$38.1bn of New Fabs, and SK hynix Fell 4.88% the Same Day (August 2026): The First Cleanroom Opens in June 2029 — and the $93.9bn Already Locked at Floor Pricing

On 7 August 2026 SK hynix's board approved roughly ₩54 trillion — about $38.1 billion — for two new fabs: Yongin Y2 for DRAM at ₩35.2 trillion and Cheongju M17 for NAND at ₩19.1 trillion. The shares fell 4.88% the same day, and the KOSPI closed down 0.60% at 6,258.77. That is a fourth layer on the three this page already separates, and the cleanest one yet: in a commodity business, capacity is not a growth announcement, it is the mechanism that ends a price cycle — and the company published the calendar. Y2 breaks ground in July 2027 and opens its first cleanroom in June 2029; M17 breaks ground in February 2027 and opens in December 2028. The shortage SK hynix's own chief executive called the worst in the industry's history is scheduled for 2027. The relief arrives in 2029. Everything between those two dates is decided by pricing, by contracts already signed at floors, and — as the same session showed — by customers redesigning their products around memory they cannot get.

The positioning story below is still the right explanation for July's violence, and the contract-structure story is still the right explanation for the 5–6 August collapse. What 7 August added is the supply side of the same argument, on a published timetable rather than a forecast.

What actually happened on 7 August: ₩54 trillion committed, 4.88% lost

SK hynix's own announcement is short and almost entirely calendar. The board approved ₩35.2 trillion for Y2, a DRAM fab inside the Yongin semiconductor cluster whose phase-one infrastructure it describes as 99% complete, and ₩19.1 trillion for M17, a NAND fab at Cheongju. Both are executions of the mid-to-long-term strategy the company set out in June 2025 rather than a new idea, and the company's stated market assumption is that DRAM and NAND both compound at about 19% a year through 2030. The Korea Herald puts the two approvals inside a far larger frame: ₩600 trillion earmarked for the Yongin cluster over its full build-out, and a further ₩100 trillion for Cheongju.

The market's response was not to the money. The KOSPI closed 7 August at 6,258.77, down 0.60%, a second consecutive decline after the 4.58% collapse documented below. SK hynix fell 4.88% and was the single largest drag on the index; SK Square fell 3.20%. Foreign investors sold a net ₩863 billion, extending their selling streak, while institutions bought ₩579.1 billion and individuals ₩267.4 billion. And Samsung Electronics rose 0.22% in the same session — the detail that rules out a general risk event and points at something specific to the company that had just committed ₩54 trillion.

Why a capacity commitment reads bearish inside a shortage

This is the least intuitive mechanism on the page, so it is worth stating carefully. Capital expenditure in a commodity business is a signal about two different things at once, and they point in opposite directions.

Read forward, it is confirmation: a management team that believes the demand curve has structurally shifted spends against it, and ₩54 trillion is an expensive way to express a view. That reading is supported by everything the company has said. Kwak told Reuters that demand would exceed capacity “even beyond 2030”, and the announcement itself assumes 19% annual market growth to that date. An SK hynix official framed the decision as one “made to seize opportunities in line with the market's growth speed”.

Read backward, it is the beginning of the end of the price cycle — and the reason is arithmetic rather than sentiment. Today's memory prices are a function of capacity that was decided three to five years ago, when nobody in the industry expected this demand. The capacity being decided today arrives into a market shaped by every producer making the same decision at the same time, against the same forecasts, at the same peak in prices. The industry has run this loop before; what makes 2026 unusual is only the size of the numbers. South Korea's government has separately floated support of ₩400 trillion each for SK hynix and Samsung, which tells you the scale being contemplated across the sector, not just at one firm.

Neither reading is a forecast, and the honest position is that both are visible in the same announcement. What an equity price does with them depends on which horizon it is discounting — and the tape on 7 August says the market was discounting neither, because it was busy with something else entirely.

The other half of 7 August: fewer stacks per chip

The proximate reason SK hynix fell was a report that its largest customer may put less memory on each accelerator. TrendForce published the underlying note on 4 August: Nvidia had held 12-Hi HBM4e as the baseline design for Rubin Ultra from 2025 through the first half of 2026, and from the third quarter of 2026 widened its evaluation to include 8-Hi HBM4e, 12-Hi HBM4 and 8-Hi HBM4. No final specification has been determined. The stated reasons are supply-side, not demand-side: an overall DRAM shortage limiting how much wafer capacity can be allocated to HBM, and uncertainty about HBM4e validation schedules and yield ramp. The performance consequence is explicit — whether HBM4e validates determines roughly 14–16 Gbps of I/O speed versus 11–12 Gbps achieved by optimising HBM4 instead.

Here is why that matters more than its size suggests. Memory demand is units multiplied by content per unit. Almost every headline about the AI build-out is about units — how many accelerators, how many racks, how much capital expenditure. The Rubin Ultra story is about the other term. A shortage severe enough that the buyer redesigns the product around it reduces content per unit, and it does so precisely because the shortage is real. The shortage partially cures itself through the customer's bill of materials.

That is the same mechanism this page already documented at the opposite end of the market. Sandisk's Consumer revenue fell 32% sequentially because ordinary buyers could not pay the price; Rubin Ultra's memory configuration is under review because the largest buyer in the world cannot get the parts. Demand destruction at the cheap end looks like a shrinking segment line. Demand destruction at the expensive end looks like an engineering decision. Both reduce the bits the industry sells at the top of the cycle, and only one of them shows up in a revenue table.

The 5–6 August precedent: the biggest beat in the sector's history, sold immediately

Sandisk's fiscal fourth-quarter release of 5 August is worth reading as a document, because almost nothing in it is bad. Revenue of $8,965 million was up 51% on the prior quarter and 372% on the prior year, when the same business turned over $1,901 million at a 26.2% gross margin. Gross margin reached 84.6%. GAAP net income was $6,903 million, or $43.97 a share; the non-GAAP figure of $39.25 beat the company's own $30–33 guidance by a fifth. The board added $14 billion to the buyback authorisation, taking the remaining total to $15.5 billion. Chairman and chief executive David Goeckeler said the company had "established datacenter as a key growth pillar" and was positioned to "generate growing and durable free cash flow".

Then the sector fell. Sandisk itself dropped roughly 10% on 6 August, and Seoul followed: SK hynix −10.37%, Samsung Electronics −6.30%, KOSPI −4.58% to 6,296.38 from the prior session's 6,598.26, according to Seoul Economic Daily. The composition of that day repeats July's: foreign investors sold a net ₩3.3274 trillion and institutions ₩121.5 billion, while domestic retail bought a net ₩3.3388 trillion. The won barely moved, closing at 1,423.8 per dollar, down 0.7 won.

One detail separates this from an ordinary risk-off day. The KOSDAQ — Korea's smaller-company index, with almost no memory weight — rose 0.26% to 801.67 in the same session. This was not Korea being sold. It was two stocks being sold, and those two stocks are most of the index.

Why guidance that looks strong still reprices the sector

Sandisk guided fiscal Q1 2027 to revenue of $10.30–10.80 billion and non-GAAP earnings of $44.00–46.00 a share on about 155 million diluted shares. Published consensus figures reported around the print varied — roughly $10.8 billion at the low end of the estimates cited, and above $11 billion at the high end — but on any of them the guidance range topped out at or below where the street had been. Sequential growth of about 18% at the midpoint is enormous in absolute terms and a sharp deceleration from 51%.

The margin line is the more informative one. Gross margin is guided to 83.0–84.9% against 84.6% just delivered. Roughly a fifth more revenue, and the profitability of each incremental dollar is flat at best. In a shortage, margin expansion is the evidence that pricing power is still building; margin flattening at a record level is the evidence that it has been converted into something else.

That Consumer row is the most important number on this page now. In the same three months that datacenter demand doubled sequentially, revenue from ordinary consumers fell by nearly a third — and fell year on year too, in a business whose total revenue nearly quintupled. TrendForce's "affordability limit", quoted below as a forecast when this page was first written, has arrived as a reported segment result. High prices are not merely slowing consumer purchases; they are shrinking that revenue line outright.

The $93.9bn that caps the upside

The structural change is in how the industry now sells. Sandisk's release confirms it has signed ten New Business Model agreements — five announced in April, five more since, including three with new customers. On the earnings call management put numbers to the portfolio: agreements with eight strategic customers, a minimum of $93.9 billion of committed revenue assuming floor pricing, backed by $16.5 billion of financial guarantees in cash deposits and instruments, at a weighted average duration of more than four years. Those contracts cover more than half of fiscal 2027 bit shipments and roughly two-thirds of fiscal 2028.

Read that from the buyer's side and it is a triumph: four years of visibility in the most violently cyclical business in technology, with the customer posting collateral. Read it from an equity holder's side and it is a trade. The company has exchanged its exposure to the spot price for a floor. If NAND prices keep climbing through fiscal 2027, more than half of the bits do not participate at the market rate. The distribution of outcomes has been narrowed deliberately at both ends — and a narrower distribution priced off a record quarter re-rates downward, because the record quarter was the fat tail.

This is the read-through to Seoul, and it is why Korean memory fell on an American company's guidance. Samsung and SK hynix sell HBM the same way: multi-year, allocated, negotiated with a handful of hyperscale buyers rather than cleared on a spot market. The pricing mechanism that produced Samsung's ₩89.5 trillion quarter and SK hynix's 76% operating margin is being progressively contracted into fixed schedules. Nobody's demand has weakened. What is changing is how much of any future shortage the seller gets to keep.

The July setup: two sets of numbers that refused to agree

None of this began with Sandisk. Samsung's Q2 2026 release showed consolidated revenue of ₩171.5 trillion and operating profit of ₩89.5 trillion, both all-time highs, with the Device Solutions division supplying ₩89.2 trillion of that profit — essentially the entire company's earnings came from chips. Fortune put the group figure at about $62 billion, a nineteen-fold increase on the year. SK hynix's own release reported ₩60.5426 trillion of operating profit on ₩79.3187 trillion of revenue — a 76% operating margin, up 557% year on year — and said HBM4 mass shipments had begun and customer demand exceeded supply capabilities.

Now the tape. Across July the KOSPI fell 22.18%, its worst month since 2008, with Samsung down 21.89% and SK hynix down 34.47%. On Friday 31 July the index rose more than 17% in a single session, a move CNBC characterised as a record rebound from a meltdown, helped by renewed confidence in AI infrastructure spending after Microsoft's results. Then on Monday 3 August it closed down 5.13% at 6,257.41, both chipmakers off about 8.8%.

That table is the single most useful thing on this page. A −34% month inside a +516% year is not a business collapsing; it is a crowded position being cut.

The amplifier: what a 2x daily-reset fund is obliged to do

A single-stock leveraged ETF promises twice the daily move of one company. That word does all the damage, because holding a constant daily multiple against a net asset value that changes every day forces the fund to rebalance in the same direction as the market — selling exposure after a fall, buying it after a rise. Two consequences follow, and neither requires anyone to have an opinion about HBM4.

The first is volatility decay. A stock that falls 10% and then rises 11.1% is exactly flat; a 2x daily fund tracking it falls 20%, then rises 22.2%, and ends at 0.80 × 1.222 = 0.978 — down 2.2% on a round trip that cost the underlying holder nothing. Chain a month of double-digit swings together and the divergence stops looking like a rounding error. That is how the Samsung KODEX SK Hynix Single Stock Leverage fund came to be down more than 80% from its 23 June peak while SK hynix fell about 34% in July.

The second is reflexivity. The rebalancing order is a market-direction trade in a single, heavily weighted stock, so the product does not merely track the move — it contributes to it. In an index where two names carry an outsized weight, a large enough leveraged complex on those two names becomes an index mechanism.

Korean authorities responded by tightening access rather than withdrawing the products. Investors must now post a minimum cash margin of ₩300,000 — about $203, roughly ten times the previous effective threshold near ₩30,000 — and proposals under discussion include cutting permitted leverage from 2x toward 1.5x or 1x. Financial Supervisory Service governor Lee Chan-jin had voiced regret about the original approval as early as 22 June: "Maybe I should have lain down on the floor to block it. I personally regret (I didn't)," he said, in remarks reported by Yahoo Finance. CNBC documented the retail reaction directly.

The second derivative, and where the cost lands

Memory is a commodity business, and commodity equities trade the rate of change of prices rather than the level of profits. On that measure the peak of the acceleration was first visible in forecasts and is now visible in reported results. TrendForce expects conventional DRAM contract prices to rise 13–18% quarter on quarter in 3Q26 and NAND flash 10–15% — still large increases, but a marked slowdown, and the stated reason is demand-side: record-high contract prices mean PC and smartphone customers "are reaching their affordability limit". Sandisk's 32% sequential drop in Consumer revenue is that sentence with a dollar figure attached.

The supply side has a wrinkle pointing the same way. HBM is what pulled capacity away from ordinary memory in the first place — it consumes roughly three times the wafer capacity of DDR5 per gigabyte. But TrendForce also notes that HBM wafer revenue was overtaken by DDR5 64GB RDIMM in 1Q26, pushing HBM profitability below conventional DRAM and giving suppliers a reason to reallocate capacity back. A shortage that becomes profitable to relieve is a shortage with a clock on it. None of which is the same as the shortage ending: Samsung's outlook still says it expects the market to remain undersupplied through year-end, and Micron guided its following quarter to $50.0 billion ± $1.0 billion at roughly 86% gross margin.

Meanwhile the cost is landing somewhere, and the most instructive line in Samsung's release is not the record — it is the loss. MX & Networks, the division that sells Galaxy phones, reported ₩33.2 trillion of revenue and an operating loss of ₩0.7 trillion, attributed to "elevated component cost pressures across the industry". One company, one quarter, one input: memory made ₩89.2 trillion on the sell side and destroyed the margin of its own device businesses on the buy side.

The same transfer is visible at Apple, where guidance rather than reported margin absorbed the shock, and Gartner estimated in February that surging memory costs would reduce global PC and smartphone shipments in 2026. The chain runs: AI capacity demand → memory reallocated to HBM → commodity DRAM and NAND repriced → device bills of materials rise → either margins compress or retail prices rise, which is where a goods-price impulse eventually meets the inflation data central banks read.

The channel to what you actually trade

The direct index channel is US-listed memory and its customers, not Seoul: Micron and Sandisk sit inside the US indices, and the AI-capex complex that buys HBM sits inside NAS100 and US500. The causality on 5–6 August ran that way explicitly — an American NAND maker's guidance set the price for Korean DRAM shares the following morning, not the reverse. That is also why the 30–31 July rebound in memory names traced back to a hyperscaler's results rather than to anything Korean, and why the capital-expenditure line in cloud earnings is the variable to watch rather than the KOSPI chart. The supply side of the same story is widening as well, with Chinese DRAM entering mainstream laptops at prices that no longer undercut the incumbents.

One structural change since July shortens that chain. SK hynix listed American depositary receipts on Nasdaq on 10 July 2026 under the ticker SKHY, pricing at $149 and raising about $26.5 billion in the largest ADR listing on record — a dual listing rather than a flotation, since the shares have traded in Seoul since 1996. The practical effect is that a Korean session no longer has to be imported through sentiment: the same news now prints in a US-hours instrument on the same day, which is why 7 August produced falls across the US-listed memory complex alongside the Seoul move rather than a day later.

The currency channel is thin and should be labelled as such. The won is not among the eight currencies the meter scores, and nothing here moves an interest-rate differential, a growth read or a terms-of-trade balance. The one live link is risk sentiment — one of the five factors — the conduit that carries equity shocks toward the yen and the Swiss franc and away from the Australian and New Zealand dollars. But apply the lesson from the mechanism: an index move sourced from a margin requirement is a weaker risk-sentiment input than an identically sized move sourced from a demand shock. The size of the candle is not the size of the information.

What would change the picture

Seven observable things, in rough order of how much they would tell you:

1. Whether Samsung and Micron match the capacity commitment, and on what timetable. One producer building is a company decision; three producers building against the same 2029–2031 window is an industry deciding the shortage has an end date. Groundbreaking dates are public and unambiguous in a way that capex guidance is not.
2. The final Rubin Ultra memory specification. TrendForce has the evaluation open across four configurations with nothing settled. Whichever is chosen sets content per accelerator for the highest-value memory in the market, and it is one of the few AI-demand variables with a discrete, announceable answer.
3. The volume-versus-price split in the next quarter's results. Sandisk disclosed a one-third/two-thirds break for Q4. If the price share of growth keeps rising while volume stalls, the cycle is late whatever the headline says. If volume takes over, the shortage is genuinely broadening.
4. How much of supply gets contracted, and at what floors. More than half of Sandisk's fiscal 2027 bits are already committed. Watch whether Samsung and SK hynix disclose comparable coverage — every bit sold forward at a floor is a bit that cannot re-rate on the next price spike.
5. Consumer and PC-facing segment revenue, not just datacenter. A shortage that destroys the demand it prices produces a smaller total market. Sandisk's Consumer line fell 32% sequentially; watch whether that spreads.
6. Hyperscaler capital expenditure guidance. Memory demand is a derivative of AI build-out budgets. Committed spending is the demand curve; anything else is inference.
7. The leveraged ETF complex itself. Assets, permitted leverage and margin thresholds determine how much mechanical amplification remains in the next move. A smaller complex means a quieter tape for the same news.

What none of those tell you is where any of these securities go next, and that is the point of separating the layers. By 7 August there were four of them stacked on the same tape: accounting, which produced the records; positioning, which produced a 22% month and a 17% day; contract structure, which quietly decided how much of any future shortage the sellers get to keep; and now capacity, which set the date the shortage stops being a supply problem — June 2029 for the first DRAM cleanroom, more than two years after the year the company itself expects to be the tightest on record. Only the first was in an earnings release. The last two are the ones that will still matter in a year. Read more about how this site approaches that separation on the about page.

Educational macro context only — not investment advice.