Aya Research
DEEP RESEARCH
中EN

SK hynix: The Company Micron Almost Bought, the Bet Nobody Copied, and the Largest ADR in History

Panoramic research report · longitudinal history + cross-sectional rivalry + synthesis Subject: SK hynix Inc. (KRX: 000660; NASDAQ ADR "SKHY" expected to list 2026-07-10) Report date: 2026-07-04 · Data cutoff: Q1 2026 results (reported 2026-04-23) + latest quote, filings and news flow through 2026-07-04 Sources: company IR releases and F-1 filing, Yahoo Finance, CNBC, Bloomberg, Reuters, TrendForce, Counterpoint Research, SemiAnalysis, The Elec, Seoul Economic Daily, KED Global, Tom's Hardware, EETimes archives, CRS reports For information and research purposes only. Not investment advice.

Before we start

Six days from now, if the calendar holds, SK hynix will price the largest American Depositary Receipt offering in the history of capital markets: 17.79 million new shares, roughly ₩45.45 trillion — about $29 billion — listed on Nasdaq under the ticker SKHY. It would surpass Alibaba's $21.8 billion New York debut in 2014, the record for twelve years.

The company selling that stock is worth about $1.12 trillion, earned a 72% operating margin last quarter, and holds a claim on roughly two-thirds of the high-bandwidth memory going into NVIDIA's next GPU platform. It is, by most operating measures, the single most successful company in the AI supply chain not named NVIDIA.

It is also a company that traded at 125 won in 2001, that survived on a $7 billion creditor bailout, that its own board saved from being absorbed into Micron by a single overnight veto in 2002, and that was bought in 2012 — for less than one-thousandth of its current market value — by a telecom conglomerate everyone said was making a mistake.

Between those two facts sits every question that matters about the AI trade in July 2026. This report is organized in three parts. Part One is longitudinal: how a bailout ward became the throne. Part Two is cross-sectional: the memory board as it stands this week — the share, the contracts, the challengers, the China clock. Part Three is synthesis: the financial snapshot, what the $29 billion listing itself tells you as an analytic object, scenarios, the valuation paradox, and the falsifiable claims on which the whole thing rests.

One housekeeping note: SK hynix reports in Korean won. We give dollar translations at roughly ₩1,480–1,540 per dollar where useful; the won has been trading at 28-year lows above 1,500, which is itself part of the story.

For readers of the earlier reports in this series, the through-lines carry over deliberately. From the NVIDIA report: the question of whether accelerator demand is a subscription or a buildout — hynix is the purest wager that it is a subscription, because its contracts assume the refresh. From the Micron report: the trailing/forward P/E scissors as the market's live estimate of cycle mortality — hynix now posts the family's second-widest spread. From the CoreWeave report: what happens to take-or-pay economics when the counterparty's own economics wobble. From the Sandisk report: the canary ordering — spot-priced NAND cracks before contracted HBM ever can, which makes Sandisk's prints an early-warning system for this stock. One company now sits at the intersection of all four frameworks. That is why it gets the fourth report.

Part One · Longitudinal: from the bailout ward to the throne (1983–2026)

1. A chaebol's spare parts: Hyundai Electronics and a shotgun marriage (1983–2001)

SK hynix began as neither SK nor hynix. It was founded in 1983 as Hyundai Electronics, the semiconductor arm of the Hyundai chaebol, headquartered in Icheon, Gyeonggi Province — the same campus where its most advanced HBM lines run today. The same year, the LG group's Goldstar Electron (later LG Semicon) was born. For fifteen years the two ran parallel DRAM businesses, both perennially behind Samsung, both burning cash through the memory cycles of the late 1980s and 1990s.

The Asian financial crisis of 1997 ended that redundancy by force. Under the government-brokered "Big Deal" industrial restructuring, Korea's conglomerates were pressured to swap and consolidate overlapping businesses. In 1999, Hyundai Electronics acquired LG Semicon for about $2.1 billion. The merged entity briefly ranked among the largest DRAM producers on earth, a direct challenger to Micron and Samsung by capacity.

Capacity, it turned out, was the problem. The merger loaded the company with debt just as the dot-com bust arrived. In March 2001 it renamed itself Hynix Semiconductor — a portmanteau of "high" and "electronics" — and began separating from the collapsing Hyundai group, a divorce finalized in 2003. The new name arrived just in time to be attached to one of the great corporate near-death experiences in semiconductor history.

2. The bailout ward: 125 won, $7 billion of creditor mercy, and the merger that died overnight (2001–2002)

In 2001 DRAM prices fell roughly 80%. Hynix lost about ₩5 trillion that year, against total debts of about ₩8.6 trillion ($7.2 billion) and a debt ratio that touched 206%. The stock traded down to 125 won. Not 125,000 — 125.

On October 31, 2001, more than a hundred creditor institutions approved a rescue package worth about $7 billion: a ₩3 trillion debt-for-equity swap (later expanded to ₩4 trillion), ₩4.4 trillion of debt rolled over for three years at half the interest rate, ₩650 billion of new loans, and ₩1.2 trillion of debt written off at 25 cents on the won by seven banks. Control passed to a creditor council led by Korea Exchange Bank. Hynix had become, in effect, a ward of the Korean banking system.

Then came the episode that should be taught in every business school. In April 2002, Micron and Hynix signed a non-binding MOU: Micron would acquire Hynix's memory operations for 108.6 million Micron shares — roughly $3.2–3.4 billion at the time — plus $200 million for a 15% stake in the residual non-memory business. On April 29, the creditors approved it with 77.73% of votes, barely clearing the 75% threshold. The world's DRAM industry was about to consolidate into a Micron–Samsung duopoly.

The next day, April 30, the Hynix board met for six hours and rejected the deal unanimously. Their stated reasons: the agreement overvalued Micron's stock as currency, was too optimistic about contingent liabilities, and left the surviving rump company with untenable cash flows. The deal died that morning. Micron walked away furious; the creditors were stunned; the Korean press split between calling it courage and calling it suicide.

Every subsequent fact in this report descends from that veto. The $1.12 trillion company preparing the largest ADR in history exists because eleven directors decided, against their own creditors, that 108.6 million Micron shares was the wrong price for the future.

3. The tariff decade: fighting Micron in courtrooms while starving (2003–2011)

Micron did not treat the jilting as final — it treated it as actionable. In November 2002 it petitioned the U.S. Commerce Department, alleging Hynix had received $11.7 billion of illegal subsidies through state-influenced banks. The U.S. imposed countervailing duties of 57.37% preliminary, 44.71% final in June 2003, later pushed to 58.11% in an annual review. The EU followed with 34.8%. Korea took both to the WTO (cases DS296 and DS299); the tariffs did not fully come off until around 2008.

So through the entire mid-2000s — the years when Samsung compounded its lead — Hynix ran a DRAM business locked out of rich Western markets at competitive prices, owned by banks whose only strategy was to find an exit. The exits kept failing. In November 2009 the sole bidder, Hyosung, walked away. In 2011 two bidders emerged, SK Telecom and shipbuilder STX; STX withdrew that September, leaving exactly one buyer at the table for a company nobody else on earth would touch.

It is worth pausing on how structurally unwanted this asset was. Between 2001 and 2011, Hynix was nearly bankrupt once, nearly absorbed by its arch-rival once, tariffed by two continents, and put up for sale by its creditors for the better part of a decade — with one final bidder. The through-line of this series has been that the market repeatedly fails to price optionality in "discarded" assets: Sandisk was thrown away twice; CoreWeave was crypto-mining refuse; Micron itself spent decades as the industry's roadkill. Hynix is the extreme case: the asset was unwanted for ten consecutive years, at prices that now look like rounding errors.

4. The acquisition everyone mocked: SK's $3 billion "mistake" (2012)

In February 2012, SK Telecom paid about ₩3.37 trillion (roughly $3 billion) for a 21% controlling stake, and Hynix became SK hynix. Contemporary commentary was brutal: a telecom carrier with no semiconductor experience buying a capital-hungry, hyper-cyclical commodity manufacturer at the bottom of a memory bust, with zero synergy to its phone business. SK Group chairman Chey Tae-won pushed it through anyway.

Weeks after the deal closed, Japan's Elpida — the last non-Korean, non-Micron DRAM maker of scale — filed for bankruptcy. DRAM collapsed into a three-player oligopoly: Samsung, SK hynix, Micron. The industry's suicide gene — adding supply into every price rally — had finally killed enough players that the survivors could earn economic profit across a full cycle. By the 2017–2018 memory supercycle, SK hynix was generating more operating profit in a single quarter than the entire 2012 purchase price.

Two structural consequences of SK ownership matter for everything that follows. First, capital: a chaebol balance sheet ended the bank-ward era and let the company invest counter-cyclically. Second, culture: SK institutionalized a doctrine that engineering judgment outranks quarterly financial optimization. That doctrine is about to matter enormously, because in 2013 the company made the least financially defensible decision in its history.

5. The niche bet, a fire, and buying Intel's NAND at the exact top (2013–2021)

In 2010, AMD and SK hynix jointly proposed a new memory architecture to JEDEC: stack DRAM dies vertically, connect them with through-silicon vias, and put the stack on an interposer next to the processor. In 2013 SK hynix built the world's first HBM chip; JEDEC adopted the JESD235 standard that October. Mass production began in Icheon in Q1 2015, and the first product to ship with HBM was AMD's Fiji GPU — the Radeon R9 Fury X — in June 2015.

Here is the part that the 2026 narrative flattens: for most of a decade, HBM was a commercial failure. It was an expensive, low-volume graphics-card technology for a customer (AMD) losing share in GPUs. Samsung evaluated the segment and largely declined to commit; the margins were unattractive and the volumes were trivial. SK hynix kept a small team funded through years when the line item made no financial sense — engineering judgment outranking the spreadsheet. There was no ten-year master plan for AI; there was a stubborn refusal to kill an elegant architecture. Optionality, again, was free to whoever was willing to carry it.

Two other events from this stretch complete the picture. On September 4, 2013, a fire broke out during equipment installation at the company's Wuxi fab in China — a plant then producing over 10% of the world's DRAM. It burned for ninety minutes. DRAM spot prices jumped about 20% that day and 36% within three weeks. The lesson — memory supply is so concentrated that one accident reprices the global market — is the same lesson that AI buyers would relearn, from the demand side, a decade later. (The second lesson: the fire raised hynix's quarterly DRAM revenue, because price did more work than volume. Remember that asymmetry when we get to 2026 pricing.)

And in October 2020, SK hynix agreed to buy Intel's NAND and SSD business for $9 billion in two phases — closing phase one ($6.61 billion, the Dalian fab and SSD unit, renamed Solidigm) in December 2021, essentially the exact top of the NAND cycle. By the first half of 2023 the consolidated NAND operation was hemorrhaging — roughly 80% of a $4.49 billion half-year operating loss — and Solidigm sat in formal capital impairment. The 2018 Kioxia consortium investment (₩4 trillion alongside Bain for a stake in Toshiba Memory) looked similarly stranded. Both purchases were mocked with the same vocabulary once aimed at SK's 2012 acquisition. Hold that thought; both return in Part Two transformed.

6. The crown: from a -67% operating margin to beating Samsung's entire company (2022–2026)

The sequence that turned a component vendor into the AI trade's vault deserves precise dating.

October 2021: SK hynix announces HBM3 first. June 2022: it becomes the only company on earth mass-producing HBM3, just as NVIDIA's H100 — the chip that would train the ChatGPT era — enters production. Every H100 ships with SK hynix memory. Samsung, which had treated HBM as a niche it could enter at will, discovers that "at will" involves years of thermal-management and yield learning it hadn't done: its rival's MR-MUF packaging dissipates heat better than its own NCF route, and its 12-high stacks keep failing qualification.

At the same moment, the broader memory market collapses into its worst downturn in a decade. Q4 2022: SK hynix posts its first operating loss in ten years. Q1 2023: a ₩3.4 trillion operating loss on revenue down 58% — an operating margin of minus 67%. Full-year 2023: ₩7.73 trillion of operating losses. The company was simultaneously the sole supplier of the most strategically important component in computing and losing more money than at any point since the bailout era. The market, pricing the commodity and not the option, let the stock languish near ₩73,000 in late 2022. Anyone who wants to argue that memory equities are efficiently priced must first explain that year.

Then the option exercised itself:

From 125 won to 2,987,000 won is a factor of roughly 23,900. From the -67% margin quarter to the 72% margin quarter is twelve quarters. And eight days after taking Samsung's crown, the company announced it would sell $29 billion of new stock in New York. Part Three will treat that decision as what it is: the most information-dense event in the memory market this year.

Before leaving the history, mark the three lessons it hard-codes into any honest model of this company. First, memory economics are reflexive at the bottom as well as the top: the 2001 bailout, the 2012 Elpida exit and the 2023 loss trough each created the consolidation that powered the following boom — the industry's crashes are its moat-formation events. Second, this specific management team's defining behavior, observable across 2002 (rejecting Micron), 2013 (funding HBM), 2020 (buying Intel's NAND into a crash), is to make time-horizon arbitrage against counterparties who price the current cycle only. Third, and cutting the other way: every one of those bets was involuntary levered to a single variable — the memory price — and the company has never demonstrated it can protect shareholders within a downcycle, only that it can position them for the recovery. The stock fell ~75% peak-to-trough into 2022 with the HBM crown already effectively won. Owning this company has always meant owning that volatility. The ADR does not change it; it exports it.

Part Two · Cross-sectional: the memory board in July 2026

7. HBM economics: how a commodity grew contracts

This series has used one ordering device across three reports: rank every AI-hardware business by how much contract cover stands between its revenue and the spot market. NVIDIA sells allocation under multi-year scarcity. CoreWeave rents GPUs on take-or-pay leases with 2030–32 renewal risk. Micron's HBM is contracted years forward while its conventional DRAM reprices quarterly. Sandisk's NAND reprices quarterly with essentially no contract cover — the complex's canary.

SK hynix now sits at the extreme protected end of that spectrum, and understanding why is the analytical core of this report.

HBM is physically a DRAM product, but economically it behaves like an aerospace component. It is co-designed with the customer's silicon (the base die increasingly carries customer-specific logic — "custom HBM"), qualified over quarters, packaged into the customer's product at an OSAT or foundry, and impossible to substitute mid-platform. Because a GPU vendor who misses memory misses its own product cycle, HBM is sold on annual and multi-year contracts with pre-agreed volumes and prices, negotiated a year or more ahead. That structure produced three facts that would have been unthinkable in any prior memory cycle:

Price coordinates, per industry estimates: HBM3 ran about $200 per stack, HBM3E about $300, HBM4 about $500. Twelve stacks or more per next-generation GPU. The bill-of-materials arithmetic explains why memory, not logic, has become the binding constraint of the AI buildout — and why the phrase "commodity memory" now describes only part of these companies.

The wafer economics deserve one more paragraph, because they explain both the pricing power and the coming squeeze. HBM consumes silicon at a brutal exchange rate: stacking, through-silicon-via drilling, base dies and yield loss mean one bit of HBM costs roughly three times the wafer area of one bit of conventional DRAM — the industry calls it the trade ratio. Every wafer reallocated to HBM therefore removes about three wafers' worth of bits from the conventional market. With HBM headed toward ~30% of total DRAM wafer input by 2027, the conventional DRAM shortage of 2026 — the +90–95% contract quarter — is not a separate phenomenon from the AI boom; it is its shadow. This is why hyperscalers began signing long-term agreements for commodity DRAM, a product that has traded spot since the 1970s, and why the tightness self-reinforces: the more HBM the leaders ship, the scarcer everything else they make becomes. A vertically integrated pricing flywheel, running until new greenfield capacity lands in late 2027–2028.

Custom HBM extends the moat one layer deeper. From HBM4 onward, the base die at the bottom of each stack migrates to logic processes — TSMC and the memory makers' own foundry-class nodes — and increasingly carries customer-specific controller logic: an NVIDIA stack and a Google stack stop being interchangeable parts. Co-design cycles run 18–24 months; switching suppliers mid-generation approaches impossible. The strategic consequence cuts both ways. For hynix, customers who co-design are customers who cannot leave, and multi-year volume commitments follow naturally. Against hynix, the same logic binds it to the fortunes and roadmap decisions of a handful of counterparties — concentration risk wearing a moat costume. Both descriptions are accurate; which one dominates depends entirely on whether the customers' own demand holds, which is Section 9's subject.

The structural point for valuation: a business that pre-sells output at fixed-or-floored prices for 12–36 months is not the business the market's memory-cycle reflexes were trained on. It is closer to CoreWeave's take-or-pay model — with the crucial difference that hynix's counterparties are the richest companies on earth buying a scarce input, not leveraged startups renting a depreciating one.

8. The board: a wounded giant, a charging American, and 58% of the pie

SK hynix holds 56–62% of the HBM market depending on quarter and counting method (Counterpoint: 56.4% in Q1 2026, ~62% of Q2 shipments), roughly 29% of DRAM overall, and — this is the fact the 2023 narrative never anticipated — is about to hold roughly two-thirds of the HBM4 supply for NVIDIA's Vera Rubin platform. Its HBM4 finished development first (September 2025), its 12-high HBM4E samples shipped to lead customers on June 18, 2026, months early, and its M15X fab came online four months ahead of schedule. Execution, for now, is monotone.

Why is the share so sticky? Because HBM leadership compounds through three loops that all favor the incumbent. Yield: stacking twelve dies with thousands of TSVs means small per-die yield edges multiply into large stack-level cost gaps — hynix's decade of learning curve is embedded in physics, not patents, and cannot be licensed. Packaging: its MR-MUF process is the industry's thermal benchmark, and packaging capacity (not wafers) is the binding constraint everyone is racing to add — hence ₩19 trillion of the ADR going to one packaging plant. Qualification: the customer's cost of switching mid-platform is measured in months of requalification on a product whose unavailability idles $40,000 GPUs. Share moves at platform boundaries, essentially never inside them; the 2026 boundary (Rubin) has already been divided, two-thirds to the incumbent. The next genuine contest is 2027's platform — which is exactly why autumn's contract headlines carry so much information.

Samsung is the wounded giant, and the wound is instructive. It passed on early HBM, then spent 2023–2025 failing NVIDIA qualification on heat and yield — the MR-MUF vs NCF packaging gap — while reorganizing HBM teams (the special HBM unit created in 2024 was dissolved back into the DRAM division in late 2025). But a wounded Samsung is still Samsung: its HBM4 passed NVIDIA qualification around January 2026, entered production in February, reached full supply by June at 11.7Gbps — above JEDEC spec — and it has reportedly secured 25–30% of Rubin HBM4 allocation plus 60% of Google TPU's HBM3E. Its 1c DRAM yield is ~70% and climbing toward 200K wafers/month by end-2026, and it showcased HBM4E at NVIDIA GTC. The "Samsung permanently broken" thesis embedded in some hynix bull cases is stale by roughly six months. What Samsung has not done is displace hynix's share leadership or its packaging edge — 2026 allocations were set with hynix at two-thirds.

The honest counterweight to the stickiness argument: it is a description of this generation's physics, not a law. Hybrid bonding — the wafer-to-wafer technique arriving with 20-high stacks around HBM5 in 2028–29 — resets part of the packaging learning curve for everyone at once, and resets are when incumbencies break. Samsung is explicitly investing to make that its re-entry point. The moat is real and it is depreciating on a published schedule; the market's 4.2× mortality pricing is, among other things, a view on that schedule.

Micron is the charging third. Its FQ3 2026 print (June 24) was the most violent earnings report of the season: revenue $41.5 billion, gross margin 81.6%, EPS $25.11, and an FQ4 guide of ~$50 billion at ~86% gross margin. Its HBM is sold out through calendar 2027 with demand visibility into 2028, HBM4 is shipping in high volume to its lead customer, and it targets ~25% HBM share by year-end — it has already overtaken Samsung in HBM share by some counts (21% vs 17%, Counterpoint Q1). Micron's report is also the reason SK hynix rallied 12% on June 25: when the number-three player prints an 81.6% gross margin and says supply is constrained past 2027, the read-through to the leader is mechanical. The single-organism behavior of this complex — the thesis of our Micron and Sandisk reports — now operates across the Pacific in both directions.

Below the HBM tier sits conventional DRAM, where the 2026 repricing has been historic: contract prices +90–95% in Q1 2026 (the largest quarterly increase ever recorded), +58–63% in Q2, with Q3 expected +13–18% as the impulse decelerates. Cloud providers are signing long-term agreements for commodity DRAM — LTAs for a product that never had them — because AI inference turned out to be a general-purpose-server workload too. Effective new supply arrives late 2027 at the earliest. Samsung 38% / hynix 29% / Micron 22% split the segment.

Two competitive wrinkles complete the board. Samsung's foundry arm — the $16.5 billion Tesla AI5/AI6 contract at Taylor, 2nm production standing up through 2026–27 — matters to the memory race indirectly but materially: a foundry that works gives Samsung both a second income stream to fund HBM price war if it chooses, and logic-process capability for custom HBM base dies that neither hynix nor Micron owns in-house. Hynix's counter is its TSMC partnership for HBM4 base dies — a best-of-breed alliance against Samsung's vertical integration, replaying a pattern (specialist + TSMC vs integrated giant) that has decided several other semiconductor wars in the specialist's favor. Meanwhile Micron carries a geographic option neither Korean rival has: U.S.-soil DRAM fabs (Idaho from 2H 2027, New York later) and a Virginia HBM packaging line, which in any deepening of U.S.–China or U.S.–Korea trade friction converts from cost disadvantage to policy moat. None of these wrinkles changes the 2026 allocation math; all of them shape who bleeds least in the next downcycle.

9. The demand engine: two-thirds of NVIDIA, the ASIC second engine, and Stargate

The demand side of hynix's book is concentrated in the handful of logos building frontier AI capacity, in three tiers.

Tier one: NVIDIA. Hynix supplies roughly two-thirds to 70% of HBM4 for the Rubin generation, as it did the majority of HBM3/3E for Hopper and Blackwell. No official disclosure exists of NVIDIA's share of hynix revenue; with HBM above 40% of DRAM revenue (last disclosed) and hynix holding ~2/3 of NVIDIA's allocation, plausible estimates put the single-customer exposure in the 25–40% band. This is the concentration risk at the heart of the story — the same customer whose own report we described as needing demand to be "recurring, not episodic."

Tier two: the ASIC buildout. Broadcom-fabbed Google TPUs, AMD MI-series, and the custom accelerators of every hyperscaler consume HBM in accelerating volume — this demand is why HBM3E prices rose 20% into 2026 contracts. Note the competitive texture: Samsung supplies 60%+ of Google TPU HBM3E and won AMD's MI455X HBM4, while hynix led Broadcom-Google orders in H1 2026. The ASIC tier is where the three-way rivalry is genuinely contested, and it is growing faster than the GPU tier.

Tier three: the option book. The OpenAI Stargate letter of intent — which SK hynix says could more than double industry HBM demand on its own — plus Microsoft, Google and Meta locking multi-year custom-HBM volumes directly, plus High Bandwidth Flash (Section 12) as a wholly new socket in inference boxes from 2027.

One demand-side shift deserves separate billing because it changed the shape of hynix's book in 2026: the migration from training to inference. Training demand is lumpy, frontier-lab-concentrated, and HBM-intensive. Inference demand is broad, continuous, and — crucially — general-purpose: serving models at scale consumes not just HBM on accelerators but ordinary RDIMM server memory and enterprise QLC SSDs in previously unmodeled volumes. This is what detonated the conventional DRAM repricing (Section 7's shadow shortage) and what turned Solidigm's eSSD book into the industry's growth story (Section 12). It also improves demand quality: inference revenue at the hyperscalers is tied to serving paying users, not to speculative model races — the closest thing the AI buildout has yet produced to recurring, unit-economics-positive demand for memory. The bear case must now argue against two demand engines, not one, and the second one has revenue attached.

The Meta Compute story slots into this framework precisely, which is why it moved memory stocks harder than GPU stocks. If hyperscalers begin reselling surplus compute, the market's first-order read is that tier-one and tier-two demand were partially double-counted — the same GPU purchase showing up as both Meta's capex and, eventually, someone's rented capacity. For a GPU landlord that is a competition problem (our CoreWeave report's subject); for a memory supplier it is a demand-quality question: contracted volumes stand, but the 2027 negotiation happens against customers who suddenly know their own utilization better. Nothing about July 2 changed a single signed contract. What it changed was the prior on how the next contract gets negotiated.

Set against this: 2026 hyperscaler capex guidance of $635–690 billion (+67–74% YoY), with the 2027 sell-side split — Morgan Stanley near $1.1 trillion, Goldman near $920 billion — defining the exact axis of the "deceleration scare" that produced this month's -14.5% day. The demand question for hynix is never whether 2027 capex grows; it is whether the second derivative stays positive enough to absorb three suppliers' expanded HBM output at contracted prices. That is the honest formulation, and it is genuinely open.

10. The contract-cover hierarchy and the July 2 stress test

On July 2, 2026, Bloomberg reported Meta was preparing to sell surplus AI compute — the "Meta Compute" story that our CoreWeave report treated as a structural break for GPU landlords. What happened next was a live experiment in how the memory complex prices risk.

SK hynix — the most contract-protected company in the entire AI hardware stack — fell 14.5%, its worst day in years. Samsung fell 9.1%. KOSPI fell 7.89%; the exchange briefly halted program selling; $290 billion of market value left two stocks in one session. The next day, Seoul repriced the panic: hynix +10.9%, Samsung +8.2%, KOSPI +5.76% — while U.S. markets were closed for July 4th, which is why our replies that week kept pointing at Sunday-night Seoul as the arbitration session.

Read the stress test against the hierarchy. By fundamental exposure, the ordering of vulnerability to an AI-capex deceleration runs: Sandisk (quarterly NAND repricing, no cover) > CoreWeave (rental renewal risk) > Micron (HBM covered, conventional DRAM spot) > SK hynix (multi-year contracts, sold out three years, price caps removed). By realized drawdown on July 2, hynix fell as much as Sandisk had the prior Friday. The most protected name traded like the least protected one.

Why? Decompose the -14.5%. First, index mechanics: hynix plus Samsung are roughly a third of KOSPI's weight, so any Korea-level de-risking flows through them regardless of story. Second, the amplifiers: single-stock leveraged ETFs tracking hynix and Samsung — products regulators had publicly criticized in June — force mechanical selling into declines, and foreign investors, who had driven the +96% half, sell the most liquid names first when they cut Korea. Third, the calendar: with book building starting July 6, every institution that wanted ADR allocation had an incentive to see (or at least not resist) a lower reference price. None of these three forces knows or cares what a take-or-pay HBM contract is. The July 3 snap-back — +10.9% on the same fundamentals — completed the demonstration: both the crash and the recovery were about who owned the stock and why, not about what the company will earn.

The resolution of that paradox is positioning, not fundamentals: a stock up 331% year-to-date, inside a KOSPI up 96% in a half (its best ever), amplified by single-stock leveraged ETFs regulators had already flagged, two trading days before a record equity offering. Contract cover protects cash flows; it does not protect multiples, and it certainly does not protect crowded trades. That distinction — cash-flow risk versus positioning risk — is the correct frame for everything in Part Three, including the ADR itself.

11. The China ledger: 36–40% of DRAM behind an annual visa, and CXMT's clock

SK hynix's Wuxi fab — the one that caught fire in 2013 — produces roughly 36–40% of its DRAM wafers (about 200K of ~550K wafer starts/month), now ~90% converted to the 1a node. Since Washington revoked the Validated End-User waiver (announced September 2025, effective December 31), that capacity operates on annual licenses that permit maintenance but prohibit expansion or upgrades. The 2026 license was granted. Each renewal is now an annual geopolitical event; the long-run trajectory is a frozen node in a business defined by node migration. The Yongin buildout is, among other things, the repatriation of that risk — at ₩600 trillion.

The repatriation program has an American wing too: a roughly $4 billion advanced-packaging plant in Indiana, announced alongside the ADR's use-of-proceeds narrative — HBM assembly on U.S. soil for U.S. customers, the same policy-moat logic Micron is building in Virginia. Between Yongin (₩600T), Cheongju P&T7 (₩19T of the ADR alone) and Indiana, the geography of hynix's next decade is being redrawn away from the annual-license regime — but on fab timelines, which is to say the exposure persists through at least 2028 while the exits are poured in concrete.

On the challenger side, CXMT's HBM3 mass-production target has already slipped out of 2026 (per Digitimes in April); its HBM capacity is a rounding error today (~5K wafer starts at end-2025, ~30K planned end-2026, ~55K end-2027) with yields below 40% expected into H2 2026 and 12-high stacks targeted only for 2027. The sober read: China is not an HBM threat before 2028, but CXMT and YMTC (already expanding NAND aggressively, ~13% share) are building the conventional DRAM/NAND capacity that will deepen the next downcycle — a 2028 problem the market will start pricing in 2027. Add the unresolved question of whether CXMT joins the Entity List, and China functions in this story as slow-moving supply risk plus fast-moving policy risk on hynix's own fab.

12. The second act nobody prices: Solidigm's turn, the Kioxia stake, and HBF

Three "failed" capital decisions from Part One re-enter here with different signs.

Solidigm (the $9 billion Intel NAND purchase that closed at the top): after the impairment years, it returned to annual profit as AI eSSD demand exploded — hynix + Solidigm hold 30.2% of enterprise SSD, the top position, with QLC leadership precisely where AI inference storage concentrates. Sell-side models put Solidigm around ₩11.8 trillion revenue and ₩1.4 trillion net for 2026. The asset bought for the wrong cycle turned out to be positioned for the right one.

Kioxia (the mocked ₩4 trillion consortium ticket from 2018): now a listed stake worth about $39.7 billion — roughly 15× cost, visible in Q1's net-income-above-operating-income anomaly. Strategically stranger: hynix is a major shareholder of its own NAND competitor, which is itself Sandisk's manufacturing twin. The memory industry's ownership graph is a web, which is one more reason the complex trades as one organism.

High Bandwidth Flash: with Sandisk (initiator) and hynix as co-standardizers under OCP since February 2026, HBF proposes NAND stacked HBM-style — ~1.6TB/s bandwidth with 8–16× HBM capacity — targeting the inference memory wall, first samples H2 2026, device prototypes early 2027. For hynix this is a hedge on both sides of its own franchise: if flash invades the memory tier, hynix co-owns the standard; if it doesn't, hynix's DRAM tier is untouched. Our Sandisk report called HBF "the option that isn't in the multiple"; the same clause applies here, with the difference that hynix's multiple needs no options to justify it.

The pattern across all three: this management's worst-timed purchases keep resolving into strategic assets one cycle later. That is either luck twice, or evidence that buying storage assets when they are hated is a repeatable institutional skill. The distinction matters for how you read the ₩1,100 trillion capex program.

About that program, disclosed June 29 and easy to misread as a single number: ₩600 trillion for the Yongin cluster (four fabs, ~₩120–150 trillion each, adding roughly 360K DRAM wafer starts per month when complete — a two-thirds expansion of current company-wide capacity), ₩100 trillion for Cheongju including the P&T7 packaging complex, and ₩400 trillion for a new southwestern cluster, all spread across a decade-plus. It is not a 2026 commitment; it is a declaration of intent to own the next two capacity cycles, funded in tranches of which the ADR is merely the first public one. The correct comparison set is TSMC's Arizona program and Samsung's Taylor bet: the memory leader is now spending like a foundry — which is to say, like a company that believes its product has stopped being a commodity. Whether the belief is right is Scenario C's question; that the belief is genuinely held is no longer in doubt, because ₩1,100 trillion is not a hedge.

Part Three · Synthesis: the snapshot, the $29 billion tell, and what has to be true

13. Where it stands: the July 2026 snapshot

The operating trajectory, in one table (K-IFRS, trillions of won):

Quarter Revenue Op. profit Op. margin Net profit
Q2 2025 22.23 9.21 41% 7.00
Q3 2025 24.45 11.38 47% 12.60
Q4 2025 32.83 19.17 58% 15.25
Q1 2026 52.58 37.61 72% 40.35

Read the table's slope before its levels: revenue roughly doubling every two quarters while the margin climbs 41% → 47% → 58% → 72% is not a demand story alone — it is the arithmetic of contracted price increases landing on a fixed cost base, the same operating-leverage signature Micron printed a quarter earlier and Sandisk two quarters earlier. The complex is one machine, reporting in three accents and two currencies.

Four consecutive quarters of margin expansion, each one setting a company record; Q1 2026 revenue up 198% year-on-year and 60% sequentially — the first ₩50-trillion quarter in company history, roughly $35.5 billion, earned at a margin that microprocessor monopolies would envy. (Q1 net income exceeds operating income on non-operating gains, mostly the Kioxia mark; use operating profit as the clean earnings-power line.) Cash of ₩54.3 trillion, net cash ₩35 trillion (~$25 billion) — before the ADR adds $29 billion more.

Q2 2026 reports on July 29 — nineteen days after the ADR lists. Sell-side estimates cluster around ₩82 trillion revenue and ₩63–66 trillion operating profit (76–77% margins). For full-year 2026, the sell-side spread is itself the story: operating profit estimates run from ₩148 trillion (Mirae) to ₩289 trillion (NH, post-revision) — a nearly 2× disagreement about the current year, in July. Brokers' price targets sit at ₩4.0–4.1 million against a ₩2.43 million close: the street is telling you it believes the aggressive scenario while the tape prices the conservative one.

Two translation notes before the valuation table. First, currency: the won at 1,500+ per dollar — a 28-year low — flatters every dollar-denominated comparison of hynix's costs and inflates the won value of its dollar-denominated revenue; a 10% won recovery would mechanically trim reported margins even with unchanged chip prices. The ADR's $29 billion inflow, when converted, is itself large enough to move the currency (the announcement alone moved USD/KRW by 30 won in a session), a small reflexive loop between the offering and the earnings it is being priced on. Second, the balance sheet: unlike the 2001 vintage of this company — 206% debt ratio, creditor committees — the 2026 vintage runs ₩35 trillion of net cash before the raise. Whatever the cycle does, the balance-sheet failure mode that defined memory investing for three decades is not on the menu this time; the downside cases are about multiple compression and earnings normalization, not survival. That asymmetry is new to this cycle and underpriced in casual bear cases.

Valuation, against the family (July 3–4 data; Korean trailing multiples computed from TTM net income, so treat as estimates):

Company Market cap Trailing P/E Forward P/E Trailing ÷ forward
NVIDIA $4.72T 29.8× 15.3× 2.0×
SK hynix ~$1.12T ~22.9× ~5.5× ~4.2×
Samsung ~$1.32T ~24.4× ~4.9× ~5.0×
Micron $1.10T 22.1× 6.5× 3.4×
Sandisk $258B 59.5× 8.6× 6.9×

The trailing-to-forward ratio is this series' one-line instrument: it is the market's real-time estimate of the cycle's mortality — how much of today's earnings power the market expects to die. NVIDIA trades at 2× (persistence mostly believed). The Korean memory pair trades at 4–5×: the market is paying trillion-dollar prices while simultaneously pricing the earnings as mostly transient. Note what hynix's 5.5× forward actually embeds: consensus-forward net income around ₩310 trillion, the aggressive end of that 2× sell-side spread. The cheapest-looking number in the AI trade is cheap only if the most optimistic 2026 model on the street is right. That is the honest fine print under "5.5× forward."

14. The $29 billion tell: reading the ADR as an analytic object

Strip the ceremony off the listing and describe it plainly: the most informed participant in the global memory market has chosen to sell 2.5% of itself for $29 billion, in New York, at prices within ~20% of an all-time high reached after a +331% year, with pricing set for July 9 and trading on July 10. Book building starts July 6 — the Sunday-night Seoul session our replies flagged as the post-crash arbitration.

The bear reading writes itself, and we should state it without flinching: record offerings mark tops. Alibaba's $21.8 billion in September 2014 preceded a 60% drawdown within a year. Glencore listed at the exact peak of the commodity supercycle in 2011 and spent half a decade below issue. When the smartest seller in a cyclical industry monetizes at scale, that is information. Add the tactical texture — a -14.5% single-day crash two sessions before book building, a KOSPI regulators already worry is levered, a won at 28-year lows — and the listing looks like urgency.

But the precedent file has a second drawer. Visa priced the largest U.S. IPO in history in March 2008 — into the teeth of a financial crisis — and compounded relentlessly for fifteen years, because what it was selling was a toll booth, not a cycle. Saudi Aramco's record listing in 2019 marked nothing at all. The "record offering = top" rule is really a rule about what kind of earnings are being monetized: cyclical peak earnings sold at peak multiples mark tops; structurally growing earnings sold at reasonable multiples do not. Hynix's offering is the interesting hybrid — arguably peak-cycle earnings, but at 5.5× forward rather than Glencore's or Alibaba's premium multiples, to fund capacity that is contractually pre-sold. The multiple is the bulls' best exhibit: tops are usually sold expensive, and whatever else this is, it is not being sold expensive.

The bull reading is more specific, and on the evidence, stronger:

  1. The uses are named and real. The F-1 earmarks the proceeds: ₩9.41 trillion for Yongin cluster Fab 1 (ground broken February 2025, cleanroom Q1 2027), ₩19 trillion for the Cheongju P&T7 advanced-packaging plant (the HBM bottleneck is packaging, not wafers), and about $7.7 billion of EUV tools delivering through 2027. This is capacity for contracts the company says are already oversubscribed three years out — pre-selling output and pre-funding the plant is the same risk posture, applied to both sides of the balance sheet.
  2. The dilution is trivial against the program. 2.5% of shares for $29 billion, against a ₩1,100 trillion long-horizon capex plan (Yongin ₩600T + Cheongju ₩100T + a new southwestern cluster ₩400T). Equity at 72% operating margins is the cheapest risk capital this company will ever see, and — unlike 2001 — it is raising from strength, from strangers, with no covenants.
  3. The structural motive is durable. A Nasdaq line converts the "Korea discount" — decades of governance-and-geography haircut — into direct access to the world's deepest capital pool, hedges the Wuxi/annual-license geopolitical file with a U.S.-investor constituency, and gives future capital raises a dollar-denominated instrument. Ten ADSs per share also does something subtler: it turns a ₩2.4 million stock into a ~$170 ticket for the retail flow that made "memory" a household trade this year.

Mechanics worth knowing before the tape starts printing. The deal is a primary offering of new shares — no insider is selling; SK Group's stake dilutes alongside everyone's, which removes the classic "smart money exiting" reading in its narrowest form. Four bookrunners (BofA, Citi, Goldman, JPMorgan) carry an unusual 0.5% discretionary incentive fee — a tell that the issuer knows execution in this tape is not automatic. The implied reference price (₩45.45T ÷ 17.79M shares ≈ ₩2.55 million) sits above Friday's ₩2.43 million close — the offering was sized against pre-crash prices, so either the book comes in hot enough to close that gap by Thursday, or the raise shrinks, or the range cuts. Each of those three outcomes is a distinct, legible signal about institutional belief in the 2027 question, delivered before Q2 earnings on July 29. Rarely does a market schedule its own referendum this cleanly.

Both readings are true simultaneously; the question is weighting. Our weighting: the listing is rational opportunism — a company pre-funding a decade of committed capacity at the best prices it has ever been offered — but its timing is still a sentiment thermometer. Companies do not get to choose whether their record offerings become historical markers; the cycle chooses. What the analyst can do is watch the tape: if the largest ADR in history prices at the bottom of its range, or breaks issue price in week one, the market will have voted on the 2027 capex question with $29 billion of real money. That makes July 10 the single most information-rich data point of the summer for the entire AI hardware complex — more than any earnings print, because it aggregates all of them.

15. Three scenarios for the year ahead

Numeric anchors first, so the scenarios are falsifiable rather than moods. The 2026 consensus corridor is roughly ₩290–330 trillion revenue and ₩148–289 trillion operating profit (that sell-side spread again). The forward multiple of ~5.5× is computed against roughly ₩310 trillion of net income — the top of the corridor. Every scenario below is really a statement about where 2027 lands relative to that anchor.

Scenario A — the supercycle holds (the street's implicit base case). 2027 hyperscaler capex lands near the Morgan Stanley ~$1.1 trillion track; 2027 HBM contracts settle up sharply (TrendForce's "multiples" read proves directionally right); conventional DRAM stays tight into 2028 as HBM absorbs 30% of wafer input. Hynix earns something in the NH-shaped range, the forward multiple proves to have been under 6× on real numbers, and the stock's ₩4 million targets stop looking exotic. In this world the ADR is remembered like Visa's 2008 IPO — a record offering that marked the beginning, not the end. Watch for: Q2 print on July 29 clearing ₩63T operating profit, 2027 contract headlines through autumn, SKHY trading above issue.

Scenario B — the plateau (deceleration without collapse). 2027 capex grows ~20% (the Goldman track); HBM stays sold out but 2027 pricing settles flat-to-modestly-up as Samsung's and Micron's added supply meets decelerating incremental demand; conventional DRAM/NAND begin normalizing in 2H 2027. Earnings stay enormous but stop growing; the trailing multiple compresses toward the forward one the slow way — sideways. The stock, up 331% YTD, de-rates 20–40% on positioning alone (July 2 was the preview). This is the modal outcome across semiconductor history and it is not a disaster; it is what "cycle" means. The ADR still funds the fabs; buyers at ₩2.9 million wait years.

Scenario C — the air pocket. A capex deceleration announcement from a top-three hyperscaler (a Meta-Compute sequel with numbers attached), an inference-efficiency shock, or a macro break turns 2027 contract negotiations from a price-setting exercise into a volume-confirmation exercise. HBM contracts protect stated volumes but new capacity (hynix's own Yongin among it) arrives into falling utilization elsewhere; conventional DRAM gives back its historic 2026 repricing at the speed it came; CXMT/YMTC's low-end supply deepens the trough. Hynix's earnings do not vanish — the contracts are real — but 2027 estimates halve, and a stock owned on momentum finds its multiple pricing the next trough, as it did at ₩73,000 in 2022. The canary order still applies: Sandisk's spot-priced margins crack first, Micron's conventional book second, hynix's contracts last. If SNDK's margins are printing fine, the air-pocket case for hynix is not live.

We decline to assign false-precision probabilities; the honest statement is that A and B together dominate C on current evidence (contracts, sold-out status, hyperscaler cash flows), while the stock's one-year path is hostage to positioning in all three — the July 2 session proved the multiple can move 15% in a day with zero fundamental news.

A word on the asymmetry between the scenarios, because it is not symmetric around the current price. In A, the forward multiple has room to double and still sit below Micron's, on earnings that are themselves growing — the compounding case brokers' ₩4 million targets encode. In B, earnings hold but the multiple's mortality assumption is validated: the stock marks time or gives back a third while the trailing number catches down to the forward one. In C, the 2027 estimate — not the 2026 one — halves, and because the current price is anchored to that estimate, the drawdown mathematically resembles 2022's ~75% at the extreme, cushioned this time by net cash and contracts. Expected value across the three is decent; the path through any of them includes July 2-sized days as a recurring feature, not a bug. Position sizing, not scenario selection, is where most of the risk management in this name actually lives. (We state this as analysis of the instrument's character, not as advice.)

16. The valuation paradox, fourth verse

Our Micron report described the memory paradox: record earnings, single-digit forward multiples, and a market refusing to capitalize a cycle's peak. Sandisk was the paradox's loudest rendition (59.5×/8.6×). SK hynix is its most consequential, because the absolute numbers are now systemic: a ~$1.12 trillion company at ~5.5× forward earnings, simultaneously the largest equity issuer in ADR history.

State the paradox precisely. If the market believed 2026-level earnings persist even three years, hynix at 5.5× forward would be the cheapest trillion-dollar asset ever listed. It does not believe it — the 4.2× trailing-to-forward ratio says most of this earnings power is priced to die. Yet the same market pays a trailing 23× — more than seven points above NVIDIA's forward — for the privilege of holding the position while it decides. The trade crowding into KOSPI (+96% in a half), the levered single-stock ETFs, the $29 billion of new paper being minted into the debate: none of this resolves the paradox; it arbitrages the volatility the paradox generates.

Three things are genuinely different from every prior memory cycle, and one thing is not. Different: (1) contract architecture — output pre-sold years out, price caps removed, custom silicon binding customers; (2) demand character — a capex program measured in fractions of a trillion dollars, funded from the operating cash flow of the five richest companies on earth rather than from leverage; (3) supply discipline — three consolidated players, all of whom lived through 2023's -67% margins, adding capacity against contracts rather than against hope. Not different: the marginal buyer's psychology, which in July 2026 produced a 14.5% single-day repricing on a newspaper story. The fundamentals of this cycle are the strongest the industry has seen; the positioning around it is the most crowded. Both statements are true, and the second one governs the next quarter while the first governs the next three years.

Step back once more and read the four P/E scissors as a single sentence, because the complex is speaking in unison. The market says: NVIDIA's earnings are half-durable (2.0×); Micron's mostly die (3.4×); hynix's and Samsung's mostly die (4.2×, 5.0×); Sandisk's almost entirely die (6.9×). Now overlay the contract hierarchy from Section 10 — hynix the most protected, Sandisk the least — and notice the market's mortality estimates run in exactly the same order as fundamental protection, except priced as if the protection barely matters. Either the market is right that contracts cannot survive a demand break (in which case NVIDIA at 2.0× is the mispricing, because its customers' contracts are its revenue), or the contracts are worth roughly what they say (in which case the Korean pair's 4–5× mortality pricing is the anomaly). The four multiples cannot all be correct simultaneously. Some pair of them is wrong, and the resolution — visible first in Sandisk's spot prints and the 2027 contract headlines — is the single largest source of expected repricing anywhere in the AI hardware complex over the next twelve months. That, rather than any single stock call, is this report's central analytical claim.

17. What would have to be true, the risk matrix, and five signals

For the bull case (A) to pay from ₩2.43M / the ADR price, all of the following must hold:

  1. 2027 hyperscaler capex grows meaningfully (>25%) rather than merely persisting;
  2. 2027 HBM contract pricing settles up, validating the removed price caps;
  3. Samsung's recovery adds share at the edges (ASIC tier) without breaking hynix's ~60% NVIDIA allocation;
  4. Conventional DRAM tightness holds through 2027 as HBM wafer absorption starves supply;
  5. No China policy shock touches Wuxi's annual license.

Falsify any two and you are in Scenario B; falsify (1) plus any other and C is live.

Risk matrix (probability × severity, one-line each):

A note on what is deliberately absent from the matrix: balance-sheet risk (₩35 trillion net cash plus the raise), customer-credit risk (the buyers are the five largest cash generators alive), and technology-leapfrog risk inside the visible roadmap (HBM4E samples already shipped early; the 2027–28 roadmap through HBM5 is published and hynix leads each node). The risks that remain are almost entirely price risks — of the product, of the stock, of the 2027 contract — which is precisely what you would expect of a company at maximum operational execution. When everything the company controls is going right, everything that remains is what it cannot control. That sentence is the entire risk section, compressed.

Five signals, in the order they would fire (our tracking dashboard for the complex):

  1. SKHY July 9–10: pricing within range and week-one trading — the market's aggregated vote on 2027; below-range pricing or a broken issue is the earliest systemic warning available anywhere in AI hardware.
  2. Sandisk's NAND prints (quarterly, spot-priced, no contract cover): the canary. If SNDK gross margins hold near 80%, the demand floor is intact; hynix cannot have a demand problem SNDK doesn't have first.
  3. Q2 2026 earnings, July 29: ₩63–66T operating profit is the bar; the mix disclosure (HBM share of DRAM revenue, if given) matters more than the total.
  4. 2027 HBM contract headlines (autumn 2026): pricing direction and whether Samsung's allocation share rises above ~30% — the first hard read on both the pricing and the share thesis.
  5. Hyperscaler Q3 2026 guidance (October): the 2027 capex second derivative, the variable every scenario branches on.

Signal design notes, briefly. The five are ordered by information arrival, not importance — the ADR fires within a week, the capex guidance in October. They are chosen to be pairwise independent: the ADR reads aggregate institutional belief, Sandisk reads spot demand, July 29 reads hynix's own mix, the contract headlines read the 2027 price, hyperscaler guidance reads the 2027 quantity. A bull whose thesis survives all five arriving hostile was never holding a thesis; a bear whose thesis survives all five arriving friendly is in the same position. We will track each publicly as it lands.

The bottom line. SK hynix is the best business in the memory industry's history at the moment of its maximum earnings power, selling the largest ADR ever printed to fund capacity its customers have already contracted. The equity costs ~5.5× a forward estimate that requires the street's aggressive case, inside a positioning structure that just demonstrated 15% single-day fragility, six days before a $29 billion supply event. The company that was worth 125 won because nobody believed anyone would want its chips is now worth $1.12 trillion because everybody believes everyone will. The 2002 board meeting teaches the only durable lesson in this file: the price of memory assets is set by conviction about the future, and conviction — in both directions — has always been this industry's least reliable component. Watch the five signals; they are designed to tell you which scenario is arriving before the multiple does.

Sources

Report generated by the Aya Invest research pipeline. Every claim above traces to a public source; where figures are computed (Korean trailing multiples) or estimated (NVIDIA revenue share), the text says so. For information and research purposes only. Not investment advice.

FAQ

What is the SK hynix Nasdaq ADR listing (SKHY) and why is it historic?

SK hynix plans to issue 17.79 million new shares as American Depositary Receipts (10 ADS = 1 share) on Nasdaq under ticker SKHY, raising about ₩45.45 trillion (~$29B) — the largest ADR offering in history, surpassing Alibaba's $21.8B in 2014. Book building starts July 6, pricing July 9, trading July 10, 2026. Proceeds are earmarked for the Yongin cluster Fab 1 (₩9.41T), the Cheongju P&T7 advanced packaging plant (₩19T) and ~$7.7B of EUV tools. Dilution is only ~2.5% of shares outstanding.

Why does SK hynix dominate HBM, and can Samsung and Micron catch up?

It bet on HBM with AMD in 2013 when the technology was a money-losing niche, giving it a decade of yield and packaging learning (its MR-MUF process is the thermal benchmark). It was the sole HBM3 supplier for NVIDIA's H100 in 2022 and holds roughly two-thirds of HBM4 allocation for the Rubin platform. But the gap is narrowing: Samsung passed NVIDIA qualification in early 2026 and supplies 25-30% of Rubin HBM4 plus 60% of Google TPU HBM3E; Micron is sold out through 2027 and targets ~25% share. Share shifts at platform boundaries — the 2027 contract round is the next real contest.

SK hynix trades at ~23x trailing but ~5.5x forward earnings — what does that spread mean?

The trailing/forward P/E spread is the market's real-time estimate of the cycle's mortality — how much of current earnings power it expects to die. SK hynix's ~4.2x ratio sits between Micron (3.4x) and Sandisk (6.9x): the market pays a trillion-dollar price while pricing most of the earnings as transient. The twist: hynix is the most contract-protected name in the complex — HBM sold out three years forward, price caps removed from long-term agreements. Either the contracts cannot survive a demand break (making NVIDIA at 2.0x the mispricing), or the mortality estimate is too high. The four multiples cannot all be right.

© Aya Research · About · invest.aya-ai.org · For information and research purposes only. Not investment advice.