50B. A finished trade does not look like that. The honest synthesis: the easy money in multiple expansion is probably made; the earnings growth is not. That distinction is the entire reason this site prices SK on earnings and treats multiples as the biggest single audit risk. On the "long cement" idea: the underlying observation is sound — datacentre shells, substations and power plants consume enormous concrete, and 22GW a year of new capacity is a construction story before it is a chip story. But three problems for us, stated plainly. (1) It is the same bet, not a diversifying one — cement demand here is downstream of the identical AI-capex driver, so it correlates with what we already own (the Nittobo lesson). (2) It fails the PEGY screen: cement and aggregates are mature, regionally priced, low-growth businesses trading at mid-to-high multiples, against SK at ~5-6× forward — the anomaly gap runs the wrong way. (3) Cement barely trades as a contract — it is heavy, cheap per tonne and uneconomic to ship far, so there is no liquid forward market; "long cement" in practice means owning regional producers or negotiating physical offtake, neither of which is available as a clean position. And the robot-construction leg is a 2030s thesis, priced today at 2020s multiples — the opposite of the setup we look for. Verdict: interesting observation, not a better use of capital than what it would replace.

What they say, ranked by evidential weight rather than volume.

1. Kwak Noh-jung (CEO) — 9/10. 2027 will be "the worst year in the industry's history from the supply perspective"; demand above SK's capacity even beyond 2030. Why it scores highest: it is against interest in the near term — telling customers the shortage is permanent invites exactly the substitution, qualification of rivals and vertical integration that hurts him. And it is falsifiable on a date. Corroborated by his own actions: capex raised to high-₩40T, M15X pulled forward, P&T7 accelerated with ₩7.09T six days before the print.

2. Chey Tae-won (Chairman) — 7/10. Wafer shortage to ~2030, deficit over 20%, "at least four to five years" to add meaningful wafer capacity. The number is the value: it converts a mood into arithmetic that Meritz independently reproduced (75-80% of demand met now, ~60% by 2027). Discount applied for the same conflict as Kwak, plus the ₩944B divorce ruling creating personal liquidity pressure at the holding-company level — logged in bear row 13, not ignored.

3. Song Hyun-jong (President), on the call — 8/10. The quietest and most useful: customers keep asking for more, and SK is signing further LTAs to reduce exposure to price swings. That last clause is a seller voluntarily capping his own upside in a shortage. Nobody does that unless they believe the current spot regime is temporary in one direction or the other — it is the clearest possible signal that management is trading peak price for durability. It is also, precisely, the de-cyclicalisation the multiple argument depends on.

4. Elon Musk — 8/10 on the observation, 4/10 as forecast. Calls the memory spike "the biggest price jump in anything I've ever seen," echoes Tim Cook's "hundred-year flood," says the US has not a single high-volume memory fab and new domestic plants will not reach volume until ~2028-2030, and that for Tesla "memory is an even bigger limiter than AI logic." Highest weight of any outside voice because he is a buyer talking his own book against himself — a customer publicly announcing he cannot get supply strengthens his suppliers' pricing power. The Tera-Fab idea and his public thanks to Micron for "making room for Tesla in the years to come" are the same signal in two forms: the scarcity is real enough to justify vertical integration. Upgraded Aug 7: now with SpaceX Nvidia-exclusive at 10-GW scale and his sharpest formulation yet — demand +200%/yr against supply +20%/yr: Economics 101 would suggest that the price increases. It does not decrease. And the arc completed Aug 7: Terafab posted its first DRAM engineering role — the buyer became a pleader became a builder, the strongest duration testimony an insider can give.

5. The investor bloc (Gerstner, Aschenbrenner, Baker) — 6/10, and honestly labelled. They are positioned, which makes them the most conflicted narrators in the set: Gerstner sold Microsoft to fund NVIDIA and SK; Aschenbrenner had SK at ~6.5% of book pre-ADR then cornerstoned the deal; Baker pitched it at Sohn and added in his 13F. Their value is not their opinions but their revealed cost of conviction — and none of them can show what they did in the crash until the Nov 14 filings.

The absence worth naming (E8 salience check). Nvidia never says this out loud. Its position is expressed only structurally — HBM stacks at the centre of every roadmap, memory named first on Morgan Stanley's constraint list after meeting Huang, and a co-development LTA with SK. The customer with the most to lose from high memory prices behaves as though they will stay high. Behaviour outranks all five quotes above.

The disconfirming voice, kept in view: SK and Samsung both warn PC and phone customers cannot secure DRAM and are cutting orders or downgrading specs. That is demand destruction at the bottom of the market — real, and the leading edge of the price ceiling the bear case needs.

The initiation wave, tabulated: Cantor $300 ("bit demand for both DRAM and NAND well in excess of supply into CY29, if not longer") · Barclays $330 ("materially higher 2027 revenues driven by HBM pricing uplift"; cash >40% of market cap by end-2027 funding buybacks) · Rosenblatt $320 (HBM dominance + "a structural change in the memory industry's ability to boost capacity" — scarcity as a feature) · BofA $250 · Stifel $240 · RBC $200. BofA's five reasons, verbatim structure: (1) strong Big-Tech chip orders long-term; (2) stable ASPs through 2027-28 with "no hard landing" — the contract-floors thesis in a bank's words; (3) super-cycle operating profit at an annualized run-rate above ₩300T from Q3 2026 — sitting at-to-above our flood FY26, conservatism receipt #5; (4) leadership across HBM, LPDDR5, eSSD; (5) a P/E of just 4× on 2027-28 earnings — our anomaly, now a bank's headline stat. The ADR premium (20% to local) is explicitly blessed via the Micron comp. The full-street map, refreshed: 39 analysts, 38 Buy / 1 Hold / 0 Sell, average PT ₩3.16M (+122%), low ₩1.2M, and a new street maximum of ₩5.3M — above Korea Investment's ₩4.7M and above our 12-month band's top: the D-branch of our fan now has professional company. Prediction P6 recalibrated honestly: with the full 39-desk set visible, dispersion baseline restates from 3.2× (10 desks) to 4.4× (max/min, Aug 12); target restated: <2.2× by Nov 15, condensing upward — and the US cluster at $240-330 is the condensation already beginning on the dollar side. Nancy's two Simply Wall St items, decoded: (a) the ₩5T (~$3.5B) semiconductor ecosystem fund targets the materials, parts, equipment and fabless firms around SK — not SK's P&L directly, but a state-funded de-bottlenecking of the supplier base beneath the fab ramps (M15X, Yongin) and another brick in the government's long-life manufacturing-hub commitment: ramp-execution risk drifts down. (b) the technology-leak sentencing cuts both ways and both are logged: it proves the IP is worth stealing and that Korea now jails people for stealing it — deterrence hardening the process moat — while adding a visible internal-controls workstream to the execution-risk row. Signposts adopted from the article: megaproject allocation tranches naming SK hubs; any further leak cases.

The seven-point argument, from the primary reports: (1) The Q2 "miss" decoded exactly as we did: operating profit ₩60.54T ran 6.4% under the ₩64.7T consensus because Q2 DRAM price capture lagged expectations — but per the company, this was shipment-timing deferral of high-value products into H2, not demand slowdown: the revenue shifted right, not away, and becomes a Q3 improvement factor. Meanwhile the 76% operating margin — a record — proved the fundamentals (their words: "solid fundamentals proven by results"). (2) Q3 DRAM ASP +20% QoQ — they doubled their own prior +10% estimate. This is now the second Korean desk independently printing ~+20% beside Hyundai Motor Securities' +19.9%: two brokers, two methods, one number — our Rung-1 input is a street consensus forming. (3) Bit growth upgraded to mid-20%s for 2026 (from 20%) on 1c-nm SOCAMM2 shipment expansion — a new server-LPDDR product lane entering our books today. (4) Earnings upgraded: 2026 OP ₩270T (+10%) · 2027 OP ₩418T (+11%). (5) The anchor: ASPs held high on up-to-five-year LTAs, with HBM4 mass-production shipments beginning in earnest from Q3 stacking HBM price effects on top. (6) Their diagnosis of the correction mirrors ours: the fall reflects "market skepticism about AI investment durability, not fundamental change" — and they expect the shortage to deepen through H2 and into next year. (7) Their closing line could hang on our wall: "focus on the company's value and the direction of earnings, not short-term price moves."

The convergence check against our model: KI's 2026 OP ₩270T sits just under our flood chain's ~₩289T-equivalent; their 2027 ₩418T sits below our ~₩480T — they are mid-migration toward the flood, one +20% quarter at a time. Their ₩4.7M target prices ~11.5× their own 2027 earnings — Micron-parity-plus, the multiple journey we model, already in use at one Korean desk. And the marker that matters: the street's maximum target has now entered the top of our flood-weighted band (now ₩4.7-5.6M after v9.2) — the first professional number to arrive where our evidence already lives. Calendar hardening from the same file: the 2027 HBM contract pricing finalizes end-September (the settlement binary now has a date, a month before the Q3 print), and the expanded shareholder-return framework lands end-Q3 per both JPM and KI — the returns catalyst just moved forward from our December assumption. Monitor hygiene, logged: the 7am job dated this raise "~Aug 11" off X amplification; the filing date was July 30 — single-source social dating gets a verification pass before it moves anything, and today it earned one.

The content ladder, on record: an ordinary car carries ~16GB of memory; an L4 autonomous vehicle 300GB+; and Micron's CEO put the next rung in writing: "Humanoid robots carry 10 times the memory of an average L2+ vehicle… We expect a sustained, substantial multi-decade memory demand cycle to begin in the latter part of this decade" — adding "we currently do not have line of sight as to when memory supply will be able to catch up." The robot brain already exists as a product: Nvidia's Jetson Thor ships with 128GB of LPDDR5X at 273 GB/s — a memory board that walks. The wave is measured, not predicted: humanoid shipments hit 19,100 units in H1 2026, +272% year over year; compounding that rate points at ~180K units in 2027 and ~650K+ in 2028, with Optimus Gen-3 production-intent, Fremont pilot lines running this year, and a 1M-unit/yr target at $20-30K. And the bigger lever arrives first: fleet training — world models, video-heavy synthetic data, continuous retraining — runs in datacenters and scales with fleet size, not chip-count per robot: HBM demand landing 2027-28, before most robots even ship. Ceiling math for scale: 100M robots × 300GB = 30 exabytes ≈ 75% of all 2026 global DRAM output. Korea's own $1T megaproject (SK+Samsung $585B of new fabs, humanoid deployment by 2028, DRAM output doubling in five years) is the state-level confirmation of both the demand and — honestly noted — the 2029-30 supply response row 7 already watches.

The correction this forces: the year-path's end-2028 assumed the multiple compresses to 9.6× as markets pre-price the factory wave — an old-cycle reflex written before the floors printed, before Goldman's record-FCF 2029, and before robotics entered the file. Four reasons it no longer holds: contractual floors with 60-70% LTA coverage blunt the fear mechanism · buyers arrive at the wave flush (Goldman's chart) · a new multi-decade demand wave begins precisely in the window (Micron CEO's words) while Korea's new fabs produce nothing before 2029-30 · HBM4E/HBM5 qualification moats hold the premium mix. Adopted: 2029E EPS ₩560K→₩600K (upsized again to ₩630K in v9.2 as the AV + fleet-training lanes were added), end-28 multiple 9.6→10.2 — end-2028 rises to ~₩6.1M, above end-2027. The 2029 wave-year breathing stays, honestly, at ~₩5.4M with the reposition plan intact — history gets its respect exactly one year, not two.

The event: for the first time, non-human traffic exceeded human traffic on Cloudflare's network — the company that fronts roughly a fifth of the web. Prince's projection: machine traffic 1,000× human within five years. Why this is a memory number wearing a networking costume: a human browsing a page consumes essentially zero AI compute; an agent visiting a page is AI compute — the page is crawled, tokenized, stuffed into a context window, reasoned over, and answered through HBM on every step. The crawl ratios quantify the intensity: Google once fetched 2 pages per human visitor delivered; last year 15; AI labs now fetch 250-6,000 pages per visitor — machines reading at hundreds to thousands of times human intensity, all of it inference load. The transaction ceiling, cross-checked against our ledger: Prince's 10-100 million agentic microtransactions per second (500-5,000× Visa's peak) would, at even ~1,000 tokens of inference each, generate 0.9-9 quadrillion tokens a day — 2-23× today's entire verified token economy from the monetized slice alone — and Prince expects that slice to be only 1-10% of total agent traffic, which is how his 1,000× reconciles. Our Chain #6 assumes HBM serving bits merely ×2.4 by 2028; the ladder of independent operator calls above it now reads Goldman ×24, Google ×330, Cloudflare ×1,000 — every one of them dwarfs our input. Two second-order effects, both memory-positive: (1) if agents must pay per page, operators will cache aggressively — prefix caches, agent memory stores, retrieved-context hoards — turning an access fee into a memory-capacity arms race (the MinIO/Neo4j agent-memory products are this market forming); (2) the machine-payment rails themselves (fraud scoring, sub-cent settlement, stablecoin ledgers at Nasdaq-multiples of throughput) are in-memory workloads — a DRAM/CXL lane, not just HBM. The economic loop closes too: per-request pricing gives tokens direct revenue attach — the monetization floor beneath the $100B/GW frame and Goldman's FCF harvest. The internet's business model is repricing from human attention to machine cognition, and cognition is the thing that lives in memory.

The allocation, now multi-sourced: Nvidia has assigned SK ~two-thirds to 70% of its entire HBM4 demand for the Vera Rubin platform — "well above" the ~50% the industry expected — because SK's paid samples "cleared final validation without issues" and its mass-production system has run since September 2025. Samsung, having passed qualification only in January, fights for the remaining ~30%; Micron runs third. Our Nvidia pass-through lane was built on a 56% blended share — the growth segment's verified share is 70%, a ~25% headroom note on the ₩44T lane, logged without double-counting the flood chain. The structure behind it: the June 2026 co-development pact runs through 2030 and includes advance payments from Nvidia to SK — the buyer of the decade prepaying its supplier, joining the industry's $38B prepay ledger by name. Jensen's own framing: SK "will continue to be its largest partner," with co-development extending into memory manufacturing technology itself (blocksandfiles) — the customer helping sharpen the supplier's tools, a moat Samsung cannot copy from outside the room. And the base logic die rides TSMC 12nm/5nm, an integration barrier that rewards the first mover. The de-risking item nobody headlined: the Wuxi, China fab — 30-40% of SK's global DRAM output — completed its 1z→1a nanometre process upgrade across ~90% of its 180-190K monthly wafers, which the trade press says "eliminated market concerns about potential supply disruptions" under export controls. A background tail-risk on a third of DRAM capacity, quietly closed. The company's own near-term words: SK expects the HBM4 shortage to worsen in H2 2026 as demand surges. The honest caveat, kept verbatim in spirit: the $500B figure is the program — Nvidia's AI-factory buildout frame — not a revenue order to SK; the benefit still routes through GPU shipments, yields, capacity and competition. What the lock does guarantee is position: first customer, largest share, prepaid, co-developing, through 2030. Footnote for the anomaly file: the ADR that fixed the access problem was the second-largest US listing in history, behind only SpaceX — and SK now comprises over a quarter of Korea's entire stock market.

What the chart shows, read plainly: the five hyperscalers' combined quarterly free cash flow climbed for fifteen years to a peak near $120-130B a quarter (2023-24) — then collapsed through 2025 into 2026 to roughly zero, with several members outright negative — and the red circle marked "You are here" sits precisely at that trough, 2026. Then Goldman's own estimate bars turn: recovery through 2027, strength in 2028, and by 2029E the tallest bars on the entire chart — new record FCF above the old peak.

Three readings for the thesis. (1) The revealed preference — the strongest "spending doesn't stop" evidence that can exist: these companies could defend their cash flow tomorrow by trimming capex; instead they chose to spend every dollar of operating cash — hundreds of billions a year — on AI infrastructure, driving FCF to zero in public view. Zuckerberg's essay (integrated this morning) is the ideology; this chart is the accounting of that ideology, across all five names at once. Memory sits inside every one of those capex dollars. (2) The V is the monetization arriving: Goldman's 2027-29E recovery is only possible if the AI revenue harvest shows up in the cash line — token revenue, the $100B/GW-class economics, the inference paybacks. The best macro desk on earth is modeling our Chain #6 as their customers' cash flow. (3) The 2029 shape is a buyer-side mitigant for bear row 7: the supply wave lands 2029-31 into customers whose free cash flow Goldman models at record highs — buyers flush with cash absorb new capacity at firm prices; buyers in a cash crunch would force the crash. This is the third independent softener on the wave (after the contract floors and the CEO's beyond-2030 calendar), now logged on row 7. The honest symmetric note, kept: the trough itself is why markets are frightened right now — "you are here" is the moment of maximum capex anxiety, and the reassurance comes from estimate bars, not printed ones. The falsifier is clean and dated: if hyperscaler FCF is NOT visibly recovering by mid-2027, the harvest thesis weakens and row 7 hardens. Until then: five companies just proved, in audited cash, that the buildout outranks their own free cash flow.

Seven extractions that matter for the thesis:

(1) The product IS persistent memory. Meta's stated goal: "Everyone will have an exceptionally capable personal agent that understands you, your goals, and everything you care about. Your agent will work 24/7 on your behalf." For billions of people. A permanent agent that knows your life is, physically, a permanent growing context — per-user memory state held hot, forever. This is Chain #6's agent multiplier and Kim's "models remember your behavior, your context" written as the roadmap of a three-billion-user company. (2) Scarcity doctrine, verbatim: "No matter how intelligent AI becomes, there will always be a finite amount of compute and therefore an opportunity cost for how we use it." And the pricing mechanism he announces: a "dynamic auction" for compute access. When the CEO of Meta says intelligence will be auctioned, he is saying compute stays scarce and price-discovered at marginal value — the exact economics behind the $100B/GW revenue frame. (3) The RSI floor: "Any lab that doesn't let their AI system direct substantial compute toward recursive self-improvement will inherently fall behind" — so labs must build for the self-improvement race and billions of users simultaneously. A structural, open-ended demand floor under capex. (4) The anti-pause clause: "AI is likely the most competitive industry in history… maintaining even a two-month advantage is incredibly valuable." Nobody can stop spending without ceding the race — the "does the spending stop" question answered by ideology, not just budgets. (5) The build-faster machinery: China cited as adding "1GW+ of nuclear every other week"; Community Compacts ($50K teacher bonuses in Richland Parish) engineered to make US datacenter construction politically frictionless. Meta is paving the road for more gigawatts, not fewer. (6) Open source resumes: "We will resume releasing some open source models soon" — open weights push inference into every device and server on earth, the long-tail memory demand Baker's device thesis describes. (7) The honest counter-current, noted: he muses an RSI system might "squeeze 100× more intelligence out of each gigawatt" — the efficiency risk acknowledged — and his own conclusion is still build maximum compute anyway, because efficiency gains get spent on more intelligence (Jevons, from the man writing the checks). Strategic read: this essay is a commitment device. Meta's ~$100B+/yr capex is no longer a budget line that can quietly shrink — it is now the physical expression of a published philosophy. Ideologically locked spending is the most durable kind.

The doctrine reversal, in audited numbers: Tim Cook built the most admired supply chain on earth on one rule — inventory is "fundamentally evil." Apple's latest balance sheet: inventory $11.09 billion, up 94% in nine months (from $5.72B), consuming $5.46 billion of cash to build. On his final earnings call before handing the company to John Ternus, Cook chose to spend it warning about one thing: "a hundred-year flood on the memory pricing" — something he has "never seen anything like" in 40+ years in electronics — with "very significant constraints currently and limited flexibility in the supply chain to remedy it," impacts expected to "increase significantly sequentially." Memory costs ate more than Apple's entire sequential gross-margin decline. Apple took its own store offline and repriced: MacBook Air +$200, MacBook Pro +$300, some products up to +25%. And the CFO's detail that matters most for pricing persistence: the inventory hedge's benefit shrinks after September — the world's most powerful buyer returns to the market as a price-taker within a quarter.

The buyer's arc is now complete across every megacap, each in its native language: Musk builds a fab · AMD buys a chip company to dodge the queue · hyperscalers prepay $38B (with $50-100B more offered) · Apple hoards $11 billion of parts against its own religion. Nobody positions like this for a shortage they expect to end. Duration corroboration keeps stacking: mainstream coverage now runs the crisis to at least 2028 with prices +30-45%, ABC (citing Kim) sizes the memory market exploding from ~$220B (2025) toward ~$890B this year (labeled as reported), and notes 90%+ of production sits with three companies. Tom's Hardware's own juxtaposition says the quiet part: "On the other side of the trade, SK hynix ran a 76% operating margin over the same three months… and guided Q3 DRAM bit shipments up only about 10% sequentially." Apple's pain is the mirror image of SK's pricing power — same transaction, two ledgers. Consumer-lane escalation tracker advances 2/4 → 3/4 (price hikes ✓, spec-skipping ✓, now megacap hoarding + margin impact ✓); step 4 — actual unit demand destruction — stays the watch, and even that mechanism reallocates bits toward AI at higher prices rather than costing SK revenue.

What the headlines missed at FMS 2026 — four items now on file:

(1) High Bandwidth Flash went from concept to published standard in six months. SK and SanDisk unveiled the first HBF standard specification: a new memory tier between HBM and SSDs — NAND stacked with TSVs like HBM — delivering up to 512 GB per device (8- and 16-high stacks) at 0.4 to 3.0 TB/s across three bandwidth grades, connecting to GPUs and CPUs through the open UCIe chiplet interface, and submitted through the Open Compute Project as an open industry standard. Read what that structure means: SK co-owns the rulebook of a brand-new product category the way JEDEC incumbents own DRAM — a fourth revenue lane (HBM · DRAM · eSSD · now HBF) being born with SK holding the pen. (2) Google joined the consortium — and Google DeepMind sat on the panel. Alongside Tenstorrent, the consortium's newest members put an anchor hyperscaler inside the standards body before the first product ships; DeepMind's senior staff engineer joined SK and SanDisk on stage for "Breaking the Memory Wall with HBF." Demand-side validation doesn't come more direct. (3) 16-high HBM4 was physically displayed (StorageReview) — the capacity-per-stack doubling that powers Citi's 434% content path, shown as hardware, not roadmap. (4) V10 375-layer wafer-bonded NAND debuted at 2.5× power efficiency, with eSSD mass production planned early 2027 — the AI-NAND leg moving to product. The keynote line that frames it all (EVP Kim Chun-sung): "What determines overall efficiency is where each memory tier — HBM, DRAM, NAND, and SSD — is positioned and how it is connected to minimize data movement." The tiered-memory story isn't a defensive hedge against HBM's importance — it's SK annexing every layer of the hierarchy while owning the top of it. Flag resolved: EETimes' "$86B revenue this year" reads as the HBM-lane revenue figure (~₩120T), consistent with this site's lane math — not a total-revenue claim.

Baker's four-reason update (All-In): the crunch persists because (1) consumer hardware — even iPhones — lacks the memory to run AI locally, forcing a device-wide DRAM upgrade cycle on top of the datacenter one; (2) TSMC is keeping capacity disciplined, killing the classic oversupply route; his 13F receipts: SK Hynix, Micron, SanDisk added late-2025 — each up triple digits since. The techflowpost detail worth framing: Google and Microsoft have reportedly offered $50-100 BILLION to prepay three years of SK's production, and SK's negotiating posture is summarized as "I won't guarantee supply — I'll just raise prices" at ~70% operating margins, unprecedented in semiconductors. The supply response is physical and dated: SK America CEO Ryu Sungsoo told Reuters the first Yongin fab is being pulled forward three months to Feb-2027 and M15X takes first wafers within weeks — "we have to support memory consumption for AI infrastructure." And the rival testifies: Samsung memory chief Kim Jaejune warned of "significant shortages through at least 2027" with fulfilment at record lows; analyst Bajarin projects HBM as the defining constraint of the decade, growing 4× to $100B+. Every leg — the buyer desperation, the seller posture, the physical fab schedule, the rival's confession — now carries a named principal and a date.

Who this is: the man who runs technology stock-picking at the largest asset manager on earth, across public and private markets, semis to quantum. What he said maps onto this site's architecture almost paragraph for paragraph.

(1) The inversion, in his numbers: chips and hardware now command ~$30 trillion of market value — triple the $10T software/services layer — because AI turned a $10,000 server into a million-dollar asset. His words: "AI happens. It's like BCE to AD — bam, 2023 happens. Everything changed." (2) The capex ceiling rises again: ~$1 trillion this year, $10 trillion over the next five — a number ABOVE Goldman's $7.6T arc, from the desk that allocates more capital than anyone alive. (3) The memory doctrine, verbatim: "Memory intensity has just skyrocketed. As you see models remember more — your behavior, your context — and as agents proliferate, memory becomes paramount. The human brain is memory-intensive, but today's AI is compute-intensive. Going forward, memory in concert with compute will become even more important." That is this site's Layer-1 thesis — context is the growing part, intelligence converges toward the brain's memory-heavy architecture — spoken by BlackRock, under his own coinage: "the RAMpocalypse." His summary image: if compute is the engine, memory is the fuel tank — and only three companies on earth can fill it. (4) Duration, independently instrumented a third time: he cites a 3-4 year mismatch between fab buildouts and today's demand — BlackRock's version of the CEO's beyond-2030 calendar and Goldman's capex arc. (5) De-commoditization, institutionalized: his segment titled "The 20-year lie: compute is just a commodity" is the multiple-re-rating argument — the JPM "raw-material supplier → long-term partner" migration — delivered as investment doctrine. And his three-horizon playbook puts memory first in the "NOW" bucket, ahead of everything else he covers.

The detail to frame: recounting the July week when IBM had the worst week of its 113-year life (−26%) as client spend fled software for physical infrastructure, Kim's chosen counter-exhibit was SK Hynix rising 27% in a single day. When the largest allocator on earth explains the new world order, your company is the picture he holds up. The discovery ladder is now complete: retail headlines (Motley Fool) → sell-side language (JPM, Citi) → the top of the institutional food chain naming the thesis out loud. Undiscovery — the anomaly-gap's engine — is ending on schedule.

THE RECEIPTS (added Aug 10, from BlackRock's own factsheets): talk is one thing — position sizing is the buy-side's real price target, and here is Kim's. BGF World Technology Fund (the flagship, holdings 31-Jan-2026): SK HYNIX IS THE #1 POSITION AT 7.64% — ABOVE NVIDIA (7.60%), above Broadcom (7.57%), above Apple. Samsung sits #5 at 5.32% — the two Korean memory names alone are ~13% of BlackRock's flagship tech fund, and with Lam Research (memory-capex equipment, 4.38%) the memory ecosystem approaches a fifth of the book. The US Technology Opportunities Fund: SK Hynix top-5 at 4.78% — a Korean listing earning a top-five slot in a US-domiciled fund. (A later third-party snapshot dated 31-Mar shows SK nearer 4%, consistent with profit-trimming into the June peak or share-class differences — both dates pre-date the current pullback; the primary factsheet is BlackRock's own PDF.) Firm-wide, BlackRock has additionally been a historical 5%+ holder of SK Hynix across its index and active books. On "does BlackRock have a target price": buy-side firms never publish targets — they publish positions. Making a stock the single largest holding of your flagship fund, ahead of Nvidia, then going on a podcast to coin "the RAMpocalypse" about its product, is the buy-side equivalent of a Strong Buy at maximum conviction. The sell-side targets stand separately: Citi ₩3.1M · consensus ₩3.32M · Korea Investment ₩4.7M · BofA $250 on the ADR.

What's real: Terafab is not just a logic fab. The SpaceX S-1 describes a "vertically integrated closed-loop single plant spanning lithography mask design, fabrication of logic and memory chips, and advanced packaging." Phase 1: $16.8B in Grimes County, Texas — 100 million sq ft planned, up to $119B if all phases proceed, Intel aboard for the logic process, a stated target of one million wafer starts a month by 2030. And three days ago the first concrete memory step: Tesla posted a "Memory Process Integration Engineer" role covering end-to-end DRAM process integration, sub-20nm node development, capacitor design through volume manufacturing, plus MRAM, RRAM and 3D DRAM. The ambition is genuine.

Did our numbers factor it? As demand, yes; as supply, zero — and both are correct. Demand: every chip Terafab ever makes — AI5/AI6 inference processors, Optimus brains, Cybercab computers, orbital modules — needs memory bolted on, and until (and mostly after) Terafab exists, Tesla and SpaceX buy it: they're inside the demand ledgers already, and Musk publicly thanked Micron for allocation. Supply: nothing before the 2030s, for four physical reasons. (1) One job posting is engineering step zero of the hardest manufacturing climb on earth — the DRAM capacitor process is the most jealously guarded IP in semiconductors, and CXMT with a decade and $50B+ of state backing remains 2-3 generations behind. (2) Tom's Hardware ran Musk's own targets through fab math: the memory portion alone needs ~12 DRAM fabs and ~$240B of front-end investment (the Big Three's entire fleet is ~36 modules built over 25 years); Jensen Huang's on-record warning: Tesla "may underestimate the years of expertise required." (3) The tool queue is our tripwire, and it's public: no memory fab exists without years-out ASML lithography orders — any real Terafab DRAM line will surface in the order book our capacity tracker already watches, long before a wafer moves. (4) Even full success is captive supply for Tesla's own new demand (a million Optimus robots is memory demand that doesn't exist yet) — in a market fulfilling <70% of orders, self-supplying your own increment isn't taking SK's customers; at the margin it slightly relieves the commodity lane after ~2031, which folds into bear row 7's window as a minor contributor, now formally noted there.

The symptom read — the strongest one yet: follow one man's arc across our file: Musk thanks Micron for making room (buyer) → declares memory Economics 101, prices only rise (pleader) → hires DRAM engineers rather than trust the market (builder). Add AMD buying Taalas to dodge the HBM queue and $38B of customer prepayments, and the pattern is unanimous: the people closest to the shortage are planning for it to outlast the decade. Checkpoints armed, fire-in-sequence: ASML orders attributable to Terafab memory · a DRAM team scaling past ~100 engineers or a big-three process lead poached · pilot-line announcement with a node spec · first functional die claim · any HBM/TSV hiring wave (none exists today — HBM is a second mountain behind the first).

Yes, Korean brokers cut targets — eight of them after Q2. Read their reasons before reading their numbers. Bucket 1 — mechanical catch-down (most of them): Mirae Asset cut ₩4.2M→₩2.8M while its own analyst wrote the correction is "excessive relative to fundamentals" — but the peer-linked valuation framework forces the target down when the whole sector's multiple deflates. He simultaneously noted DRAM spot prices rising for 50+ straight sessions past their previous high, limited China impact, and large additional shareholder-return capacity. Hankyung's meta-analysis says it outright: "not fundamental damage but valuation re-adjustment; the oversold zone makes dip-buying valid." And the market-wide tell: July produced 580 downgrade reports against 351 upgrades — the first downgrade-majority month this year, across everything. This is a sector multiple event, and targets chasing a fallen price (괴리율 narrowing) is what sell-side desks mechanically do. Bucket 2 — the consensus-miss optics: Q2's record ₩60T operating profit still missed an "extremely high" street bar (the 5-6% miss and 48% capture this site logged in real time). Bucket 3 — the genuine bears, counted: NH (₩3.4M — China tech-gap narrowing could pressure share/ASP) and BNK (₩1.48M — the supply-glut school, the true bottom of the fan). Two houses out of the market carry an actual bear thesis; the rest cut numbers while keeping Buy. Bucket 4 — the countertrend: Korea Investment raised to ₩4.7M; Kyobo, IBK and SK Securities held; and Kiwoom's "downgrade" note literally recommends adding for the rebound, calling the valuation "a very attractive zone" for 2027-28 HBM growth. Target dispersion now runs ₩1.48M to ₩4.7M — the street has published its own outcome fan, and its shape matches ours: thin bear tail, fat upper body.

Flow-pressure status, honestly: the margin-loan unwind ran in two waves (the Aug 3 −8.79% wave, then the Aug 6-7 −10.4/−4.9% wave off the ₩5.24T record balance — the largest on the KOSPI). The pattern is late-stage — forced selling typically exhausts within 2-3 sessions of each trigger, and the Aug 5 +5.77% bounce marked wave one's end — but it is not certifiably over until three things print: (1) the daily margin-balance series drops meaningfully (watch ~₩4T), (2) this Thursday's options expiry (Aug 13) passes — the last dated derivatives-pressure window, and (3) a green close on rising volume. Until then, mornings stay vulnerable to residual forced flow and the tape stays untrustworthy as a fundamentals gauge in either direction.

What it is, plainly: TileRT compiles a model's entire token-generation loop into one giant permanent program living on the GPU — no more launching thousands of tiny programs per token, which is the overhead that made GPUs feel slow for instant-response AI. Result, independently benchmarked: ~500 tokens/second per user on a standard 8-GPU B200 box — 1.9-3.6× faster than the best conventional setups, ~4.5× better end-to-end latency, already in production at Xiaomi and Z.ai. In SemiAnalysis's own image: it doesn't build a new rocket — "it straps solid rocket boosters to the Metro Bus" you already own.

Why this is quietly excellent for SK — three mechanisms. (1) The speed tier stays on HBM. The premium "fast modes" users are proving they'll pay extra for can now be served from the ordinary GPU fleet — and at batch-size-1, every single token reads the model's full active weights (21-42 GB per token) out of HBM: ultra-fast serving is the most HBM-bandwidth-hungry mode per token that exists. The market's premium tier just became a heavier HBM workload, not a lighter one. (2) The specialist threat shrinks. SemiAnalysis says it directly: Cerebras, Groq and SambaNova "are no longer competing against a clumsy kernel-launcher. They are competing against their own execution model, running on fungible hardware, reallocated by a config file." If the speed tier is a software setting on HBM machines, the case for buying fixed SRAM fleets narrows to the extreme top — bear row 14's severity decrements, and the checkpoint "hyperscaler deploys >10K specialist units" becomes less likely. (3) Fungibility protects the general fleet. A GPU pool shifts between training, prefill, normal serving and speed-tier by scheduler, hour to hour; an ASIC fleet fixes its ratios the day the purchase order signs. Buyers guessing wrong on GPUs rebalance in software; guessing wrong on specialist silicon strands capital. That logic keeps capex flowing to the HBM-carrying machines. The honest limits, kept: software can approach the memory-bandwidth ceiling but cannot raise it — Cerebras still owns the absolute top speeds; TileRT supports only a handful of models today with ASIC-style per-model compilation pain; and one passing SemiAnalysis aside doubles as pricing telemetry: memory latency hasn't improved in a decade "even as HBM prices continually increase." Their agent benchmark detail seals the context doctrine: real coding-agent traces run a median 140,000 tokens of input — the conversation is the workload now, and it lives in memory.

The contract revolution, now industry-wide and quantified. The three memory makers collectively hold ~$38 billion of customer advance payments for capacity through 2030 (Micron alone: $18B cash collected of $22B contracted across 16 partners; SanDisk $16.5B). Samsung's own July-30 call: 60-70% of its total capacity will run under long-term contract volume — five major data-centre customers signed, five more closing — "nearly all customers are requesting multiyear supply agreements." Five-year terms are the new standard, and many contracts carry a rolling clause that adds a year as each first year expires: a perpetual five-year commitment. And the asymmetry, per BofA via Chosun: take-or-pay clauses and price FLOORS that cap quarterly declines while permitting rises of 10-20% or more. Heads the suppliers win; tails they don't lose. This is the convexity finding — first spotted in SK's ceiling removal — now confirmed as the structure of the entire industry, with Samsung inside it as umbrella, not underminer. JP Morgan names the consequence: memory makers are moving "from cyclical raw-material suppliers to long-term partners with multi-year protection." The re-rating leg, in a bulge-bracket's words.

The shortage tax, visible on shelves: DDR4 spot runs +77% above July's contract price; locked-out smaller buyers pay it, and Dell, HP and Lenovo have raised PC prices 15-20%. Even Samsung's own Galaxy S27 Ultra may skip its memory upgrades — the AI buildout eating the flagship phone's parts bin. The duration telemetry: the digest carries a Reuters-attributed remark that SK's CEO expects 2027 to be the worst memory shortage ever, with demand exceeding capacity beyond 2030 (flagged for primary verification next sweep) — consistent with his November framing and with Goldman's capex-to-2031 arc; if confirmed, it is the second independent counter-telemetry on bear row 7's timing. The honest symmetric note stays: the same analyses warn the prepayment cushion could erode by 2029 when new fabs reach full output — which is precisely row 7's window, unchanged.

Citi's numbers, two of them: the 20%+ pullback is a buying opportunity across the complex (upcycle "still in its early stages") — and the content bomb: HBM per AI system to grow 434%, from 20.7 TB to 110.6 TB. Note what that does to our conservatism ledger: the token ledger's content slider assumes 2.0× per GPU; Citi's system-level path is 5.3×. FMS 2026 closes the loop on the full-stack story: SK keynoted the conference with "Tiered Memory" — "no single memory type can economically serve agentic AI workloads alone" — the exact architecture (HBM + DRAM + CXL + AI-NAND) this site mapped through Storage-Next; Samsung answered with zHBM (memory stacked directly atop accelerators), 400+ layer NAND and its first processing-in-memory chip. And a software echo: MinIO and Neo4j both shipped agent-memory platforms this week — persistent context becoming a product category is the demand side of the same doctrine: the conversation is the part that grows.

The specs, verified: each Groq 3 LPU carries 500 MB of on-chip SRAM at 150 TB/s — roughly 7× the bandwidth of a Rubin GPU's HBM4 (22 TB/s). A full LPX rack: 256 LPUs, 128 GB total SRAM, 40 PB/s aggregate. Spectacular numbers — until you look at the other column. Nvidia's own Ian Buck: the LPU has "1/500 of the memory capacity" of a Rubin GPU. A whole LPX rack holds 128 GB against the 20,700 GB of HBM4 in one Vera Rubin NVL72 — 0.6%. A trillion-parameter model needs ~8 entire LPX racks just to hold its weights once, with zero room left for context — which is why The Elec notes it cannot run frontier models alone, and why it ships only paired with Rubin.

The detail that answers the demand question — from Nvidia's own architecture docs: in the pairing, the two processors co-compute every token, split by workload: attention decode — the KV cache, the per-user growing context, the fastest-expanding memory demand in AI — stays on Rubin's HBM; only the feed-forward layers (fixed weight-streaming math) offload to SRAM. Nvidia architecturally assigned the unfreezable workload to HBM and the frozen workload to SRAM. Our doctrine, drawn as a wiring diagram by the counterparty itself.

Net effect on SK demand, honestly: (1) Zero displacement — every paired NVL72 keeps its full 20.7 TB of HBM4; the LPX adds beside it. (2) Net positive — 35× tokens per megawatt means more revenue per gigawatt, which accelerates gigawatt construction, which multiplies Rubin racks, which multiplies HBM (Chain #5). Faster librarian, more books. (3) The capacity gap widens, permanently: SRAM bit cells stopped shrinking years ago (below ~5nm they barely scale), while HBM capacity doubles per generation — physics forbids SRAM from ever catching up. (4) The LPU is fabbed by Samsung's foundry on 4nm (Huang publicly thanked Samsung) — a foundry win for the rival, not a memory threat to anyone. Tripwire status: the co-compute design means LPX structurally cannot displace HBM sockets — it requires them. The watch stays armed only for a hypothetical future standalone SKU.

Hyundai Motor Securities: with HBM4 contribution rising, DRAM average selling price projected +19.9% next quarter even as the LTA share grows — a named broker independently printing our Rung-1 assumption (~+20% realised) almost to the decimal. SK Securities writes the multiple thesis in its own words: HBM fulfilment running below 70% of demand; the LTAs' core value is a "mutual binding structure" ensuring profit sustainability; and the ₩100T net-cash goal may arrive ahead of schedule (Kioxia asset-sale dividends + ADR proceeds) — bringing the returns policy forward, which "helps the market recognise the value of the LTA model and further drives up the valuation." That is the re-rating leg, published by a Seoul desk. Mainstream arrival: the Motley Fool ran the anomaly gap as a headline — "SK Hynix supplies more than half the world's HBM. Its stock costs under 4 times next year's earnings" — citing the 58% share, ~10 LTAs, HBM4 ramping and HBM4E samples already out, and the $500B+ group partnership with Nvidia. Also confirmed: the expanded Q3 returns story is real and building (TheNextWeb), and the full week's tape for the record: −8.79% · +0.64% · +5.77% · −10.37% · −4.88% to ₩1,730,000. The A-list remains silent — still nothing fundamental for them to react to.

What actually started the fall (the margin loans were the amplifier; this was the spark): Korean forum chatter (~Aug 5-7, unconfirmed) claimed SK cut a "special deal" pricing 2027 HBM4 for Nvidia near $2.8/Gb — "only" ~+50% against the ~+80% the market had come to hope for — allegedly in exchange for SK Group getting priority Nvidia GPU allocation. The market traded the rumor as fact: −10.4%, −4.9%.

Now the decomposition that matters. (1) The "disappointing" number is far above our book. The printed record (Dealsite): SK's current HBM4 contract runs at mid-$500s per stack ≈ $1.9/Gb — the exact figure every revenue chain on this site is built on. The rumored $2.8 is therefore ~47% above our assumption. If it prints, our 2027 HBM lane, the Nvidia pass-through (₩44T), and Rung 1 all get upgraded, not cut. The market crashed toward a number that would raise these forecasts. (2) If it's false, the reversal is violent. Citi (read-through): the price is "not locked, still negotiating" — Buy reaffirmed, target held ₩3.1M, 2026-27 operating-profit estimates nudged up; Samsung is quoting HBM4 at $4.5-5/Gb and (Dealsite) explicitly targets the same price as SK, refusing to undercut — the second supplier acting as a price umbrella, not a price war. Digitimes forecasts 2027 HBM4 at $4-5+/Gb. Every printed and quoted number in the record sits between our $1.9 and $5. (3) The same-day tell: on the −10.37% day itself, Moneytoday reported GPU makers' HBM4 cost burden nearly doubling and being passed through, with ASPs staying firm and both suppliers' earnings runs intact — the tape crashed against its own day's fundamental reporting. (4) The allocation-swap wrinkle, read straight: if SK really traded some price for priority GPU allocation, that is the MaaS rung-2 strategy becoming visible — the group securing scarce Rubin for its own 2-GW AI factory. A strategic purchase, not a surrender. (5) The second rumor (HBM4 defects / a Vera Rubin timing slip / an analyst's access restricted) stays rumor-grade: watch for an IR denial; note Nvidia already trimmed next-gen specs on component scarcity by Forbes' account — a different mechanism than defects. Korean retail has turned sour while global sell-side stays Buy/Overweight — historically a contrarian tell, and the A-list sources (Patel, Baker, Kindig) are silent, which is itself information: nothing fundamental has changed for them to react to.

The tape, day by day: Wed Aug 5 +5.77% to ₩1,668,000 (the bounce) · Thu Aug 6 −10.37% (the flush resumed) · Fri Aug 7 −4.88% to a ₩1,730,000 weekly close. That close sits ₩2,000 — 0.14% — above the ₩1.730M thesis-intact line: desk prediction P4's weekly-close test was kissed, not breached. Armed and decisive next week. The extraordinary part: on Friday itself the company approved the Yongin·Cheongju expansion (reported at ₩38-54T) — in a shortage regime, paid-up future revenue — and announced expanded Q3 shareholder returns, the early edge of the very catalyst our spring schedule dates to Dec 31. The stock fell 4.88% anyway. A tape that sells a returns announcement is not pricing fundamentals; it is a flow regime — and the spark is now identified (see the Aug 9 signal above: an unconfirmed $2.8/Gb HBM4 rumor that, even if true, sits ~50% above this site's assumptions) — amplified by the margin unwind, plus two fresh overhangs: Solidigm Nasdaq-listing talk and Chongqing plant stake-sale speculation — both, note, value-crystallisation events if real, not value destruction). Where the street sits: consensus target ₩3.32M (individual reports an extreme ₩1.48-4.70M spread — the market publishing its own outcome fan), forward PER ~4.1×. At ₩1.730M the stock trades at ~3.5× our 2027 clean EPS while carrying +44% on entry. Falsifier check, run honestly: nothing on the list fired — pricing intact (UBS hikes standing), demand intact (the same week delivered SemiAnalysis's turbine-counted 10GW validation and Nvidia's $91B guide), capex accelerating. The gap between the tape and the tape-measure is now the widest since entry.

The validation, physical: our #1-ranked source independently concludes SpaceX's 10-GW-by-end-2027 is real — not by trusting the call, but by counting: the Southaven power plant grew from 27 turbines (~495MW) in February to 69 turbines (>1.2GW) in July, satellite-tracked; the MiniHard greenfield is reaching ~500MW in five months; Colossus 1 took 122 days. Their read on the financing answers yesterday's open question — Nvidia vendor financing is likely why Musk declared Nvidia-exclusive (SpaceX evaluated TPU and AMD; the money argument won), paired with "value-based pricing" at $30-50M/MW/yr on 3-5-month lead times, the fastest delivery in the industry.

The economics layer — the tokens finally priced: their Inference Simulator (real production coding traces on simulated silicon) finds OpenAI and Anthropic can each generate >$100B per GW per year selling frontier API inference, against ~$12B/GW/yr of rental cost — an ~8× revenue-to-cost ratio; Opus-class margins 85%+, now the consensus number; a leaked DeepSeek call confirms 10-month GPU payback. "Serving inference tokens is unbelievably profitable." Combine with our counted stacks and the leverage stat writes itself: the $2.2B of HBM in a gigawatt is repaid by ~8 days of the inference revenue it enables. That is why buyers do not blink at memory price hikes — and why the elasticity in our bridge is so low. The demand side doubles: Microsoft has "awakened" — 10 GW of binding contracts signed year-to-date (~$300B), the April renegotiation dropped OpenAI's 20% revenue share, and SemiAnalysis models Azure accelerating from ~42% toward 100%+ growth serving OpenAI models at $100M/MW/yr with zero training cost. SpaceX's path: ~$300B ARR by end-2027 with only half its incremental compute monetised.

The pass-through ladder — Nvidia's future revenue, translated to SK: Nvidia is tracking ~$400-430B this fiscal year (Q1 filed $81.6B, Q2 guided $91B), and the street now models next fiscal year (FY28) near $561.5B — revised up 123% in fourteen months. Apply the filing-anchored ratios (cost 25.1% of revenue; HBM ~39% of cost ≈ ~10% of Nvidia revenue — the same dollars as the famous "52→62% of BOM" stat, which measures the parts list only, before packaging and test): FY28 implies ~$56B of HBM flowing through Nvidia alone → SK's ~56% ≈ $31B ≈ ₩44T from one customer — and adding the non-Nvidia accelerator wave (TPU, Trainium, Jalapeño, MI-series ≈ 40% of HBM demand) lands total SK HBM in the ₩60-75T lane. That is the fourth independent route to the same number (token ledger · elasticity bridge · BofA TAM · now the Nvidia income statement), and it sits inside the revenue book the extrapolator already carries. Thesis status: reconfirmed at every joint it touches — repricing quantified (row-12 mitigant now has margin math), the buildout self-funds (8× ratio, 10-month paybacks), and the largest new buyer's numbers survived tier-1 physical audit.

The buildout, from the call: SpaceX ended the quarter at 1.4 GW of nameplate compute, expects >2 GW by year-end and "closer to 10 than 5" GW by end-2027 — with a tentative 20 GW of power and cooling ordered ahead of the chips ("close to 15 GW" even if a quarter of projects slip, by Musk's own haircut). All of it exclusively Nvidia Vera Rubin — Kyber NVL72 racks on the ground, Space-1 Rubin modules on the Starmind satellites launching next year — and the line that matters most: "Our understanding with NVIDIA is that we will receive a very significant percentage of their GPUs next year." A single buyer claiming a major share of 2027 Rubin output. The economics that fund it: the CFO says cloud rental deals carry sub-one-year payback, with $6.7B contracted in the first weeks of Q3 — and only ~10% of capacity rented out, the rest feeding Grok. Musk's monetisation guess, labelled as his guess: $30-50 per watt of Rubin. Sub-year payback on rented compute is the receipt that makes the Dwarkesh repricing thesis investable at infrastructure scale — and it is why the buildout self-funds (the bear-row-12 mitigant, printed by a third named operator). The Terafab ($20-25B) is framed by Musk as "survival infrastructure" for his own shortage, not competition — and Forbes notes the scarcity is severe enough that Nvidia has trimmed specifications on next-gen products over component supply.

The ledger math (Chain #5, single buyer): +8.6 GW of compute in ~17 months ≈ ~4.3M Rubin-class GPUs at the Naver calibration ≈ 1.24 EB of HBM4 ≈ ~$19B of HBM — SK's ~56% share ≈ ₩15T from one customer's one-year ramp, roughly 2.7% of 2027E revenue, before anyone else's guide. The waterfall, reconciled (asked Aug 8 — a fair challenge): HBM is 52-62% of Nvidia's cost to build the accelerator, not of the price customers pay. Per gigawatt: system capex ~$50B → Nvidia collects ~$22.5B for the accelerators → full manufacturing cost ~$5.8B at Nvidia's printed ~74% gross margin (margin is stated on revenue — settled by the filing itself, FY27 Q1 ended Apr 26: revenue $81.6B, gross profit $61.1B = 74.9% GAAP, cost of revenue $20.5B = 25.1% — a ~298% markup on cost; a 100% markup would be only a 50% margin, and would predict $40.8B of cost against the $20.5B actually filed) (component BOM ~$3.5B of it; CoWoS packaging, test, yield and freight the rest) → the HBM inside is ~$2.2B — 62% of the component BOM, ~38% of full cost, 10% of Nvidia's revenue, 4.4% of system capex: same dollars, four honest denominators. The cleanest anchor skips percentages entirely: count the stacks — 500K GPUs × 8 × ~$550 printed contract price = $2.2B/GW, no margin assumptions required. The ~$16.6B/GW between cost and price is Nvidia's gross profit — the wedge on the table, and the entire multiple-legs campaign (custom HBM with logic base dies, ceiling-free pricing, MaaS) is SK's bid to migrate points of it across: each 1% of system capex captured ≈ $0.5B per GW. Sanity check that catches the error instantly: at $25B of memory per GW, the world's ~40 GW/yr of builds would need ~$1T of memory annually — against an entire 2026 HBM market of $54.6B. And one honest enrichment the challenge produced: the ₩15T was HBM-only — add host LPDDR/SOCAMM, CXL tiers and AI-NAND and the single-buyer full-stack memory prize runs ~$24B, SK share ≈ ₩19T. And the memory quote, verbatim: "Memory output is increasing by around 20% per year… ask yourself, is demand increasing by 20% a year? No, demand is increasing by 200% a year, maybe higher. So if you have demand increasing much faster than supply, then Economics 101 would suggest that the price increases. It does not decrease." His numbers (200 vs 20) sit far above our elasticity-bridge defaults (28 vs 18) — the engine's conservatism now has an insider bound to measure against.

The capex frame around it (I-O Fund, verified): Big-4 H1 actual $301B; full-year guides now $732.5B (+79% YoY, a 15-point acceleration; 39% above what the street expected nine months ago); and Goldman's long arc: $7.6T cumulative 2026-31, still growing to $1.64T in the final year — which doubles as counter-telemetry on bear row 7: the 2029-31 capacity wave may land into demand that is still rising. The anomaly-gap note: all of this printed in a month when memory stocks fell — retail sentiment on the ADR reads "bearish" while the largest new buyer in history claims half a year's GPU output and calls memory the binding constraint. That spread between tape and tape-measure is the original thesis of this site, widening.

What happened: AMD reached a definitive agreement to acquire Taalas (Toronto, founded 2023, $219M raised; price undisclosed; closes Q4-26) — the startup that etches a model's weights into the chip's final metal layers, so the circuit is the memory. Its HC1 test chip on TSMC 6nm served Llama-3.1-8B at ~17,000 tokens/second — claimed 48× a GPU and 8.5× Cerebras (vendor's own February numbers, unverified independently). On-chip SRAM carries the KV cache and small adaptation layers. Core design: no HBM. And it isn't alone: AMD integrated Cerebras wafer-scale engines into its systems in July — both GPU giants now own an SRAM/etched specialist (Nvidia-Groq, AMD-Taalas/Cerebras). Google shelved this exact idea for Gemini; Taalas shipped it for pinned open models — the segment difference explains both choices.

The scenario, segmented. Etched silicon can serve only workloads that are simultaneously stable (model frozen for months — the metal-mask cycle plus qualification is weeks-to-quarters at best), small-to-mid (HC1 bakes an 8B model; multi-die needed beyond ~70B), and short-context (SRAM caps the KV cache hard). That intersection — moderation, translation, high-volume pinned assistants, per-enterprise fine-tunes via adaptation layers — is real and generously reaches 25-35% of raw token volume by 2028-29, but a far smaller share of HBM-attached value, since those tokens already run on the cheapest silicon while frontier and agentic work anchors the HBM fleet. Displacement math: if MSICs capture half that segment by 2029 (aggressive — requires the Q4 close, Helios integration, per-model volume economics, hyperscaler qualification), serving-HBM bits displaced ≈ 4-8% — roughly five to eight weeks of demand growth at Chain #6's rate. And through ~2028 the regime is supply-constrained: substitutes at the low end serve demand SK's wafers couldn't reach anyway; every bit SK can make still sells.

The symptom read — the part the headline hides: AMD sits behind Nvidia in the HBM allocation queue. Assembling an HBM-light inference portfolio (Taalas + Cerebras + MK1) is partly a workaround for the shortage itself — the acquisition is evidence of scarcity as much as a bet against it. Taalas dies are fabbed on mature 6nm; they compete for none of SK's supply.

The sentence survives, amended: you can now freeze the weights commercially — you still cannot freeze the conversation (SRAM-capped context excludes the fastest-growing workload), you cannot freeze frontier iteration, and you cannot touch training. Where the real risk lives — the coupling: MSICs matter most exactly when bear row 7 matters — if the 2029-31 capacity wave lands just as etched silicon caps the stable-segment price, the two rows compound. Bear row 14 opened (severity LOW through 2028, MEDIUM 2029-31), jointly conditioned with row 7 on LTA coverage. Checkpoints, dated: deal closes Q4-26 → first Helios-Taalas SKU (~mid-2027 watch) → first hyperscaler deployment >10K units → demonstrated model-refresh <8 weeks at volume → >5% of inference tokens on non-HBM silicon. Fire in sequence and the row upgrades; until then, the 12-month numbers are untouched — zero 2026-27 volume exists.

Yes — it's the same story, and the group has named it. Chey Tae-won is publicly exploring "Memory as a Service" (MaaS): just as Amazon and Google lease computing, SK would supply memory resources plus the operating software as an ongoing service — pooled, composable, billed per need. The CXL plumbing, the "stranded memory into tokens" push, and the AI-N storage tiers are the technical floor MaaS would stand on; the customer's payoff is exactly the tokens-per-watt utilization gain, and SK's payoff would be recurring service revenue on top of chip sales. Chey's own rationale: memory is the bottleneck, every customer needs a different memory mix, and "a single solution cannot solve every problem… you need to adjust the memory stack to each market, and you also need special software that can implement it."

The three rungs, honestly graded. Rung 1 — already earning (quasi-rent): the ceiling-free, volume-locked LTAs with customer prepayments are the closest thing to rent a manufacturer can write — contracted offtake, floating price, cash in advance. This is live now and it is the convexity finding by another name. Rung 2 — being built: the SK-NVIDIA Korea AI factory (the 2-GW SKT venue) is explicitly framed around GPU-as-a-service bundling SK memory into a managed offering — SK participating in service economics through the group's own data centres — alongside CXL modules and AI-N sold as co-designed capability, and a stated plan to double wafer capacity over five years underneath it all. Rung 3 — the full MaaS vision: pooled memory billed like cloud. Chey's own honesty, kept intact: "an idea at this point" — no pricing model, no scope, no commercialisation timeline disclosed. What rent would mean for the stock: service revenue is valued differently from cyclical hardware — every rung climbed compresses the cyclicality discount that has always capped memory multiples. This is the same third leg of the multiple argument (platform economics), now with the group chairman describing the destination out loud. It does not make SK a hyperscaler; it makes SK progressively harder to value as a commodity.

Vision ladder for the record: CES 2025 "Full-Stack AI Memory Provider" → AI Summit Nov 2025 "Full-Stack AI Memory Creator" (co-architect, three pillars: custom HBM, AI-D, AI-N) → Jul 2026 MaaS exploration. Each step moves the company one layer up from the die.

Exclusive one — Frozen v2, the fuller picture. Google is designing a server chip that hardwires Gemini's architecture into silicon — 6-10× more tokens per watt than its latest TPU, deployment ~2028 — born from a compute shortage severe enough that Google Cloud has been turning away customers. The detail that matters most sits in what they didn't do: the original plan to etch the weights into the chip was put on hold because models update too fast — it would shorten the chip's lifespan. So the weights stay loadable (in memory), the project is framed internally as a trial run, and it is explicitly "not meant to replace TPUs" — TPUs which themselves carry 288GB of HBM each. Exclusive two — the partner-hedge. Nvidia's pattern: rather than fight the specialists, absorb them — the $20B Groq deal (a non-exclusive license plus the CEO and team, structured, per two senators' letter, to sidestep antitrust review) and now a system combining Nvidia hardware with d-Matrix, the SRAM in-memory-compute startup. And here is Huang's own description of what the Groq LPX racks do inside a Vera Rubin system, verbatim: "we're just going to add a whole bunch of Groq chips, which expands the amount of memory it has." The SRAM specialists are being deployed as memory tiers inside HBM-anchored racks.

The structural read — what can be frozen and what cannot. Specialization attacks the stable parts of the workload: fixed architectures, small hot weights, predictable decode. Three things it structurally cannot touch: (1) context — the KV cache grows with every agentic step, is different for every user, and cannot be etched into anything; it lives in HBM and DRAM, and it is the fastest-growing demand vector in the entire stack; (2) iteration — frontier weights change weekly, which is exactly why Google shelved weight-baking; (3) training — untouched entirely. You can freeze the architecture; you cannot freeze the conversation. Meanwhile the custom-silicon wave itself — Google TPU, OpenAI's Jalapeño inference chip, Amazon, Meta, and now Anthropic reportedly discussing a chip partnership with Samsung — multiplies the number of HBM-buying accelerator programs, diversifying SK's customer base beyond any single buyer. Sponge accounting, honestly kept: this is the fourth and fifth efficiency instrument logged this year (after K3's brute inefficiency, LUT-B, and the cost-per-token curve); each previous one was outrun by usage. Tripwires armed rather than assumed: weight-baking revived into production (ESCALATED Aug 6 — AMD acquired Taalas; five dated checkpoints now govern, see the Aug 7 signal above; decremented Aug 10 — TileRT shows the speed tier becoming a config file on ordinary HBM hardware, squeezing the specialist TAM from the software side) · Frozen displacing TPU-with-HBM orders at scale post-2028 · SRAM racks displacing rather than accompanying HBM sockets in Nvidia BOMs · standalone Groq/d-Matrix hyperscale wins outside paired racks.

The frame: as hyperscalers shift from building to monetizing, the metric that separates winners is tokens per watt — tokens are the revenue side, watts the cost side, and power is no longer abundant (ERCOT: ~390 GW of data-center interconnection requests against 23 GW added in two years; chip power H100 700W → Blackwell ~1,300W → Rubin ~2,300W; hyperscaler PUEs already squeezed to 1.09-1.17, so the cooling lever is spent). The claim that matters for us, in their words: "when it comes to improving tokens per watt, memory is the biggest constraint" — GPUs sit idle waiting for KV cache, and the memory wall data is stark: over twenty years compute grew 3× every two years while memory bandwidth grew 1.6× and capacity 2×; even Rubin's 5× inference leap rides on HBM bandwidth up only 2.8× and capacity 1.5×. The receipt — the single best value-capture statistic on this site: HBM already represents ~52% of a B200's bill of materials, rising to ~62% in Rubin. The accelerator is becoming a memory product with a GPU attached; Nvidia sells the box, SK supplies the majority of its cost. The offload-engine response — and why it is additive, not substitutive: Nvidia's CMX (proprietary) and the open CXL ecosystem both aim to "turn stranded memory into tokens" — feeding KV cache from DRAM tiers below HBM. That expansion memory is standard server DRAM at ~$40/GB (Marvell's print — against >$200K/TB for HBM bought through GPUs), which means the utilization war creates a new attach lane for SK's commodity bits exactly as they were being written off to CXMT: CXL expander modules, SOCAMM, pooled DRAM. Same complements logic as Storage-Next — the tiers feed the HBM, they don't replace it (the wall guarantees that). One honest watch item: Nvidia's 35×-tokens-per-MW showcase pairs Vera Rubin with the Groq 3 LPX — an SRAM-based decode engine carrying no HBM. Today it's framed as a co-processor beside HBM-anchored Rubin racks; if SRAM offload ever starts displacing HBM-carrying sockets rather than accompanying them, that's a genuine substitution channel — now fully decomposed in the specialization-wave signal above, where Huang's own words describe the Groq racks as memory expansion. Telemetry armed, alongside CMX-vs-CXL adoption and SK's CXL/SOCAMM design wins.

The tape: Monday Aug 3, SK closed −8.79% (₩1,567,000) with Samsung −8.76% in lockstep; Tuesday morning another −4.8% to ₩1,492,000 by 10am, Samsung −4.5% — while the Nasdaq rose +2.13% and the semiconductor index +1.05% overnight. Korean chips falling into an American chip rally is the signature of a local, mechanical flow, not a fundamental repricing. The mechanics, in order of weight: (1) The margin-loan complex (빚투). SK carries the largest margin-loan balance of any stock on the KOSPI — ₩5.24T — built up into the July highs. When collateral ratios breach, brokers force-liquidate (반대매매) at the open regardless of price; flow reports describe exactly these 8-9am forced sells running now, and each forced sale pressures the next account's collateral — the self-feeding loop that has nothing to do with HBM. (2) A rumor traded like a fact. Korean retail media circulated Aug 4 as a shareholder-return announcement date; the company promised a plan "within the year," never a date. Expectation built, the date passed unconfirmed, and the disappointment sold — a phantom catalyst failing on schedule. (3) The amplifiers: the 2× leveraged ETF complex rebalances with the move into every close, and institutional profit-taking joined midday. Options are the minor voice here — the Korean monthly expiry is Aug 13 and the ADR chain is young; this unwind is a cash-margin story, not a derivatives one. What it is not: no operating fact changed. No falsifier condition is near firing. The same week this flow runs, the September price-hike stack (UBS +32%) and the Storage-Next architecture both strengthened the fundamentals. Level check, honestly: ₩1.49M sits back inside the crash gap zone (₩1.52-1.64M reclaim lost), with the ₩1.42-1.40M confluence — the rising 200-day line, the technical thesis-intact level — 4.9% below. A weekly close under it without a fundamental falsifier is the one technical event that earns a seat at the fundamental table (desk prediction P4). Position vs entry: +59% at Monday's close, ~+51% intraday.

What happened today: at the Future of Memory and Storage conference NVIDIA open-sourced its cuFile API — the direct pipe that lets GPUs read storage without the CPU middleman — and launched Storage-Next, an industry initiative built on SCADA (scaled accelerated data access) pulling storage makers, controller vendors, cooling and standards bodies into one architecture for GPU-driven storage. NVIDIA's stated reason is our thesis in their words: mixture-of-experts models "think extremely quickly, rapidly pulling data and firing tools for agentic workflows" — the token machine needs a bigger fuel line.

SK's position inside it, with dates: the joint program (NVIDIA codename Storage Next, SK codename AI-N P) targets samples by end-2026 at 25 million IOPS on PCIe Gen 6 — roughly 8-10× today's best enterprise SSDs — and mass production by end-2027 at 100 million IOPS, a 30-50× leap. The family has three lines: AI-N P (performance), AI-N B — High Bandwidth Flash, co-developed with SanDisk (alpha spec 2026, evaluation samples 2027), and AI-N D (capacity). Phison's CEO supplies the market backdrop: NAND pricing sharply up, 2026 capacity effectively sold out, relief not expected before late 2027.

What it means, decomposed. (1) The third lane gets an AI story. NAND — the least-loved ₩119T of the stack, already showing eSSD revenue doubling QoQ — becomes an active tier of the AI memory hierarchy: model parameters and agent context that cannot fit in HBM stream from AI-NAND at memory-like speeds. (2) It makes HBM MORE valuable, not less. These tiers are complements: a faster fuel line means GPUs starve less, so every HBM bit does more work — and context offload is precisely what enables the long-running agents that then demand more HBM per step. The substitution fear runs backwards. (3) The moat extends. An AI-SSD co-designed into NVIDIA's reference platforms carries the same qualification hysteresis as HBM — the E13 lock-in structure now reaches into part of the NAND lane. (4) Value-chain position. SK is selling co-designed capability — capacity plus bandwidth plus the I/O path — under long-term co-development agreements, not commodity SKUs. That is the third leg of the multiple argument: ceiling-free LTAs (convexity), ~10 signed LTAs (de-cyclicalisation), and now full-stack co-architecture (platform economics). Honest boundary: this does not make SK a hyperscaler and it will not earn a hyperscaler's P/E — what it does is keep compressing the cyclicality discount that has always capped memory multiples.

Announcement-stage architecture — watch, do not price. Signposts: AI-N P sample delivery (due end-2026), HBF evaluation units (2027), NAND-lane margin prints, eSSD growth rate. First revenue relevance 2027; first margin-mix relevance 2028.

The migration completed: the buildout was first counted in chips, then dollars, then gigawatts — and now in tokens, the actual output of the machine. The numbers, verified: consumption runs ~370 trillion tokens a day (~135 quadrillion a year) — already 2.4× Dell's 2028 forecast, and Dell's COO had already revised that forecast up 57-fold from his original model and "still suspected it was too low." Tirias missed its 2024 forecast by 33×. Goldman (May 2026): ~47Q tokens/month by 2028 (~565Q/yr, 10× Dell) rising to ~120Q/month by 2030 — a ×24 multiplication driven by agentic AI, which becomes "the large majority" of volume. Google alone: 3.2 quadrillion tokens/month — 330× in two years, 7× YoY, with Gemini at 900M users. Why this is a memory story before it is anything else: a generated token is a read event — the model's active weights and its entire growing context are pulled from HBM for every step, and decode is bandwidth-bound. Token growth is HBM-bandwidth demand growth with three multipliers stacked: more tokens, more tokens per task (agent chains), more context per token (KV memory). The sponge, quantified by Goldman itself: cost per token is falling 60-70%/yr — and the Jevons receipt printed anyway: Google's volume rose 7× through that price collapse. Chain #6 on the engine page now runs this ledger live with three independent cross-checks (fleet ≈ installed base, implied revenue ≈ lab run-rates, 2028 ≈ Goldman) — output: serving HBM bits ~×2.4 by 2028 after conceding the full efficiency curve, before training, ASICs and build-ahead add their layers. And Goldman's Schneider hands bear row 12 its second named mitigant: raising gross margins raises operating cash flow, "and that gives you more headroom to spend."

What happened: consumer DDR5 spot quotes slipped for the first time in weeks on buyer resistance — the report circulating this weekend. What the primary tape says, updated yesterday (Jul 31): DRAM contract prices continue to rise, with manufacturers raising quotes on tight supply and expectations of negative production growth; PC-lane gains have narrowed into a buyer-seller stalemate; notebook shipments are dropping sharply as retail prices suppress consumer spending; and original manufacturers' PC supply is being crowded out by server demand — with sellers retaining bargaining power on rigid baseline demand. The precedent that already ran this movie: German DDR5 retail fell 7.2% in March — the first monthly decline after eight straight gains — was called a top by some, resolved as a consumer-driven, short-term adjustment with contracts firm… and by July the same spot chip printed $61 highs, with the shortage cascading so hard that DDR4 scarcity pushed buyers down to DDR3, lifting DDR3 prices. The mechanism, with its number: HBM consumes roughly 3× the wafer capacity of equivalent commodity DDR5 — every wafer SK pivots to the fortress lane removes three wafers' worth of consumer supply. Apple saying it cannot get enough, and OEMs shipping thinner memory configs, is not weakness — it is what rationing looks like from the buyer's side. Per-unit consumer content falls precisely because the bits are worth more elsewhere; in a supply-determined regime, SK's revenue follows the wafers, and the wafers are walking up the value stack. The honest ladder, position marked: step 1, consumer spot wobbles (now — benign, recurring) → step 2, PC contract gains narrow into stalemate (now printing) → step 3, server/enterprise contract resistance (not present — 2027 talks "progressing smoothly") → step 4, two consecutive quarterly contract declines (the armed falsifier — RED if it fires). We are at step two of four, in the lane that matters least. Superseded: the verified Q3 stack (UBS +32%/+18%, Samsung ≤20%, TrendForce 13-18%) lifted the extrapolator defaults to +20% Q3 / +10% Q4 — mid-range by rule, below UBS — moving E[V] to ₩3.90M. The consumer-spot wobble itself still touches nothing.

The arithmetic, from the post itself: frontier-lab revenue is compounding ~10× a year while lab compute supply grows only ~3× — so some mix of higher lab margins, higher compute prices, and more compute shifted to inference must happen, and all three are already printing. The receipts: spot GPU prices are up 40%+ from the February trough; Google is paying ~$920M a month for ~110,000 GB200/GB300 GPUs from SpaceX through June 2029 (~$30B total) — roughly 2× the spot rate — with Google's own explanation being "unexpected demand" for its AI products; and Anthropic's inference margins reportedly climbed from 38% to over 70%. His thought experiment for where this goes: an H100 hosting a human-level software engineer justifies >$250K a year in rent — ~15× today's spot. Baker's response: called it "fairly wild" while admitting even he hadn't fully assumed rental prices would keep rising — and added the mechanism that matters most to our bear file: spot now rents at ≥2× contracted rates, which means hyperscalers are currently UNDER-earning their own fleets; as old contracts roll off and reprice, operating cash flow accelerates sharply — funding the capex wave without the debt spiral. That is a direct, dated mitigant to bear row 12, with telemetry: watch cloud OCF growth steepen through 2026-27 as contracts reprice. Why this is our story: when the rent on compute doubles, the value concentrates in the scarce physical inputs that gate it — and Baker names the stickiest one explicitly, citing a Micron/Meta white paper showing a ~38× slowdown when working data spills out of DRAM to SSD: "the DRAM bottleneck almost certainly lasts the longest," undersupply ranked HBM > DRAM > NAND — the exact ordering of our two-lane chart. Agentic workloads, long contexts, KV caches and test-time compute all intensify it. SemiAnalysis's frame from the March deep-dive completes the loop: memory supply grows only ~20-30%/yr against faster AI demand, and ASML EUV output becomes the binding constraint by ~2030 — our transmission chain, in their words. And note the buyer-class datapoint hiding in plain sight: SpaceX is now a hyperscale compute landlord — 110K+ GPUs of HBM bought by a company that wasn't in this market a year ago. Vet appendix (Aug 3, video version circulating): Dwarkesh flags his own weak link honestly — "there's no deep reason why the trend needs to continue… it's ultimately a question about AI capabilities" — the same conditionality as our Scenario D, which is why this corroborates rather than re-triggers. The sharpest published counter (AI Weekly): his pricing logic rests on lab preferences, which open-weight models and efficiency gains will constrain. Vetted: the counter attacks the price leg but concedes the volume leg — open-weight models need identical compute to run (K3 exhausted its maker's GPUs in 48 hours), and every efficiency instrument this year has been outrun by usage. Memory wins on either leg: high prices concentrate value in the scarce input; high volumes ship more bits. And the closing convergence, from a sceptic: a former OpenAI researcher who publicly disputes the essay lands here anyway — "the people making out really happily here are Nvidia and Micron, since they're the ones selling the chips to everyone in the race." When the argument's critics arrive at your position as their conclusion, the position no longer depends on the argument being right. Meanwhile the tape ran the waterfall live again: Wall Street read the same essay as a customer-cost story and sold Meta −10% and Google −8% — the funders fall, the arms dealers stand.

What's flagged for Q3 (Jul-Sep): TrendForce expects DRAM contract prices up 13-18% QoQ and NAND 10-15%, with supply "extremely constrained" but consumer softness capping the pace. Samsung is reportedly asking customers for up to 20% on DRAM and more than 20% on LPDDR; Goldman lifted its Samsung DDR5-RDIMM expectation to +13% QoQ from a prior +5%. Our extrapolator's Q3 assumption of +15% sits mid-range of the forecasts and below Samsung's ask — built in, and deliberately not at the top.

The bigger finding: ZDNET-sourced reporting says SK Hynix has eliminated price ceilings in its long-term supply agreements — locking customer volumes while letting prices float with the market — in contrast to Micron, which keeps both floors and ceilings. Analysts in the same reporting estimate SK's DRAM gross margin already passed 90% in Q2. Read the structure: this is the strongest possible seller's contract — guaranteed offtake, uncapped price — and it reconciles the apparent tension in Song's call language: the LTAs damp volume risk, not price upside. Consequence for the model: our 50% pass-through damping (validated at 48% in Q2) is correct for the contracts that governed Q2 — but as ceiling-free agreements phase in, the damping factor should fall and realised ASP should track spot more closely. Discipline: the default stays 50% because it is the validated number; a tripwire now arms instead — if Q3 realised pass-through prints above 55%, the ceiling removal is confirmed and the slider moves. One more wafer fact logged: HBM will consume 23% of total DRAM wafer output in 2026, up from ~19% — the commodity squeeze that keeps conventional prices bid is itself a product of the HBM pivot.

Post-print consensus: KRX 12-month PT ₩3,251,340 across 37 analysts (high ₩5.3M, low ₩1.2M) — 36 buy, 1 hold, 0 sell — implying ~90% upside from ₩1.718M. The ADR consensus sits at ~$245-252 against $143.73. House read, per the standing rule: analyst targets are lagging confirmation, never signal — but their distribution is information. Three weeks ago our ₩3.86M E[V] sat above the street; the street has now migrated to bracket it (consensus ₩3.25M, high ₩5.3M), which converts our number from an outlier into the upper-middle of a crowd — epistemically safer, and exactly what the "street arrives underneath the model" pattern predicted on Jul 23. Also logged from the weekend sweep: Samsung is closing on ~30% of Nvidia's 2026 HBM4 allocation (supports our audited 56% SK default rather than the old 70% chatter), and Nvidia has requested 16-hi HBM4 deliveries by Q4 2026 — the mix-richening our GB-per-stack slider already assumes. Nothing in the weekend flow moves a model line; everything in it confirms one.

The five-session tape, in full, because the shape is the lesson: Jul 27 close ₩1,816,000 → Jul 28 ₩1,550,000 (−14.65%) → Jul 29 ₩1,401,000 (−9.6%, intraday low ₩1,246,000) → Jul 30 ₩1,322,000 → Jul 31 ₩1,718,000, limit up +29.95%. The KOSPI rose 1,001.89 points, +17.91% to 6,595.45 — its largest single-day gain in history, and SK hynix hit its price limit for the first time in seventeen years. The trigger came from overnight America: Microsoft's earnings beat restored confidence in the durability of AI capex, and memory ran — SanDisk +26%, Micron +18%. Samsung closed +26.81%. SK's 2× leveraged ETF gained as much as 55% intraday; foreign buying led the tape; market cap recovered to ~$849B, 18th in the world. Even after the limit-up move the stock sits ~42% below its ₩2,987,000 high.

Read what the week actually proves. The intraday low of ₩1,246,000 was −43% from the ₩1.730M stamp — reached on the day the company printed the best operating quarter in its history. Four sessions later, on no company news at all, the price was 38% higher. No operating fact changed in either direction. This is the amplifier we have been documenting since Jul 12 running in both directions at full gain, and it is the single strongest piece of evidence for the discipline the site is built on: the tape carries almost no information about the thesis at daily resolution.

The actuals, from the company: revenue ₩79.3187T, operating profit ₩60.5426T, operating margin 76%, net profit ₩93.9226T — all-time highs, with revenue +257% and operating profit +557% YoY, and first-half revenue crossing ₩100T for the first time in company history. Against consensus (₩84T / ₩64T) that is a ~5.5% miss on both lines, and the stock fell 9.6% after being down as much as 15% intraday.

Why it missed — and this is the whole story: Reuters attributes the shortfall to HBM4 shipments below expectations, pushing revenue recognition into later periods. That is our NH-Securities whisper from Jul 13, logged as the one real fundamental question in the file, resolving as TRUE — and as deferral, not destruction. HBM4 mass shipments began in Q2 with the full ramp planned for H2. The revenue moved right on the calendar; it did not disappear.

Disaggregating the headline, per house rule: net profit of ₩93.92T contains ₩63.3T of investment-asset gains from the Kioxia stake sale — a balance-sheet item crossing into the income statement, not operating cash creation. Strip it and tax the rest: clean quarterly net ≈ ₩46T, clean EPS ≈ ₩62K, annualising to ~₩247K. At ₩1,718,000 that is ~7.0× annualised clean earnings — and roughly 6.4× the full-year 2026 clean number the run-rate implies. Note the trap this creates for anyone reading a screen: Naver shows a trailing P/E of 16.4× on distorted earnings, while the operating business trades at less than half that.

What the operating detail actually said. DRAM ASP +~30% QoQ, NAND ASP +mid-50%, enterprise SSD revenue doubled sequentially. Long-term agreements now signed with about 10 key customers to lock mid-term supply visibility. Gross cash ₩88T, net cash ₩69.4T. Q3 guide: DRAM bit shipments +~10% QoQ, NAND low-single-digit; full-year demand growth mid-20% DRAM, high-10% NAND. Capex raised to the high-₩40T range, accelerating M15X mass production and pulling forward Yongin Phase 1 (cleanroom early 2027), with P&T7 and M17 behind it. NAND 321-layer to reach ~50% of domestic capacity by year-end.

The Q&A, where the forward information lives. On HBM4 competitiveness (SK Securities): leadership rests on time-to-market, yield, quality and customer trust accumulated since HBM2E; HBM4 yields and quality are near HBM3E maturity; HBM4E samples delivered with volume production in 2027; iHBM in preparation for HBM5 to improve thermal management. On 2027 HBM pricing (UBS): discussions with key customers progressing smoothly, with pricing to reflect not just conventional DRAM trends but the higher resource needs, technical complexity and value of HBM — the de-commoditisation argument, stated by the seller. On shareholder returns: management is reviewing additional measures and aims to share a plan with the market within the year. President Song Hyun-jong: customer demand remains robust, buyers keep asking for more memory, and the company is pursuing further long-term contracts to reduce exposure to price swings.

What actually happened. (1) CXMT listed on Shanghai's STAR Market — an ~$8.6B raise, Asia's largest of the year — and surged 466%, which Kiwoom notes may pull global capital toward the Chinese name and weaken relative flows into Korean chips. (2) Apple is reportedly testing CXMT's DRAM — the first credible sign of a top-tier Western customer qualifying Chinese memory. (3) The Information reports China has begun producing domestic immersion DUV lithography machines — delivered this year to SMIC, Hua Hong and CXMT. ASML fell ~5.7-7%, its worst day since June.

Now the decomposition, which the tape did not do.
Scale: CXMT is the world's fourth DRAM maker at 8% share — against Samsung 36%, SK 29%, Micron 24%. The lithography programme targets ~5 machines in 2026 and ~20 in 2027, against ASML's ~130 immersion systems a year at 98.7% share. Tom's Hardware puts the tool at roughly ASML's 2008-era NXT:1950i — 28nm-class single exposure, sub-10nm unlikely before 2030 — and some critical components are still Japanese.
Product: this is commodity DRAM, not HBM. HBM needs TSV stacking, MR-MUF-class packaging and the advanced nodes that EUV enables. ASML's own CEO said rising DRAM litho intensity partly reflects customers replacing multipatterning with cheaper single-exposure EUV — so a DUV-only path means more masks, worse overlay, lower yield and higher cost per bit. China can build a lot of DDR5; it cannot cheaply build the memory that sits on a Rubin package.
Politics: CXMT sits on the Pentagon's list of firms with alleged military ties, US lawmakers have signalled interest in restricting American purchases, and a bill moving through Congress would cut all three named recipients off from ASML sales and servicing by statute. Analysts note CXMT stays constrained by export controls and is unlikely to ease the near-term shortage. Addendum, verified Aug 3 (SemiAnalysis): the low-cost-flood fear has not materialised — CXMT's Q1 DRAM ASP ran only 5-10% below the big three while its per-bit DDR5 cost is >30% higher, leaving it no incentive for a price war at the expense of profit. A high-cost producer pricing near the leaders is a shortage participant, not a shortage breaker. Threat decrement logged.

What we concede, because it is real. A well-funded CXMT with domestic tooling is a genuine 2028+ threat to the conventional DRAM line — roughly ₩100T of our ₩357T model, not the HBM engine. One sell-side path has CXMT at ~18% share by 2028. And the Apple qualification matters more than the IPO: it is the difference between capacity and accepted capacity. Second-order effect worth stating: if China takes commodity share, SK's rational response is to shift more bits to HBM — which tightens HBM further. The threat compresses one lane while narrowing the other.

The rule, written Jul 13 and honoured today: ">20% unrecovered below the ₩2,180,000 stamp = ₩1,744,000 → the mechanical price rule." Seoul opened Jul 28 at ₩1,641,000, −9.64% — 24.7% below the stamp, 5.9% below the tripwire, and through the chartists' 200-period 4H EMA (₩1,730,887) that sat within 1% of it. Both systems broke the same line, as designed. Samsung −8.07%, SK Square −8.76%; Monday's US session had SKHY at $145, below the $149 IPO price for the first time. Your position: +66% vs entry (from +122% at the peak).

What AMBER means here, precisely: verify and tighten — not sell. The thesis engine is unchanged; what changed is that the price has crossed a line we pre-committed to respecting, and there is now a real new datapoint underneath it rather than pure flow (see the CXMT signal below). The verdict returns to GREEN when either the print clears the HBM4 question and price recovers above ₩1.744M, or the CXMT/lithography threat is bounded by evidence. It goes to RED only on a falsifier: a hyperscaler capex cut, two quarters of contract price declines, or an HBM4 shortfall confirmed by the company.

The timing is almost absurd: the Q2 print lands tomorrow (Jul 29, 9:00 Seoul), and Monday's US session already priced ADRs below the offer. Every open question in this thesis — the NH HBM4-shipment whisper, the LTA structure, the Microsoft deal report, the 2027 supply guide, and now the CXMT response — gets answered by the company in one session, roughly 24 hours after the tripwire fired. The ladder's hard date is the same day.

Asked directly whether the chip sector is due for a bust, Huang: "No, not for a while" — then, offered the phrase by his interviewer, he embraced it: "This time is different because this is not demand driven… this is industrially driven, meaning the fundamental technology of computers is changing." He argues the world needs a whole new infrastructure layer alongside energy, roads and the internet, and that the semiconductor industry "probably needs to be somewhere between five to 10 times larger… it is way too small today." The argument that matters most to us is his inversion of the glut risk: "We basically are constrained in every single direction, in every single way. That constraint is good. That constraint is what holds the system back. So that gives us plenty of time to go build out these infrastructure." Credibility split, per house rules. The observation — constrained everywhere — is not his opinion; it is independently verified by five physical sources this month: TSMC (CoWoS and N3 sold out into 2027), ASML (2027 book effectively full, memory segment +75%), SK's own board (₩7.09T to accelerate P&T7), Meritz (suppliers meeting 75-80% of DRAM demand, ~60% by 2027), and the ledger of GW projects whose binding constraint is power and construction labour. Score 9/10. The conclusion — therefore no bust for a while — is a forecast from the single most conflicted narrator in the industry, wrapped in the exact phrase that flagged 1999. Score 3/10 as prediction; logged as a top-marker in both directions, the same treatment given to Chey's "memory has permanently changed." And Fortune's framing is the honest context: chip stocks have sold off hard despite strong earnings, guidance and persistent shortages — which is precisely the fundamentals-versus-flows divergence this site has been documenting for two weeks. What we take: the constraint mechanism is a real, evidenced mitigant to bear row 7 (the 2029-31 capacity wave) — physical bottlenecks stretch the buildout over more years and lower the odds of a synchronized glut. What we refuse: treating "no bust for a while" as a reason to loosen a single falsifier.

NAVER / NVIDIA / Brookfield, $10B: the GAK Sejong DSX factory goes 55MW (1H-2027) → 200MW (2028) → 1GW long-term, running Vera Rubin and Blackwell. Structure matters: Brookfield up to $9B as exclusive capital partner (a nonbinding term sheet), NVIDIA $1B equity, Naver the rest — infrastructure capital, not operating cash, funding AI factories. The number that upgrades our model: 200MW will house about 100,000 GPUs — a hard density calibration (500 GPUs/MW) that our own SKT rack build-up (320/MW) was conservative against. It now anchors the new gigawatt ledger. Japan: Huang's Tokyo week produced a physical-AI coalition with Toyota, Fujitsu, Fanuc, Kawasaki Heavy and Kioxia — with Kawasaki, Fanuc and Yaskawa already on NVIDIA's stack — plus a 140MW Vera Rubin AI factory expected 2028. His framing: the next frontier of AI is the physical world, a once-in-a-lifetime opportunity for Japan. Aggregate: Korean companies have now closed close to $950B of AI partnerships with US firms (including Samsung-Broadcom ~$200B for memory and 2nm), and reporting adds two SK items to verify at the print: an SK hynix long-term memory supply deal with Microsoft and SK Telecom working with Anthropic. Two structural notes. First, NVIDIA's own segment data — datacentre $75B last quarter, +92% YoY; the edge/robot-training segment +29% — shows physical AI is real but still small: our Physical-AI adder card is sized correctly, and the near-term robot money is the training cloud, not onboard memory. Second, NVIDIA and Amkor are expanding advanced packaging in the US — the fifth distinct pipe-widener logged this month, and the one that most directly raises how many stacks 2027-28 can physically absorb.

Announced at the San Francisco AI Summit alongside President Lee, and confirmed on both companies' own newsrooms: a $500B-plus partnership spanning AI factory construction and AI memory supply, signed as Letters of Intent. Concrete pieces: SK Telecom builds a 2-gigawatt Vera Rubin DSX AI factory, first online 2027, powered by SK hynix HBM4, and NVIDIA and SK hynix establish a long-term partnership to secure and co-develop next-generation AI memory including HBM. Huang's own framing to Lee: over $500B "of business together." The arithmetic we can actually do: 2GW ÷ 1.25 PUE ≈ 1.6GW IT ÷ ~180kW/rack ≈ 8,900 VR NVL72 racks ≈ 640,000 Rubin packages × 288GB = ~184 petabytes of HBM ≈ $6.6B, or ~$10-11B of total SK memory content in that one factory — roughly 2% of the headline. The rest is NVIDIA silicon (~$40B), construction, land, power — and much of the flow runs from SK to NVIDIA. So: $500B implies ~15GW of eventual build or a decade-long horizon, and is not a revenue line. What genuinely changes: demand security (a co-development LTA with the customer that is ~half of world HBM demand), vertical integration — SK now owns the memory and the AI factory, which is the venue MaaS always lacked — and Korea as sovereign-AI anchor (AI Frontier Lab, KAIST Korean LLM, NVIDIA's 260,000-GPU pledge to Korea). Broader package: NVIDIA investing in Naver, Hyundai self-driving Genesis, LG power and robots, and chip/memory design partnerships spanning SK hynix and Samsung. Model treatment: zero incremental revenue booked. Confidence in the 2027-28 demand floor rises; the revenue lines do not move until definitive agreements and volumes print.

Alphabet's print delivered four new hard anchors for the model: (1) 2026 capex lifted to ~$195-205B (our bear-row-12 table carried the old $180-190B guide) with $811B of future purchase commitments disclosed — and the tape's reaction was the thesis in miniature: Big Tech stocks plunged on the spending while Asian chip stocks climbed, funds rotating into Micron and SK, Kospi back above 7,000. The customer's equity pays; the supplier collects. (2) Alphabet ≈ 7%+ of SK's revenue today (Bloomberg) — at our ₩357T 2026 line that's ~₩25-27T (~$17-18B) a year from one customer, and roughly 8-9% of Alphabet's capex flowing to SK alone. Triangulating the $811B: at ~15% memory intensity and SK's ~55-60% share, SK's slice of Alphabet's committed future spend runs ~$65-75B cumulative (~₩100-110T) over the commitment horizon — fat error bars, stated as such, but the order of magnitude is now anchored in a 10-Q, not a guess. (3) Long-term contracts heading to 25-30% of SK revenue in the next few years — a quarter-plus of the income statement de-cyclicalized by structure, the strongest quantification yet of the volatility-damping shift. (4) The leverage datapoint that outranks all of it: Alphabet, Microsoft and Meta offered to help fund SK's fabs and EUV in exchange for capacity — and SK refused the equity-like money, converting their urgency into better supply terms instead. The supplier that can decline the hyperscalers' checkbook holds the whip hand. Bear-row-12 status, updated honestly: the capex-cut tripwire not only didn't fire — capex was raised — while the FCF strain path (Kindig's >44%-growth line toward negative 2027) is confirmed on track: growth is running ~+115%. The race between monetization and financing fatigue is fully on; SK gets paid either way until a guide actually cuts.

The week's real signal is board money: at its 8th board meeting (Jul 22, all six independent directors present), SK approved an additional ₩7.09T for P&T7 — 37.3% of the plant's ₩19T total — explicitly "to accommodate increased production needs and in connection with the acceleration of the opening schedule of the cleanroom." Read the structure: P&T7's ₩19T is 95% of M15X's ₩20T — packaging capex now rivals wafer capex. The capital bottleneck has moved to exactly where our TSMC chain and ASML sub-linearity said it lives. And the timing is its own message: a company seeing demand softness does not accelerate a ₩19T packaging plant six days before it reports. This is the highest-ranking signal class in the house — dated, physical, board-resolved. The Intel decode: Tuesday the stock spiked +9.26% intraday to ₩2,006,000 — briefly through Zone 2 — on a report tying SK to Intel's Ohio site; SK denied it had "sought or made a decision on any such acquisition," and the gain evaporated (close −0.33%, ₩1,830,000; a ₩176K intraday swing). What the episode reveals: the market wants SK to buy US wafer capacity — Kwak already has a US fab "under consideration" on the record — and will pay ~9% on the rumor alone. File under optionality, not event. The rest of the week: foreigners net-buying three straight days; Hanwha's PT at ₩4.3M (from ₩1.63M — "transforming into one capable of consistently generating high levels of profit"); SWS analyst FVs migrating ₩2.71M→₩3.13M; HBM4E samples now shipping to major AI customers; and the company on record expecting DRAM constraints to persist to 2030 while signing long-term contracts "aimed at reducing memory market volatility" — the de-cyclicalization structure, in their own words. ADR recovered to ~$165.

The deep-dive on Vera Rubin NVL72's first benchmarks (CoreWeave engineering rack): 5.4× performance per MW and 5× per dollar vs GB200 (2025 baseline); against the current GB300, ~2× at low speeds widening to ~4× at 200 tok/s/user — and above 300 tok/s/user only Rubin can serve at all. Where does the gain come from? SemiAnalysis says it plainly: Rubin's advantages "lie in a higher HBM capacity, higher CPU DRAM capacity, and greater HBM bandwidth" — 2.8× the memory bandwidth via 3D-stacked HBM4 — making it the machine for multi-trillion-parameter models (Fable 5, Gemini, Kimi K3). Translation: "fast mode" — the agentic speed tier the market is moving to — is a memory product, and SK sells it. The TCO table completes the loop: Rubin costs more per GPU-hour ($3.57 vs $1.84-2.36) — much of that the fatter HBM bill — yet lands cheaper per token at every speed. Customers will rationally pay the memory premium; that is HBM pricing power, written in a spreadsheet. Ramp risk falls too: a cableless tray design plus Blackwell-kernel reuse (SM100 runs on SM107) means faster production ramp and no software air-pocket — HBM4 pull-through accelerates through 2H26-27. The honest new item — LUT B, logged into the efficiency-sponge family: Rubin's 3-bit lookup-table weight format stores weights at 3.125 bits — their own Kimi-K3 worked example: 1,487GB → 1,094GB of weights, ~4 HBM packages instead of ~6. A real weight-capacity reducer, in silicon. The offsets, also theirs: no accuracy data published, CoreWeave didn't even enable it, KV-cache and activations (the part exploding with agents and long context) are untouched — and fewer bytes per weight means faster decode, more tokens served, more usage: the sponge feeds the rocket again. Feynman (SM_140) sneak peek: a genuinely new architecture with AMD-style 3D stacking — 2028's watch item. And InferenceX gets Nvidia-verified Rubin numbers plus Google TPUv7 and AMD MI455X in Q3 — objective cross-platform memory-intensity telemetry incoming.

The date, from the source: SK hynix's own SEC 6-K and DART disclosure schedule the Q2 call for Jul 29, 9:00 AM Seoul — the same day the 17.79M new shares list on KRX. One day now carries the earnings print, the HBM4-whisper verdict, the new-share conversion test, and the ladder's hard date. Aggregators showing Jul 22 were stale; the calendar error here is corrected and logged. The premium is deflating in an orderly way: from ~38% at the feral peak to ~22% — 41% of it gone in two sessions — with the ADS at ~$152, just 2.2% above the $149 offer. The falsifier (<3% premium) remains far away; what's dying is the leveraged froth, not the structure. Valuation marker worth framing: per Barron's, SKHY now trades at 5.71× forward vs Micron's 5.93× — the HBM leader priced below the number-three player. New shortage quantification (Meritz, Kim Sunwoo): suppliers are meeting only 75-80% of DRAM demand today, falling to ~60% in 2027 — a 40% shortfall, the hardest number yet put on Kwak's "worst year" call. Chey, Friday: "Memory chips will continue to be needed, so their value will trend upward over time." And from the IPO filings, a number for our ASML card: ₩45.5T of Korean facility capex plus ₩11.9T specifically for EUV scanners through 2030, on a disciplined mid-30% capex-to-sales ratio — SK's slice of Veldhoven's order book, in won. Position markers from Q1: HBM share 56.4% (#1), DRAM 29.1% (#2), NAND 18.5% (#2); Q1 gross margin 79.3%.

The verification matters more than the headline. The circulating analysis assumes Google bakes Gemini's weights into silicon, eliminating weight-storage memory. The Information's actual reporting: the design shifted away from the earlier "fully hardwired" approach that baked in model weights to a "flexible hardwired" strategy that hardwires the architecture — precisely because frozen weights die with every model update. Google hit the wall our physics predicts: frontier weights (hundreds of GB to TB) cannot fit in silicon — even a 10B-param model at 4-bit needs ~2,000mm² of SRAM, multiple reticles, just for weights — and models now iterate monthly (K3 shipped last week). So Frozen v2 is an architecture-ASIC: 6-10× tokens-per-watt vs the latest TPUs, deploying ~2028, with weights still loaded from external memory. The memory implications, corrected: (1) the memory wall gets relatively taller — when compute overhead shrinks 6-10×, memory traffic becomes a larger share of each token's cost and power; MS's memory-wall thesis, cast in silicon. (2) Jevons, again: Google is building this because its compute shortage is so severe that Google Cloud is rejecting external customer orders — 6-10× efficiency means they serve 6-10×+ more tokens, each still reading weights and KV from memory. Total bits go up. (3) The chip threatens Nvidia's and TPU's share of Google inference — an intra-compute-layer fight; memory is the arms dealer on every side (TPUs already carry SK HBM via Broadcom). The surviving headwind kernel, honestly bounded: if future Frozen generations revert to weight-baking for the nano/distilled tier, that tier's weight-storage demand shrinks — but that's the tier with the least memory content per unit anyway, and it joins the same family as Qualcomm HBC, HBF, and TurboQuant compression (which spooked memory stocks in March). Threat: 2/10, bounded by mask economics and model velocity. Tape note: TurboQuant's March precedent says headlines like this can hit memory stocks on the wrong first read — same DeepSeek-echo pattern flagged on K3.

Moonshot's Kimi K3 (Jul 16): 2.8 trillion parameters — first open-weight model in the 3T class — 1M-token context, native vision, weights public Jul 27. Beat Fable 5 and GPT-5.6 Sol on Frontend Code Arena, third on the Artificial Analysis index. Baker's viral read (he's 13F-long SK): "negative for Anthropic and OpenAI while net positive for essentially every other company in the world. I mean that very literally." His mechanism is the margin waterfall: 2-3 closed labs at ~90% inference margins trap the economics at the model layer; an open-weight model needs identical compute to run, so "every dollar that isn't captured as model-layer margin flows downstream instead" — to power, chips, datacenters, and memory. That's our layer. Then reality staged the demonstration: within 48 hours demand pushed Moonshot's GPUs to their limit and the company paused new subscriptions ("Our GPUs are feeling it"). The Decoder's line deserves framing: "So much for the idea that open source cuts computing needs." The memory math makes it ours specifically: K3's MoE keeps all 896 experts resident — ~1.4TB of weights even quantized — and serving takes ~8× H100/H200-class GPUs ≈ 1.1TB of HBM per instance, before the 1M-token KV caches. K3 may be the most HBM-intensive artifact ever handed to the public — and from Jul 27 anyone can self-host it. Baker's honest nuance, kept: K3 is token-inefficient (50-70% costlier to run than GPT-5.6 per Artificial Analysis) — the true "Sputnik moment" would be open and efficient. For memory, even that cuts our way: inefficient models that still get adopted consume more bits per unit of intelligence. This is live evidence for Scenario D's Jevons line — cheap access → usage explosion → aggregate demand up. One caution, logged: DeepSeek's Jan-2025 release produced a −17% Nvidia day on the wrong first read before demand exploded; a K3-echo tape scare into this fragile tape is possible.

The uncomfortable identity: SK's revenue is its customers' capex — and the capex is now outrunning the cash that funds it. The printed numbers: Amazon is already FCF-negative (Q1 −$17.2B; LTM FCF $1.2B, −95%; capex +78.5% vs OCF +53%; debt +166% to ~$154B). Microsoft guides to roughly breakeven FCF this year ($190B capex vs OCF barely ahead). Meta is on the brink if spend scales abruptly (debt +190%). Google holds out longest (net cash $90B) but raised $20B of debt plus an $84.75B equity raise, and goes negative in 2027-28 if capex grows >44% then >18%. Combined 2026 guides: ~$710B across four companies — which, noted for the record, already clears the $700B/yr telemetry bar our Vahdat prediction row set for mid-2027. The regime change buried in the piece: capex is no longer funded from cash flow but from capital markets (~$200B of raises announced in months). That converts the binding constraint from cash generation to investor willingness — echoing the 1999-2001 telecom structure, where the market's mood, not demand, ended the buildout and took suppliers down 90%. The differences that matter, stated honestly: these customers each generate $100B+/yr of growing OCF from profitable core businesses (telecoms didn't); AI monetization is a lag, not an absence (Google Cloud +63%, MSFT AI ARR $37B, AWS AI ~$35B ARR, Meta Advantage+ $60B+); net cash is still positive at all four; the buyer base is broadening to sovereigns who aren't FCF-constrained (Japan, Gulf); and HBM is the last line item cut — it's the binding input to the inference revenue they're all chasing, and much of SK's 2026-27 output is pre-contracted with customer prepayments. Sequence if it breaks: raises get punished → guides flatten (tripwire #1) → SK's 2028-29 book softens into the capacity wave = the glut case. That's a 2027-28 transmission risk with clean telemetry, not a 2026 event. Closing irony, on the record: I/O Fund's own answer to this risk is to own the suppliers receiving the capex — memory included. The author's hedge is this portfolio.

The morning-TV hit (Jul 15, CNBC International): Lee dismissed Korea bear-market fears outright, calling the five-to-six-week momentum downswing "temporary breathing room within a multi-year AI expansion" and Monday's Seoul plunge "a healthy pause following extraordinary year-to-date gains." The two quotes that matter: the cycle is a "massive massive industrial build that is going to last for years" — and "all those dips are buyable." This caps a consistent run: Jul 8-9 he named Samsung and SK Hynix as buy-the-dip with HBM the category he's most bullish on ("the sell-off is cyclical and sentiment-driven; the demand is structural and accelerating"); on crash-day Jul 13 he reiterated S&P 8,000 by year-end while discussing the SKHY ADR. His Jul 16 webinar frame: June CPI soft, 68% of the CPI basket deflating from peak, playbook stays buy-the-dips with gains back-ended into year-end. (The refreshed Top 5 list itself sits behind Fundstrat's email gate — whether SKHY is on it is unconfirmed.) House rules apply: street voices are lagging confirmation, never signal — but Lee moves retail flows, and his repeated on-air naming of SK through the worst week in its history is flow-relevant. The honest counterweight he supplied himself: he expects Aug-Oct to "feel like a bear market" (four triggers incl. the SpaceX lockup expiry; margin debt flashing a consolidation signal for later 2026). That window overlaps our Aug 26 Nvidia/Kyber test and the leveraged-ETF fragility — logged as a tape-risk overlay, not a thesis item.

The desk that was bearish into the ADR now writes the bull case in structural language: AI's core constraint is shifting from a compute wall to a memory wall. The number that is the wall: DDR5 bandwidth grows +14% over two years (44.8 → 51.2 GB/s) while monthly inference tokens grow ~320× (10T → 3,200T). Memory/storage now runs to 73% of a CPU server's bill of materials; DRAM sits near 30-year price highs after Q1 contracts jumped ~90%+; MS just raised its Q3 forecasts again (server DRAM +13-18% QoQ, PC DRAM +15-20%, eSSD +18-23%). Storage's share of cloud capex heads from 12% → 40%, cloud storage spend to $418B by 2030, servers to 59% of all DRAM demand by 2028 (from 37% in '23). Citrini's supply math: a 28.7-exabyte DRAM gap by 2030 — ~18% of demand — even if every announced fab, Chinese ones included, comes online; conventional DRAM ASPs $1.5-2/Gb "for years." MS names our favorite: "We expect a steeper upcycle with rapid gains in DRAM, HBM, NAND, and legacy memory," favoring Micron, Samsung and SK. Per the house rules this is lagging confirmation, never signal — the physical layer said it first — but confirmation at this size moves flows. The honest counterweights, logged: CXMT's ~$9.8B Shanghai IPO to fund expansion (share 8%, doubled in a year; could grow output two-thirds by 2028 — bear row live), MS's own "blind-buying phase is fading" caution with consumer OEMs cutting orders at the price ceiling, and the software-efficiency sponge (KV-cache compression, quantization) that MS itself flagged in March.

TSMC: record quarter — $39.6B revenue, +36% YoY, above its own guidance high-end; June alone +67.9% YoY, breaking a four-year seasonal pattern (AI has rewritten the calendar). The parts that extrapolate straight onto SK: CoWoS and N3 sold out through end-2026 with lead times into 2027; CoWoS capacity tripling-plus from ~35K wpm (end-'24) toward 125-130K wpm by end-2026; the packaging supply-demand gap narrowing 20% → ~10% — a fifth pipe-widener; capex $52-56B (largest ever, +37%), spending at the high end. ASML: raised full-year guidance a second time to €43-45B (from €36-40B at the start of the year), guided Q3 to €11-12B, and — the line that matters most here — memory-segment revenue set to grow 75% in 2026, with EUV capacity +30% planned for 2027, another +30% under study for 2028, and 2027 already close to fully ordered. CFO Dassen: DRAM is using more EUV layers per node — tool spend converts to bits sub-linearly, a physics-enforced floor under pricing. The extrapolation chain: every CoWoS wafer exists to marry logic to HBM — at the 2026 run-rate (~1.2M wafers/yr average, ~15 packages/wafer, ~6-7 stacks/package) TSMC's packaging math independently implies a ~$100-120B HBM market this year, landing within ~20% of the engine's number by a completely different physical route. Push the same chain into 2027 — CoWoS units +60-80%, GB-per-stack +30-50% (HBM4, 16-hi), $/Gb +25-100% (Goldman) — and HBM dollar demand roughly doubles, which is Scenario B's exact shape and the Nvidia-chain card's ₩130-160T SK HBM output. What today's prints do: physically underwrite the demand side of Jul 29. What they don't do: answer NH's SK-specific HBM4-ramp whisper — that's execution, and only SK's own print clears it.

The full tape since Monday's record fall: Tuesday Seoul +8% (foreigners net-bought ₩2.33T, buy-side sidecars fired, close ~₩1.99M — exactly retesting Zone 2 from below); ADR ripped +27% to $193.92 on thin-float + leveraged-fund demand. Wednesday ADR gave back 5% to ~$184.50 in group-wide profit-taking. Thursday Seoul −10.95% to ~₩1.842M, KOSPI −5% triggering the year's 37th sidecar, now a technical bear market — a follow-on from the US overnight chip rout (MU −8%), not Korean news. The stated causes are all structural: crowding (semis = 20% of the S&P 500, vs ~8% at the dot-com peak — Kondratev), AI-spend nerves, and the leveraged wrappers — four SKHY single-stock ETFs now trade (SKUU/SKDD/SKHL/SKHU), and Korean regulators met today specifically to review single-stock leveraged ETFs: the amplifier is officially on the regulator's desk. Meanwhile the physical layer said the opposite out loud: ASML raised full-year guidance a second time (€43-45B) and is ramping EUV output, and the same CNBC piece notes HBM demand still exceeds supply with pricing power intact. The ADR premium has gone feral — implied ~30-40% over Seoul on a 2.4%-of-company float chased by leveraged products; two prices of one company pulling apart. Your Frankfurt line tracks Seoul.

The Tokyo swing, sorted by what it means for memory: (1) Physical AI = a second-stage demand rocket. Robots pull memory twice — onboard (Jetson Thor carries 128GB of unified LPDDR per robot; Huang calls humanoids "a multitrillion-dollar opportunity") and in the cloud, where every robot fleet needs continuous training and simulation, which is straight HBM demand now. The onboard lane is SOCAMM-shaped: SK's second lane again. (2) Sovereign AI broadens the base. SoftBank (GB200 NVL72), KDDI, Sakura (~10,800 GPUs), GMO et al., METI-backed — every GB200 rack lands with ~20TB of HBM, and national buildouts dilute the top-5-hyperscaler concentration that is 60% of Nvidia's revenue. (3) Nikkei: Nvidia and Mitsubishi Heavy exploring cooling + energy-management partnership for AI factories — Nvidia recruiting heavy industry to widen the deployment pipe below the chip. Fourth data point in the pattern; every pipe-widener raises how many racks (and bits) 2027-28 can physically absorb. (4) Jensen reaffirmed Vera Rubin "progressing as planned" (third reaffirmation — consistent with our chip-ships/rack-changed adjudication) and teased a possible future Rapidus opportunity — a years-out watch item that would dilute TSMC concentration risk if it ever hardens. Nothing here changes numbers this quarter; everything here confirms the shape: demand widening, deployment pipe being industrialized, memory the constant.

Aschenbrenner (Situational Awareness, ~$20B AUM): SK Hynix was already ~6.5% of the fund's disclosed book before the ADR — then he cornerstoned more in the offering alongside Baillie Gifford (an SK holder since 2000) and Coatue: up to $7B indicated, ~$5B allocated, ~25% of the deal. The fund's edge is exactly this thesis — early IREN, CoreWeave, Bloom Energy on the "AI infrastructure bottleneck" read; backers include Nat Friedman and Daniel Gross. The order book behind them: ~$171.5B, 7×+ covered, from long-onlies, sovereigns and Asia specialists. Street echo: Daishin's Ryu lifted his PT ₩3.4M → ₩3.9M ("the valuation discount will quickly narrow"). No post-crash statements from Aschenbrenner, Gerstner or Baker have printed — and none should be expected: disclosure mechanics mean the Aug 14 Q2 13F shows only Jun 30 books (pre-ADR); the cornerstone buys and any crash-week adds won't print until the Q3 13F on Nov 14. Between now and then, conviction is only visible through what they already did: buying 25% of a $26.5B deal. DART: no new insider transactions filed since the listing. Moonshots (Diamandis/Blundin, eps late-Jun→Jul 11): nothing SK-specific, but the ambient read is pure Scenario D — a June episode literally titled around "Recursive Self-Improvement Arrives", US government gating frontier model releases, and Blundin's standing line: "there's just not enough energy, compute, infrastructure, anything." The people modeling the singularity and the people writing $5B cornerstone checks are converging on the same object: the bottleneck.

Seoul closed ≈₩1,845,000 (−15.4%, largest one-day fall on LSEG records), dragging KOSPI −9% and tripping the year's seventh circuit breaker. SKHY fell 7.9-9.3% to ~$152-155 — a few dollars above the $149 offer. MU −6.4%, SNDK −8.4%, SOX −3.6%. SK Hynix issued no profit warning and no adverse operating disclosure. Five forces, none of them new physics: (1) sell-the-news profit-taking after a 3× year — bear-file row 1, which had this exact window flagged AMBER-live Jul 10-30, fired on schedule; (2) benchmark confusion — Yuanta's Yoo: "everybody's really confused about… where the fair price is" now that a 25%-premium ADR anchors the same asset (TSMC steady-state is 13-14%); (3) 2.5% fresh share supply from the raise; (4) the one real fundamental whisper — NH's Ryu: the Q2 HBM4 shipment increase investors expected "does not appear to have materialised" — unverified, single analyst, resolves Jul 29; (5) the leveraged-ETF amplifier — Korean single-stock leveraged products hit 84% of KRX volume with ~$9B of mechanical rebalancing per 1% move, and GraniteShares launched SKUU (2×) / SKDD (−2×) on SKHY today. The deep-structure card wrote this mechanism down on Jul 12: leveraged wrappers don't change the organism, they amplify its variance — sell-offs overshoot. It overshot. Bulls on record same day: Yoo (temporary; "right direction" over 6-12 months), Blancato ("not the end of the run… demand into late 2027, early 2028"), Wool (rebalancing, not deterioration).

What it is: Qualcomm puts its inference accelerator (XPU) underneath an LPDDR stack, TSV-connected — near-memory compute. Claims: 133 TB/s per AI250 card (18× its LPDDR5X-based AI200), 6× bandwidth-per-watt vs HBM, 200× capacity-per-watt vs SRAM, 768GB theoretical. Gen1 ships mid-2027 in the AI250; Gen2 2028. The physics audit: (1) the headline number is apples-to-oranges — 133 TB/s is internal bandwidth between a stack and the die glued to it, not an external interface feeding a separate GPU like HBM4's 3.3 TB/s per stack. When a claim needs a redefined denominator, discount it. (2) Thermal ceiling is structural: DRAM refresh degrades on top of a hot die — that's why HBM sits beside the GPU. The compute die under the stack must stay low-power, which caps the compute it exists to feed → many-small-units inference architecture, not a Rubin competitor. (3) Segment: cloud inference TCO tier only; training and frontier inference stay HBM. Qualcomm's accelerator share today rounds to ~1%. (4) The kicker — someone must supply the LPDDR, and the candidates are Samsung, Micron… and SK, the LPDDR leader already shipping SOCAMM to Nvidia. Same two-lane structure as pooled memory: if inference tilts LPDDR, SK sells LPDDR stacks instead of HBM stacks — a margin-mix question, not a revenue loss. (5) It joins a crowded family — SOCAMM, Apple unified memory, Intel stacked DRAM, and HBF (which SK is inside). Everyone attacking the memory wall is demand-side proof the wall is real, announced during what Kwak calls the worst supply year coming.

SemiAnalysis (Jul 6): the Kyber NVL144 rack slips >12 months to 2028 — its 78-layer PCB midplane (three 26-layer boards laminated into one) can't be made at production yields. Nvidia's four-word reply: "Our roadmap is intact." Then Huang went to Morgan Stanley's roadshow and denied the delay — while MS's Moore simultaneously confirmed "original Kyber rack plans are being replaced by better solutions." Resolution: the chip ships, the rack changed. Rubin Ultra silicon stays 2027 (Jensen's claim, credible — he owns the schedule); the Kyber rack-as-designed is dead-or-2028 (SemiAnalysis's claim, credible — physics doesn't negotiate). This is the third instance of the same pattern: CoWoS-L warping killed the 4-die package, T-glass gates substrates, now PCB lamination kills the backplane — the deployment pipe below the chip keeps breaking, never the silicon, never the demand. For us: neutral-to-positive. Rubin Ultra carries HBM4E either way; a cable-heavy fallback rack needs more copper and materials (Mitsui lane); and SemiAnalysis itself projects Nvidia DC revenue 20% above consensus in 2H FY27. The demand context from the same roadshow: quarterly revenue approaching $100B (Q2 guide ~$91B), FY27 growth +52% (MS, PT $288; street to $500 at Baird), cumulative Blackwell+Rubin orders ≥$1T through 2027, and AI labs that were mostly ASIC-based now running ~50% Nvidia. And Moore's closing sentence is our thesis from the demand side, verbatim: Nvidia's challenge is "how to translate demand into deliverable system revenue under multiple constraints such as memory, networking, power and data center space." Memory. Named. First.

Chairman Chey, at the bell: "We could be memory servicers, memory as a service. In the future, that is another area where we could actually focus." The concept: stop selling chips once; rent memory capacity + optimization software the way Amazon rents compute — tied to SK Group's ~$1T, 15-20GW AI-datacenter build. Why it matters structurally: the 10× Micron-ceiling exists because memory is cyclical hardware. A real service layer converts boom-bust chip sales into recurring revenue — cloud-infrastructure businesses trade at 20-30×, not 10×. And the technical substrate already exists in our ledger: optically-pooled memory (COUPE/CXL) makes memory disaggregatable, and disaggregated memory is rentable — MaaS is the business-model wrapper on the pooled-memory topology, with SOCAMM2 + custom HBM + "special software" (Chey's words) as the stack. The honest scoring: Chey himself calls it "an idea at this point" — no pricing, no scope, no timeline. Execution credibility today: 3/10. Strategic-intent signal: 8/10. Also the flip side, logged: a service model moves utilization risk onto SK's own balance sheet — in an oversupply, rented memory sits idle at SK's cost, and it part-competes with the hyperscalers who are its biggest customers. NOT in the engine until it has a price and a customer — but priceable as an option: ~35% probability of concrete form within 12 months, +15-25% pop if announced ≈ +6-9% expected value embedded today. ROIC audit required the day terms land.

Launched the same day as the ADR, three tokenized versions went live: Backpack's $SKHY (redeemable 1:1 into an actual share entitlement via ACATS — the strongest structure), Kraken/xStocks' $SKHYx (cash-settled claim, multi-chain: Solana, Ethereum, and TON inside Telegram's wallet), and Ondo's $SKHYon (total-return, dividends auto-reinvested, cross-chain via LayerZero to BNB). The stock is now tradable 24/7, globally, without a brokerage. Honest sizing: combined day-one token volume was $1.18M — a rounding error against a $1.27T company. This is a distribution-rails signal, not a flow event: Solana's tokenized-equity segment grew 6× in H1 with 262K holders, and the SpaceX precedent hit 13K token holders in a week. What it does change today: price discovery never sleeps — weekend token prints will now leak into Monday opens, and the share becomes usable as DeFi collateral. Risks logged: three issuers fragment liquidity; xStocks' cash-claim model is the FTX-era design (Backpack's CEO's critique, on the record); a weekend depeg on a thin venue makes headlines even when it means nothing. Watch metric: token AUM (not volume) — and the conservation law that matters: every token minted is a custodied ADR pulled from tradable float. A slow float sponge, plus a new instrument: weekend token prices now lead Monday opens — tracked as a prediction below.

Open $170 (+14.1% vs the $149 offer) → high $174.45 → close $168.01 on ~88M ADRs; market cap $1.27T, #11 in the US (below Tesla, above Eli Lilly). The pop beat SpaceX's debut (+11%). Cornerstones (Baillie Gifford, Coatue, Situational Awareness) received ~$5B of allocation. The number that matters: $168.01 ×10 = $1,680/share vs Korean parity ~$1,444 = a ~16% premium — exactly TSMC's steady-state ADR premium, formed on day one. Conversion restrictions mean it can persist (UBS's call, vindicated). Seoul closed flat ₩2,180,000 Friday before the US session: the ADR close implies ~₩2.53M — Monday's KRX open is the first chance for the local line to react. CNBC's desk framing for US investors: "SK Hynix is bigger, cheaper and closer to NVIDIA," with ~$14B of passive index buying estimated to follow. Mechanics: SKHYV → SKHY regular-way from Mon Jul 13 (CNBC says the switch shows Tuesday); offering settles Jul 14. Leveraged ETFs (SKHX 2×, SKHZ −1×, Direxion SKHL filed) launch Jul 13-14 — volatility amplifiers, both directions.

Three claims, each falsifiable: (1) 2027 = worst-ever supply shortage in memory history; (2) "customer demand will remain higher than our supply capacity even beyond 2030"; (3) customers sign LTAs "because they believe the shortage will last for longer." New disclosure: a US wafer fab is under consideration (US vs Japan vs Southeast Asia — needs land, power, water, skilled labour at competitive cost; nothing decided). Chairman Chey to CNBC: told customers capacity would double in five years, "all my customers said, that's not enough, man, we need more"; AI agents and physical robots "need a lot of memory chips"; also floated future additional US share issuance once the price is stable. Credibility read: the 2027 claim scores 8/10 — dated, superlative, mechanism given, corroborated by customer behaviour (LTAs with floors and prepayments are revealed belief, and per Tom's Hardware some customers have offered to buy SK's EUV machines and fund fab lines outright). The "beyond 2030" horizon scores 6/10 — consistent with Chey's March statement but too far to hold anyone to. One honest flag: Chey's "memory has permanently changed from boom-bust" is exactly the sentence that appears near cycle tops — we log it as both thesis-confirmation and top-marker, and let the 2029-31 fork discipline handle it.

Brad Gerstner (Altimeter): sold the firm's entire Microsoft position to fund Nvidia and SK Hynix — "you have to make choices in this market" — with 80% of Altimeter's capital now in memory, logic and compute, and hardware demand framed as a durable trend through 2029, not boom-bust. Gavin Baker (Atreides): "HBM DRAM will be 30-40% of ALL hyperscaler capex in 2027"; only three companies on earth can stack 12-16 dies; no fourth supplier arrives in 2027; new supply agreements lock floor pricing above prior-cycle gross-margin peaks; his 13F added SK Hynix, SanDisk and Micron in late 2025 (his Micron call is up ~14×). Add the ADR anchors — Baillie Gifford, Coatue, Aschenbrenner's Situational Awareness — and the picture is complete: the most-followed AI investors in America are concentrated in your exact position. Comfort and caution in one fact: consensus among the smart money means the easy mispricing is closing — which is the re-rate you own it for.

$26.5B raised (177.9M ADRs; 10 = 1 share) — the largest foreign listing in US history: past Alibaba (~$25B) and Saudi Aramco ($25.6B), second only to SpaceX ($85.7B) among all listings. Book 7×+ covered (~$170-200B of demand); anchors Baillie Gifford, Coatue and Situational Awareness took up to $7B (~25% of the deal); orders from $200M with several above $1B. Priced above spot into a KOSPI bear market — institutions paid up while the tape was fearful. Proceeds → Yongin fab 1 (₩31T), P&T7 packaging (₩19T), ASML EUV scanners (₩12T). Jensen Huang, June: SK stays Nvidia's largest memory partner; the shortage "will persist for a few years." Mechanics note: Friday trades when-issued (SKHYV→SKHY); regular-way settles from Mon Jul 13 — the first print is still tonight.

TSMC's COUPE co-packaged optics: 500 wafers/mo today → 10K by Q2'26 → 15K by Q4'26 → ≥25K by 2028; NVIDIA, Broadcom, AMD queued as first customers. Plain words: light replaces copper between chips, so clusters can move data far faster. Faster pipes don't reduce memory — they expose the memory bottleneck harder (a fed GPU wants more HBM, not less), and optically-pooled memory racks add a second demand lane. The nuance the analysts miss: some pooled bits may shift from HBM to cheaper LPDDR-class memory — which is exactly what SOCAMM2 is. We supply both sides of that fork. Reality check (MS): 2026 CPO lands at ~23K units vs 200K once hoped, yields 20-50% — this is a 2027-30 lane, not a today event.

The new LTA structure, in plain words: customers get a guaranteed supply and SK gets a guaranteed minimum price (floor) — but there is no maximum. When shortage pushes spot prices up, SK's contract prices ride all the way up with them. Add 10-30% prepayments and 3-5 year terms (Google: 5yr commodity DRAM with a +2yr extension tied to next-gen HBM; Microsoft: multi-year DDR5 worth tens of trillions of won). SK may be the only major supplier with no cap — Micron caps its new deals at Q2-2026 market levels. Context that makes it bite: conventional DRAM contracts rose 90-95% in Q1 and another 58-63% in Q2, and Goldman just tripled its 2027 HBM price forecast.

Reports Meta will sell surplus compute externally (i.e. enter the cloud business) sparked "AI overbuild / peak demand" fears: Kospi −10% over two sessions, our stock −14.6% Thursday to ₩2.19M, then +10% Friday. Counter-facts, same week: Korea's June exports crossed $100B for the first time with semiconductor shipments nearly tripling (physical > narrative); SemiAnalysis's 50-enterprise survey found no 2H26 budget risk; Meta reselling compute is a revenue move, not a capex cut. Amplifiers were local: Samsung+SK = ~half the Kospi now, June inflation 3.2%, mechanical foreign selling ($62B YTD; Goldman simultaneously raised Kospi target to 12,000).

₩80T for the M17 NAND fab (construction 2027, operations H1 2029) + ₩20T for Cheongju P&T7 packaging (late 2027). SK Group AI datacenters scaling 5GW → 15GW. His words: demand "soaring… as AI services take off" — said on the day the stock fell 14%. Demand conviction from the person who sees the order book.

Samsung + SK Hynix to build four new fabs under a state semiconductor program. State-funded capacity gets built even if prices fall — the 2029-31 capacity-doubling danger window is now national policy, with M17 switching on right at its start. Near-term bullish (funding certainty), long-term the single biggest input to Scenario C.

17.79M new shares (~2.5% dilution) → up to ₩45.45T. Reference started at ₩2,555,000/share, was re-based to ₩242,500/ADR off the Jul 3 close, and priced $149 on Jul 9 — done. 10 ADRs = 1 share. Underwriters: BofA, Citi, Goldman, JPMorgan + 9 others. Proceeds: Yongin Y1 fab (₩31T), Cheongju P&T7 packaging (₩19T), EUV machines (₩12T). Ticker SKHY confirmed (SKHYV when-issued through Jul 13). New shares list on KRX July 29; settlement Jul 14. This signal is now history — kept for the record.

Q2 consensus revenue ₩82.46T (Q1 was ₩52.58T). A beat on the same day 17.79M new shares hit the Korean market = maximum volatility clustering. A miss into fresh supply would be the worst single-day setup of the quarter.

TSMC's CoWoS-L packaging warps under a 4-die monster chip — dies lose contact with the substrate. Nvidia retreats to a proven 2-die design. Open question that matters to us: 16 HBM4E stacks per package → 8. Bear read: halved memory per halo chip. Neutral read: 2+2 board-level assembly + more units per rack recovers most of it. Standard Rubin (8 stacks, ships this summer) untouched — that's the volume driver.

China's CXMT signs a 3-5 year server-DRAM supply deal ahead of its IPO. CXMT went 3% → 8% of global DRAM in a year — but still far behind on DDR5 yields and has no HBM. Apple reportedly lobbying to source from CXMT too. This nibbles the commodity segment (~30% of SK revenue), not the moat.

50+ enterprise conversations: budget caps are companies growing up, not cutting back. Top-decile customers (most of the revenue) at little risk; coding → cyber → knowledge-work waves still loading. Explicit call: no material risk to 2H26 AI budgets. Separately, OpenAI halving inference cost is the Jevons pattern — cheaper AI → more AI → more memory. Efficiency headlines are noise for L1.

Open-weights models + agent harnesses now rival frontier labs on hard tasks at a fraction of the cost. That squeezes model-layer margins (Anthropic/OpenAI pricing power) — not silicon. SK sells HBM to whoever wins. Agent harnesses running multi-agent jobs for 11-20 hours with 1M-token contexts are memory multipliers. Only channel to us: if model-margin compression ever dents hyperscaler capex. Watch capex guidance, not benchmarks.

Korea's financial regulator publicly criticised leveraged ETFs tracking SK Hynix/Samsung. Stock +300% YTD, +920% in 12 months, retail leverage products proliferating, 12.5% single-day swings. None of this changes fundamentals — all of it says positioning is hot and air-pockets are possible.

Samsung shipping HBM4 and running a "Super-Gap" roadmap — parity risk unchanged, SK still a full generation ahead (HBM4E sampled). New angle: overseas-investor inquiries about a Samsung ADR surged after SK's filing. Relevant because Samsung is your designated first diversification target post-re-rate — a Samsung ADR would make that trade easier and could give it the same re-rate kicker.