The Record Quarter Was a Sell Signal — Memory's Old Timetable

On July 7, the day Samsung Electronics posted a record quarterly operating profit of KRW 89.4 trillion, its stock collapsed and the KOSPI tripped a circuit breaker (a safety mechanism that halts trading during a sharp drop). The headlines called it a "panic." But this is the old timetable of memory, a cyclical asset. This is not the first time a record quarter has read as a "sell" signal.

And a confession up front. This piece, too, cannot answer "sell now?" The stock is the first to speak of a cycle turning, and whether that speech is true stays unsettled by earnings while indicators rule on it late. What this piece can do is read the direction and set out what would confirm it — that far.

Memory prices don't climb like stairs; they rise and fall like waves. An asset whose demand and price swing hard between boom and bust — this is a cyclical asset. Memory is its archetype. Read the KRW 89.4 trillion crash as "panic or not" and you see half the event. Read it as a cyclical asset working normally and three layers appear: that the stock signals the turn first, that the signal can be wrong, and that its truth is therefore ruled on by indicators.

What Happened

On June 22, SK Hynix passed Samsung Electronics (KRW 2,088 trillion) intraday at a market cap of KRW 2,090 trillion, swapping the top-stock throne after 25 years and 7 months. A day later, on June 23, the KOSPI fell −910.71pt (−9.99%), its largest point drop on record. The year's fourth circuit breaker fired that day, and Samsung and Hynix each fell around −12%, their steepest declines since the 2008 financial crisis.

One thing to flag. June 23 was not a pure cycle event. That day's trigger ran along four lines, and three of them had nothing to do with the cycle. The prior day's decline in New York tech, fears of a U.S. rate hike, and the possibility of a failed inclusion in the MSCI developed-market index pressed down together. The cycle engine cannot fully explain that single day. So the cycle narrative in this piece takes direct aim past June 23 at what comes after: July 7.

July 2 was a "Black Thursday." The KOSPI closed at 7,648.09 (−7.89%). No circuit breaker fired that day; only a sell-side sidecar (a buffer that briefly suspends the effect of program-trade orders during sharp futures moves) was tripped, the year's 15th. The trigger was Meta. The market read its signal that it would lease out its "surplus GPUs" externally as "the AI compute shortage is over = memory demand has peaked." That day SK Hynix logged −14.57%, its largest single-day drop on record, beyond 2008 (−12.73%). The next day, July 3, brought a +5.76% V-shaped rebound, but foreigners alone were net sellers for a second straight session.

Then on July 7, Samsung Electronics disclosed Q2 preliminary operating profit of KRW 89.4 trillion (+1,810.3% year over year, a record). The KOSPI settled −4.91% that day and the year's sixth circuit breaker fired. Samsung fell to KRW 296,000 (−6.92%), breaking below "300,000-won Samsung." The following day, July 8, the KOSPI slid again, −5.35%.

Yet on July 8 signals of a different grain also appeared intraday. Early in the session the index rose to 7,791.66 (+1.77%), opening higher before giving it back in the afternoon; foreigners net-bought +KRW 331.6 billion, turning direction after 14 trading days; and the won/dollar rate reversed sharply to KRW 1,498.5 (−29.7), returning to the 1,400s for the first time in about two months. These three signals return in the closing section.

Why the Crash on a Record-Profit Day

Good earnings becoming a bad stock comes down to a defining property of asset prices. A stock is not a value that confirms a settled past. It is a value that prices in future cash flows in advance. Earnings settle "how much has been made so far"; the stock discounts "where the cycle goes next" and sets a price. So for a cyclical asset, the stock moving ahead of earnings is the rule. Not the exception.

Sources: Samsung Q2 preliminary disclosure (f18) · cyclical-structure interpretation (f37) · as of 2026-07-07

What the market prices in is not the settled KRW 89.4 trillion. It is the cycle direction that comes after. So for a cyclical asset like memory, a record quarter, far from a buy signal, tends to become the stamp confirming the stock's peak. This crash is the archetype.

The reading that this is "because of rates" or "New York's shock spreading over" is rebutted by the timeline. The overnight New York session for July 7 (July 6) had, if anything, the Nasdaq up +1.1% and the Dow crossing 53,000 for the first time. Then, once Samsung's earnings were disclosed, that same session in New York saw the Philadelphia Semiconductor Index (SOX) collapse −5.5%, Micron −7.3%, AMAT −10%. Read the order and the fuse is clear: a rising New York came first, and then Samsung's earnings dragged the index down.

Here's the snag. That the stock leads is a strong proposition. But the stock does not always lead correctly. It can lead and be wrong. So the signal the stock gives requires a test. History shows why.

The Old Timetable

Set the proposition — that stock leads earnings — on the data, and the sample splits like this, unairbrushed.

Sources: NASDAQ · KRX · each company's IR/disclosures (f30–f36) · composite calculation f37

Two leading, one simultaneous, two decoupling, one uncertain. Within this sample, the count of stocks that waited for the earnings peak and turned only afterward is 0. But don't misread that zero. The absence of lagging cases does not "prove the stock always leads correctly"; it only "disproves that the stock lags earnings." It rules out a direction without proving the accuracy of the lead. And with only two cycles in the sample, it cannot vouch for anything outside them.

The lag opens up more honestly. The direction is always the same, but how many months ahead runs irregularly, 0–18 months. In 2018 it was a 3–10-month lead; in 2021 Micron was effectively simultaneous. There is a timetable but no timestamp.

Above all, 2021 is the problem. That year the train ran onto an entirely different track. Samsung's stock peaked in January 2021, yet consolidated operating profit kept climbing for five more quarters, peaking only in Q2 2022 at KRW 14.1 trillion. The gap between stock peak and earnings peak is about 18 months. This is less a lead than a decoupling (the phenomenon of two values that moved together splitting off in opposite directions) made by the early-2021 liquidity rally. SK Hynix, too, saw earnings rise through Q2 2022 after its early-2021 stock peak, then turn to an operating loss of −KRW 1.7 trillion two quarters later.

The lesson of 2021 is not a counterexample to shove into a footnote. It is, rather, the heart of this piece. The stock can diverge even from the cycle. So this piece stands not on "the stock always leads correctly" but on "the stock speaks first, yet it can lead and be wrong." Precisely for that reason, a separate device is needed to test the signal the stock gives.

Why Samsung Fell Further

Even in the same crash, the two companies were hit in different spots, and conventional wisdom trips on the split. The wisdom goes: "Hynix, with its larger exposure to HBM (high-bandwidth stacked memory attached to AI accelerators), would have been hit harder in this crash." The actuals flipped it. Over July 7–8, Samsung fell about −13% and SK Hynix about −11.7%; Samsung, the party that reported, was cut deeper.

The substance of the split doesn't resolve along an exposure ranking. It is a difference of composition. SK Hynix is a pure-play on HBM. It leads with a 58% HBM market share (by shipments), and its customer concentration is, if anything, dispersing. Nvidia's share of revenue came down from 27% in H1 2025 to ~14.8% in Q1 2026, and a first new major customer breaking 10% has attached.

Samsung is different. Nearly all of its profit is a single memory engine. On a settled Q1 2026 basis, semiconductor (DS) division operating profit of KRW 53.7 trillion was about 94% of the company. On top of that sits exposure to non-memory losses. Q2 foundry and System LSI are estimated at about −KRW 3.2 trillion in losses (a brokerage estimate, before the July 30 settlement), and memory of +KRW 93.2 trillion is estimated to have offset that loss. A composite portfolio takes a discount relative to a pure-play.

So Samsung's excess drop does not reduce to a single cause. Several piled up: sell-the-news after a +150% rally (the pattern of selling into gains on the day a positive is confirmed), worry over KRW 400-trillion cluster capex, the composite-portfolio discount, a margin gap (Samsung 52% vs. Hynix 72%), and HBM qual (customer quality certification) uncertainty. Samsung has completed its HBM3E 12-high qual and has a green light on entering the HBM4 supply chain, but the approval timeline is still unsettled. These worked together.

So the phase decides who is hit more. In the demand-peak shock (July 2, Meta), the pure-play Hynix was hit hardest (−14.57%); in the earnings-confirmation phase (July 7–8), Samsung — the reporting party and a composite portfolio — was hit in excess. It splits by phase. If the cycle actually turns, this divergence too comes out of composition. Hynix gets hurt when HBM demand reverses; Samsung when a memory slowdown overlaps its non-memory losses. (This is not a call to buy or sell a name. It is only the grain of the direction.)

Why Korea Is Cut Deeper Than the U.S.

That the split among names amplifies into the index's drop is due to index structure. Samsung Electronics and SK Hynix together are about half of KOSPI market cap. That surged from about a quarter at the end of last year. It is a structure in which a single memory shock concentrates straight into the whole index. In the U.S., semiconductors are not this concentrated in the index.

The KOSDAQ was pushed down deeper than the large caps. Its drop over this stretch (July 1, 929.35 → July 8, 785.00) was −15.5%, deeper than the KOSPI's −12.7% over the same stretch. Two settled observations stand out. One is a shift in the sector landscape: with semiconductor equipment-and-materials names surging (Jusung Engineering 63rd → 5th, and others), the semiconductor sector's market cap overtook healthcare for the first time. The other is a concurrent weakness in bio: KOSDAQ150 Healthcare is −23.09% year to date. But this bio weakness should be seen as an independent variable on a separate axis from the semiconductor cycle. Absorb it into the cycle narrative and you pin the cause wrongly. As background for the KOSDAQ's dual fragility, the point that equipment-and-materials names track the three companies' capex (a qualitative observation; a quantitative linkage figure is not secured) and the KOSDAQ's rate-sensitive character are raised, but where there is no quantitative basis I won't nail down causation.

The Narrative Ran Ahead of the Indicators

What drove this crash was not the indicators. It was the narrative. Even the year's most favorable positive (early-July's employment shock easing the fear of a rate hike) couldn't save the AI-crowded assets. The market is pricing past the present (rates, earnings) into the future (the possibility of an AI cycle peak). The core: the narrative (the peak warnings from Meta, DeepSeek, Morgan Stanley) turned before the actual indicators and pushed the stock down. But here the narrative is a channel where the market buys and sells the story first — not the efficient price-discovery channel where public data sets the value. This distinction matters again in the test system we're about to set up. However early the narrative turns, whether it is a real cycle turn has to be separately confirmed by public data.

Nor is a gap between stock and fundamentals always a cycle signal. While Nvidia — the sturdiest of fundamentals — lagged at about +3.2% year to date, Micron rose +305%. To keep from reading even this gap as the cycle, telling true signals from false ones must, in the end, be left to a test of public data.

So, Ahead — Correction, or Entry into Descent?

Earnings (KRW 89.4tn) are not the answer. Being lagging, they can't tell what's ahead. History, too, gives no "when." The lag is unpredictable.

The stock signaled the turn first this time as well. But as in 2021 that signal can be wrong, so whether it is true is ruled on by indicators. Here the role of the indicators must be set precisely. These indicators are a test system. They are devices to confirm whether the signal from stock and narrative is a real cycle turn or noise — not a real-time trading signal. This is the place where the public-data price discovery I set apart earlier goes to work.

The test system is two-tiered by speed. At the top is one medium-speed check that lights up next-fastest after the stock, and below it are the three slow confirmations that are slower than the stock. Each of these four is a tester. The contract price of commodity DRAM (DDR5), negotiated and disclosed each quarter, is the medium-speed check. Because it is fast-cadence data, if the stock signal is true, the historical pattern says it turns here first. It ran +90–95% in Q1 and +58–63% in Q2 — still rising. Still to the upside.

Sources: TrendForce · Micron IR · each company's IR (f42 · f40 · f41 · f44 · f47) · as of 2026-07-08

There's a reason for attaching an observation channel to the table. The three slow confirmations differ in how fast they light up and in the window you look through. The HBM fulfillment rate is undisclosed, so it must be proxied by volume disclosures and backlog; Big Tech capex guidance opens only in the fixed window of Q3 earnings (late Oct–early Nov). This window sets the timing of the ruling later on.

The third tester reads in two grains. The three firms' official plans are aggressive expansion without output cuts, but suppliers' inventory sits at 3.3 weeks at the end of Q3, a cycle low, and TrendForce rated 2026 industry capex as "cautiously held · bit-supply growth limited." There's a temperature gap between plan and commentary, so read only one side and you misread this indicator.

Here the strongest objection has to be taken head-on. "HBM is different from past commodity DRAM cycles. It's locked in take-or-pay (a condition where you pay for the contracted volume even without taking the goods — like owing a cancellation fee even when you cancel a reservation) long-term contracts, so cycle pre-pricing doesn't work this time." Partly right. HBM has, until now, been priced on annual contracts, so the contract-price level reflects a commodity surge with a one-year lag. So on level alone, the contract price is no fast test signal. That far, granted.

But this buffer has begun to split by supplier. Into 2026 annual contracts are converting to 3-to-5-year long-term agreements (LTAs), and SK Hynix has removed even the price ceiling from those LTAs. In a structure where a market-price rise reflects immediately, the "one-year-lag buffer" thins out at SK Hynix (Micron keeps its floor and ceiling). And renewal negotiations can't be hidden. The HBM4 2027 contract negotiation opens in Q2, and if that volume and fulfillment collapse, HBM rides the cycle too. So the reading on tester (i) is shifted from "contract-price level" to "renewal-negotiation terms · volume." The equation "sold out through 2027, so upside confirmed" doesn't hold either. An order allocation is not a fulfillment guarantee, and the fulfillment rate is undisclosed. Nor is there evidence that HBM is cycle-immune. From Q1, the per-wafer yield of DDR5 64GB RDIMM has actually surpassed HBM.

Another objection. "This crash is just an exogenous shock touched off by macro triggers (rate fears, Nasdaq's fall, AI-overinvestment worry) and has nothing to do with the cycle." Separate the trigger from the engine. That macro touched it off is true. Macro factors were mixed into the June 23 trigger and the July 8 fuse. I already granted above that June 23 isn't fully explained by the cycle. But July 7 is different. On the day after New York's overnight session rose +1.1%, Samsung's earnings disclosure was the direct fuse. A macro risk-off (a wholesale flight from risk assets) doesn't explain this single day. Why, in the same session, the Dow and defensives (healthcare, utilities) held while memory and the Nasdaq alone were pushed deep — macro can't answer. The engine is that memory, near a cycle peak, priced the future downward.

The last objection aims at the argument's soft spot. "Isn't 'the stock leads the cycle' a post-hoc, convenient reading of a thin sample? Besides, you said the stock decoupled from the cycle in 2021." Right. The sample is thin, and the f37 actuals in fact bent the lag and the confidence. That this piece dropped "the stock always leads correctly" and stands on that humility is exactly why. The 2021 decoupling is not a counterexample that topples this logic. It is the ground that tells us why a test system is needed. And what turns this charge into something falsifiable is self-refutation ②, set out below. If the stock falls another −30% from here while the indicators hold to the upside, this piece confesses that its own cycle reading was noise. The −30% is not an arbitrary number. It is the halfway point of the −57% that Micron fell peak-to-trough in 2018.

Now back to the three signals of July 8. That day the first fine signal appeared in the fast channels. Foreigners turned net buyers after 14 trading days, the index tried a rebound to +1.77% intraday, and the exchange rate reversed sharply. But inflating this into a "crack" won't do. Foreigners' +KRW 331.6 billion is a small size against the roughly −KRW 7 trillion of cumulative net selling that had built up, and the KOSPI closed the day at −5.35% after all. That no circuit breaker fired on July 8 supports an easing versus July 7, but this isn't settled yet. It is only a hinge question. The start of a turn, or the start of a retrace? With indicators still to the upside now, which one it is stays open.

So I commit. Until early November, when the Q3 earnings season ends — Big Tech Q3 capex guidance comes out from late October into early November, so this is the ruling window — if fewer than two of the three slow confirmations light up as descent, this crash is settled as a pre-priced correction. If two or more light up, it is settled as entry into cyclical descent. Not "could be upside, could be downside," but a date (early November) and a threshold (two or more). As of July 8, 2026, all four are to the upside, so the data still supports the correction side.

The self-refutation, too, comes in two layers. One: if, by the ruling window, all three slow indicators hold to the upside and yet the cycle turns, then the testers I chose failed to catch the cycle, which means the choice of testers itself was wrong, and this piece's argument is rejected. There is no excuse of "the testers missed it but the argument holds." Because the testers are the argument. Two: if the stock falls another −30% from here (the −57% halfway point above) and yet the four indicators hold to the upside and the cycle does not turn, then stock and narrative were noise, and this piece's cycle reading was overreach.

To speak honestly, there is something this piece can't do. Indicators give only direction and after-the-fact confirmation. The answer to "sell now?" this piece cannot give. The stock is issued first and earnings arrive late; an investor who bought at the top now has already received the bill, and if the cycle actually turns, the next cost is borne by the three firms' earnings, capex, and employment — and by the downstream industries that carry the price of memory. It is the very structure by which, in 2021, SK Hynix turned from an earnings peak to a loss two quarters later.

The July 4 daily's open question — "the memory demand peak carries over to Samsung's July 7 preliminary earnings" — is answered by this piece. Earnings were not the answer. This old timetable turns once more this time as well. The answer is spoken first by the stock, and whether that speech is true is ruled on by indicators.

This topic from other angles: Collusion? A Shortage? The Real Story Behind the Memory Price Surge · What Ends the Cycle First Is Demand, Not Supply — Peak Margin Is the Starting Line of Mean Reversion · SK hynix Has Passed Samsung. The KOSPI Is Now Two Companies · The Most Famous Call Was the Most Wrong — A Post-Mortem on 2024 AI-Chip Price Targets · The Week Good News Split in Two — Records in New York, a Sixth Circuit Breaker in Seoul.