Research OS on Korea Invest Insights

In an Earnings Bubble, Estimates Get Cut Last: AI Infrastructure Crowding and the Value of a Deep-Dive

Sun, 31 May 2026 14:00:00 +0900

This is a methodology note. For the underlying pieces that fed the analysis, it helps to read them alongside: the AI substrate/PCB thesis (the system BOM’s common bottleneck), Goldman’s token demand vs. J.P. Morgan’s memory ASP peak-out, and the Thesis OS public note that explains the structure running all of this work.

TL;DR

In a recent report, BCA Research argues that the AI bubble is not a valuation bubble but an earnings bubble. It is not the P/E that swells but earnings themselves. And like every bubble it deflates eventually, though BCA adds that its own AI demand gauges show no imminent signal yet.
The defining feature of an earnings bubble is the lead-lag. In BCA’s words, in almost every case “stocks began to fall well before profit estimates were cut.” Consensus estimates are a lagging signal.
So precisely in a phase where money is crowding into AI infrastructure like this, what matters more is a deep-dive that reads system structure and leading demand indicators directly, instead of waiting on consensus EPS. Once estimates are cut, it is already too late.
This note sets out, without exaggeration, what such a deep-dive actually looks at, and how we have run it as a structure called Thesis OS.

1. What it means to call the AI bubble an “earnings bubble”

When people say “bubble,” they usually picture P/E ratios shooting up — a valuation bubble, where price rises far faster than earnings. BCA Research sees AI as a somewhat different kind. It is an earnings bubble, where earnings themselves swell rather than price.

This is not a new pattern. Homebuilders and banks did exactly this just before the financial crisis. Their P/E ratios looked low, but only because unsustainable earnings inflated the denominator (E) and made the multiple look cheap. Industries that swing hard with boom and bust — natural resources, airlines, shipping, and today’s semiconductors — are vulnerable to this kind of earnings bubble.

Right now the semiconductor revenue curve resembles that picture.

Source: an approximate reconstruction based on public WSTS annual aggregates, with 2025-2026 as estimates. Illustrative, not investment advice. The shape of the underlying data is in the same vein as the “parabolic semiconductor sales” chart presented in the BCA Research report (2026-05-28).

A revenue curve going parabolic is both good news and a warning. When earnings rise fast, the P/E looks low. But if those earnings are a product of the cycle, the very fact that the multiple looks cheap can become a trap. The old warning of cyclical industries applies here: “the most dangerous moment is when earnings are at their peak."

Make no mistake. Neither BCA nor we are saying “it deflates now.” BCA judges that its AI demand gauges — adoption rates, token spending, AI coding-tool downloads, GPU and memory prices — are mostly still at reassuring levels. The point is not timing but how this bubble behaves.

2. The real trap of an earnings bubble is the “lead-lag”

What makes an earnings bubble dangerous is not that it bursts, but the order in which it bursts.

The core point BCA makes is this: Wall Street analysts are poor at calling when an earnings bubble will deflate. And in almost every case, “stocks began to fall well before profit estimates were cut” (BCA Research, 2026-05-28).

Here is what that sentence means in practice, drawn as a picture.

This is a conceptual diagram, not actual data. It simplifies the lead-lag structure BCA described, in which the price leads and the estimates lag.

The red line (the share price) turns down first. The blue dotted line (consensus earnings estimates) is cut only much later. The gray band in between is the lag. If you hold a rule that says “I’ll sell when analysts cut their target price or estimates,” you will always move late by exactly that lag.

The conclusion follows. Consensus estimates are a lagging signal. The stock does not look most expensive when earnings peak, and by the time estimates are cut the price has already fallen. So if you watch only the estimates, you miss both the peak of the bubble and its inflection point.

3. That is why a deep-dive is needed — what does it look at

If estimates lag, what should you watch to get ahead? A deep-dive looks not at headline EPS but at the structure and leading indicators that produce that EPS. Concretely, four things.

① It reads the system structure. The linear story — “after GPU comes memory, then substrates” — is easy to trade but only half right. Real AI infrastructure is a rack-level system in which GPU, CPU, DPU, NIC, switch ASIC, memory modules, and power boards all scale together. As we laid out in the AI substrate/PCB thesis, substrates and PCBs are not the final stop of a rotation but the common denominator of the entire system bill of materials (BOM). Knowing the structure reveals where the true bottleneck is.

② It separates the variables. Looking at the same AI demand, Goldman tracks token usage (Q) and cost per token (C), while J.P. Morgan tracks the rate of rise in memory prices (P). Decompose the two forecasts into P, Q, and C and it becomes clear that the two views, which looked like a clash, are actually talking about different variables and can hold at the same time. Lumping it all into a single headline number is what hides this.

③ It tracks leading indicators directly. Instead of waiting for consensus EPS to be cut, it watches what moves earlier — HBM long-term contract prices and volumes, server DRAM contract prices, token spending, GPU and memory spot prices, adoption rates. These change direction ahead of the estimates.

④ It separates fact, inference, and speculation. “Officially confirmed fact,” “reasonable inference,” and “mere speculation” do not go in the same box. Things that are unverified — customer names, whether a part was adopted, contract terms — are clearly flagged as inference or speculation. Without that separation, you get swept up in an attractive story and buy speculation as if it were fact.

These four do not wait for estimates to be cut. That is why they suffer less from the lag.

4. Thesis OS — the structure that runs this deep-dive as a system

Doing the four things above well once or twice is not hard. The hard part is doing them every time, with the same discipline. So we entrust this work to a structure rather than to a person’s mood on a given day. That structure is Thesis OS.

Thesis OS is divided into three roles.

Role	What it does
Alpha (알파)	Evidence gathering — market data, screeners, crawlers, fact-checking pipelines
Lattice (격자)	Judgment — weaving evidence into a thesis, building forecasts, stress-testing with the counter-case
Arki (아키)	Governance — keeping the whole consistent through schemas, workflows, and health checks

The point is not flashy automation but the repeatability of discipline. Separating evidence (Alpha) from judgment (Lattice) reduces the chance that a good story runs ahead of the evidence. With governance (Arki) in place, you split fact, inference, and speculation by the same standard each time and keep tracking the leading indicators. Thesis OS is open source, so interested readers can inspect the structure itself directly.

5. Our blog’s work — plainly

We write this as a record, not a boast. The most honest evidence is what pieces this methodology actually produced.

Mapping the system structure: the AI substrate/PCB thesis — seeing AI as a rack-level system and redefining substrates as the common bottleneck.
Separating the variables: Goldman vs. J.P. Morgan — decomposing two seemingly opposed forecasts into P, Q, and C.
Earnings read-through: Marvell Q1 FY2027, Dell Q1 FY2027 — translating U.S. earnings into Korean component and materials bottlenecks.
Tracking the cost structure: AI token futures and cost per token — the axis shifting from a performance race to a cost race.

What these pieces share is that they do not rush to a “buy/sell” conclusion. Instead they map the structure, separate the variables, present the leading indicators, and treat names not as recommendations but as observation points. The aim is to give readers material to judge for themselves. We do not claim to be able to call the market top or the moment the bubble bursts. As BCA concludes, even analysts are poor at that. What we are trying to do is more modest: to understand the structure before estimates are cut, and to make ourselves watch leading signals rather than lagging ones.

6. Closing

In a phase where this much capital has crowded into AI infrastructure, the most dangerous posture is to wait for the consensus to turn for you. In an earnings bubble that signal always arrives late. The share price moves before estimates are cut.

So a deep-dive is not a flashy forecast but preparation to be less late. Understanding the system, separating the variables, watching the leading indicators directly, and distinguishing fact from speculation. To repeat that work not once but every time with the same discipline, we use a structure called Thesis OS. If you are interested, we encourage you to look not only at the conclusions but at the structure and process behind them.

This piece briefly cites, with attribution, the published core argument of BCA Research’s “Earnings Bubbles Are Still Bubbles” (Global Investment Strategy, 2026-05-28), and the charts are self-produced based on public data and concepts. It is not a recommendation to buy or sell any particular security; investment decisions and their consequences rest with the investor.

Goldman's Token Demand vs. J.P. Morgan's Memory ASP Peak-Out: Do the Two Calls Really Clash?

Sun, 31 May 2026 11:00:00 +0900

This piece follows on from AI Token Futures and Cost per Token, Samsung Electronics 2026 Deep Dive, SK Hynix: The HBM Heavyweight, and Korea AI Infrastructure After Nvidia Q1 FY27. Where those earlier posts looked separately at cost per token, individual companies, and the spread of AI infrastructure, this one merges two different calls — Goldman on demand and J.P. Morgan on memory pricing — into a single framework and lays out the investment timeline from 2026 to 2030.

TL;DR

Goldman Sachs sees an explosion in token usage (24x by 2030) plus a steep drop in cost per token (60-70% per year). J.P. Morgan sees the year-over-year (YoY) growth rate of DRAM and NAND ASP slowing from 2027. Because they track different variables, the two calls do not clash head-on.
Combine them and a single path appears: 2026 is memory ASP beta, 2027 is a peak-out in price momentum, and 2028-2030 is a rotation of alpha into the components and systems that lower cost per token.
So the investment conclusion is not “buy memory without limit.” In 2026, memory beta such as HBM, server DRAM and eSSD is favorable, but after that you have to be selective about where the bottlenecks are — ASIC, AI networking, optical, HBM leaders, eSSD, advanced packaging, and MLCC/FC-BGA.
The two most common misreadings are: (1) seeing J.P. Morgan’s peak-out and concluding “the AI infrastructure cycle is over,” and (2) seeing Goldman’s demand and concluding “memory prices keep surging through 2030.” Both go too far.

1. Two Giants, Two Seemingly Opposite Calls

Looking at the same AI era, two global investment banks have painted pictures that seem to point in opposite directions.

Goldman Sachs (official article, May 5, 2026) looks at the demand side. The core is two points.

Token usage grows 24x by 2030. The call is that monthly usage reaches 120 quadrillion tokens in 2030; working backward, that puts the 2026 base at roughly 5 quadrillion. That is about 121% average annual growth over four years.
At the same time, inference cost per token falls 60-70% per year. The drivers are improving chip efficiency and better data-center architecture.

Source: a reconstructed chart that simply indexes the figures in Goldman Sachs’s official article (2026-05-05). It is not the original chart but the original numbers — “24x tokens by 2030, cost per token down 60-70% a year” — plotted on a log scale.

In the chart above, the blue line (usage) climbs steeply, and the orange and green lines (cost per token) fall even more steeply. At a 60% annual decline, cost four years out is about 2.6% of the 2026 level (roughly a 39x improvement); at 70%, about 0.8% (roughly a 123x improvement).

By contrast, the J.P. Morgan material looks at price. The picture is that DRAM and NAND ASP rise strongly in 2026 but that, from late 2026 into early 2027, the year-over-year growth rate decelerates quickly.

Source: a session-uploaded image (DRAM from WSTS / J.P. Morgan estimates, NAND from Gartner / J.P. Morgan estimates). Detailed figures such as FY26 DRAM +53% and NAND +30%, FY27 DRAM +1% and NAND -6% are based on a secondary summary; the original table is unverified.

On the surface, “demand explodes” (Goldman) and “memory price gains roll over” (J.P. Morgan) seem to clash. But look closely and the two calls are talking about different things.

2. Why It’s Not a Clash — Separating P, Q and C

A memory company’s profit breaks down into a simple multiplication.

Revenue = shipments (Q) × average selling price (P), and the token economy = usage (Q) × cost per token (C).

Through this lens, it becomes clear that the two calls are watching different variables.

Dimension	Goldman Sachs	J.P. Morgan
Variable watched	Token usage (Q), cost per token (C)	YoY change in DRAM/NAND ASP (P momentum)
Core message	AI usage explodes long term, cost per token plunges	Memory price growth rate slows from 2027
Time horizon	2026-2030	Late 2026 to early 2027

There is a trap that must be flagged here. The vertical axis of J.P. Morgan’s chart is not the absolute level of price but the “year-over-year growth rate (YoY %)." The two are entirely different.

Take an example.

Point in time	ASP index	YoY
4Q25	100	–
4Q26	300	+200%
4Q27	315	+5%

From 4Q26 to 4Q27 the price index still rises, from 300 to 315. Yet the year-over-year growth rate plunges from +200% to +5%. In other words, J.P. Morgan’s 2027 “peak-out” may not mean prices collapse but that the pace of increase slows. The stock market usually prices in this direction of the growth rate before it prices in “record-high profit” itself. That is why share prices can cool first, right when profit is at its peak.

To sum up, Goldman watches Q (usage) and C (cost), while J.P. Morgan watches the pace of P (price). Because these are different variables, both can be right at the same time.

3. Total Inference Cost May Actually Fall

Here a counterintuitive conclusion emerges. Even if usage grows 24x, if cost per token falls 60-70% per year, then the total inference-cost burden could actually be lower than in 2026.

It is simple arithmetic. Assuming the same token mix and the same model complexity, total cost in 2030 is the usage index × the cost-per-token index.

At a 60% annual decline: 24 × 2.6% ≈ 61% of 2026
At a 70% annual decline: 24 × 0.8% ≈ 19% of 2026

This is a heavily simplified sensitivity calculation. It does not reflect token mix, model size, context length, the rise in reasoning tokens, multimodality, redundant processing, latency constraints, or memory-bandwidth bottlenecks. In practice these factors could push cost back up.

This chart is the heart of Goldman’s logic. Whoever lowers cost per token ends up making the money. More than the rise in usage itself, it is the cost reduction that makes that usage affordable which keeps the industry’s cash flow alive. That said, as the caveat notes, if context length or reasoning tokens rise quickly, the real cost curve may fall less than this chart suggests.

4. So the Timeline Shifts

Combine the two calls and the following path is logically consistent.

Period	What happens	Where it’s favorable
2026	Agent usage surges → HBM, server DRAM, eSSD and NAND allocation tightens → DRAM and NAND ASP spike	Memory ASP beta works strongly
2027	A higher base plus some supply additions and long-term agreements (LTAs) slow the YoY ASP growth rate. B2B AI memory stays firm; B2C consumer memory hits price resistance	Share prices price in “slowing growth rate” before “peak profit.” Segment divergence begins
2028-2030	Usage keeps rising, but the drop in cost per token absorbs it	Alpha rotates from broad memory beta to the token-cost-reduction stack

The core message is one. From 2027, “how much further DRAM and NAND prices rise” matters less than “which components and systems lower cost per token."

5. Example Investment Ideas (Observation Points, Not Recommendations)

Below are not stock recommendations but examples of “where to look first” along the timeline above. AI and memory stocks have already re-rated quickly, so this is not a call to buy right now but a map for when the conditions are met.

Example 1 — 2026 Memory ASP Beta

The phase where HBM, server DRAM and eSSD allocation tightens and DRAM and NAND prices spike. Memory large-caps such as SK Hynix, Samsung Electronics and Micron benefit directly. The counterargument is that once the 2027 slowdown in the growth rate begins, the share price can take multiple compression first, even if profit is high.

Example 2 — Token-Cost-Reduction Stack

The real heart of Goldman’s logic is not the rise in usage but the fall in cost per token. Customers will pay more for chips and systems that lower their cost per token. Custom ASICs, AI networking and optical, and HBM leaders fit this best. (The interconnect, substrate and power bottlenecks seen in the Marvell read-through are this same thread.)

Example 3 — eSSD / Enterprise NAND

Agent inference demands far more retrieval (RAG), logs, context, KV cache and checkpoint storage than plain training. If it is true that AI servers use roughly 3x the SSD capacity of ordinary servers, then NAND can be reclassified not as a plain commodity-cycle asset but as an AI storage bottleneck. The counterargument is the possibility that price resistance in consumer SSDs dilutes enterprise strength.

Example 4 — Advanced Packaging / MLCC / FC-BGA

If Goldman’s 2030 token call is right, the complexity of server and rack architecture keeps rising. It is not only GPUs, ASICs and HBM that grow; demand for substrate area, power stabilization, decoupling capacitors and high-speed signal quality grows alongside. High-spec MLCC and FC-BGA suppliers such as Samsung Electro-Mechanics fall here.

Different Peak-Outs by Segment

What matters is not “memory as a whole” but the differences by segment. J.P. Morgan’s peak-out logic does not apply equally to all memory.

Segment	Likelihood peak-out applies	Tension with long-term demand
HBM	Low-Medium	High (demand keeps reinforcing the bandwidth wall)
Server DRAM	Medium	Medium
eSSD / Enterprise NAND	Medium	High (structural demand possible)
Mobile DRAM	High	Low (fast consumer price resistance)
Commodity DRAM/NAND	High	Low (peak-out logic applies best)

Common condition: the examples above need more than “oversold” or “because it’s AI.” A candidate must be somewhere where actual orders, contract prices and earnings estimates are newly rising, and where bottlenecks and entry barriers are confirmed.

6. The Two Most Common Misreadings

In this phase the market is prone to two mistakes pointing in opposite directions.

Misreading 1 — “J.P. Morgan says peak-out, so the AI infrastructure cycle is over." No. A peak-out is a slowdown in the growth rate of price, not a collapse in the level of price. And it applies best to commodity memory and less to bottlenecks such as HBM, eSSD, ASIC and packaging.

Misreading 2 — “Goldman says 24x tokens, so DRAM and NAND prices keep surging through 2030 too." This too goes too far. Goldman says explosive usage and a steep drop in cost per token at the same time. If cost falls quickly, the YoY growth rate of memory prices can slow even as usage rises.

The right reading is in between. A memory super-cycle in 2026, a peak-out in price momentum in 2027, and a long-term expansion of the token-cost-reduction stack in 2028-2030.

7. Fund Manager’s Comment

The two calls are not enemies but two axes of one picture. Goldman speaks to “how much, and how cheaply, we come to use AI”; J.P. Morgan speaks to “how long memory prices keep rising fast in that process.” The investor’s job is not to pick one of the two but to not miss the inflection points on the timeline.

The two most dangerous choices are these.

Throwing out all of AI infrastructure on seeing the single word peak-out — the mistake of giving away even the winners with bottlenecks at a bargain price.
Chasing commodity memory to the bitter end on token demand alone — the mistake of taking multiple compression head-on in the phase where the price growth rate rolls over.

The signals to check now, put simply, are these.

What you watch	Goldman (demand) strengthens	J.P. Morgan (peak-out) strengthens
Token usage	High growth sustained	Growth rate slows
Cost per token	Continues to fall 60%+ a year	Decline rate slows, latency cost rises
HBM long-term agreements	Prices held or raised	Renegotiation, volume deferral
Server DRAM contract price	Further increase	Slowing growth rate, spot discounts
Enterprise SSD	CSP long-term agreements expand	Consumer SSD price resistance spreads in

In short, recognizing memory beta in 2026 while preparing, from 2027, to shift weight toward the bottlenecks that lower cost per token is the most reasonable stance available now. It is neither chasing memory to the end nor throwing it all out in fear of the peak-out.

This piece is an analysis reconstructed from Goldman Sachs’s official article (2026-05-05), J.P. Morgan-related charts and secondary summaries, and TrendForce’s official outlook. The full text of J.P. Morgan’s original report, its quarterly ASP tables, and its segment-level detail could not be verified, and company names are examples illustrating the analytical flow, not investment recommendations. Actual investment decisions and their responsibility rest with the investor.

How This Blog Is Made: Introducing Thesis OS, Our Open-Source Research Operating System

Sat, 30 May 2026 11:00:00 +0900

🔗 Go to repo: github.com/youngseongshin/thesis-investment-os — the open-source system that runs this blog’s research

Today’s post is a little different from the usual. It isn’t about a stock — it’s about how the posts on this blog actually get made. Let me pull back the curtain for a moment.

What it takes to produce a single post

The posts on Korea Invest Insights are not improvised by a person staring at a blank screen. Behind them runs a small operating system called the Thesis Investment OS. The name sounds grand, but the idea is simple.

Make investment judgment visible, testable, and honest about its own track record.

It is not an automated trading bot, not a signal-selling service, and not an “AI that picks stocks for you.” It is a framework that gathers fragmented market information into a thesis — and lets you go back later and check whether that thesis turned out right or wrong.

The structure breaks into three roles. Think of them as three people on one team.

1. Alpha — the one who gathers the evidence

Alpha is the role that collects and verifies facts.

Quantitative data: prices, volume, flows, fundamentals, filings
Qualitative data: news, filings, earnings transcripts, community signals
Narrowing down candidates with screeners, then layering on context to surface names worth watching

What Alpha produces is evidence records, market snapshots, intraday alerts, screener candidates, and research packets. In short, it is the one who honestly stacks up “what happened." It does not judge yet. It only gathers the raw material.

2. Lattice — the one who builds judgment from evidence

The name Lattice comes from Charlie Munger’s idea of a “latticework of mental models” — a mind built from many interlocking frameworks.

Its role is to take the material Alpha gathered and turn it into an actual investment decision.

Registering a thesis and organizing it into a decision card
Running a devil’s advocate review that deliberately argues the other side
Recording predictions in a prediction ledger, then revisiting them later to see if they held up

The structure you read on the blog — “here’s the conclusion,” “this is a fact and this is speculation” — comes straight from Lattice. The point is to make a call, but leave it in a form you can grade later.

3. Arki — the one who keeps the system running

Arki is the least visible role, and perhaps the most important. It is the one that keeps the whole system healthy.

Defining the schemas that hold the data and the vault layout that stores it
Managing recurring jobs and running health checks
Keeping migration logs and governing the permissions and rules of each role

If the system were a house, Arki is the one making sure the electricity, water and heating keep running while Alpha and Lattice do their work. It is not glamorous, but without Arki the other two wouldn’t last long.

What these three roles have produced — real examples

This is abstract in words, so here are two recent posts that came through this system.

Dell Q1 earnings and the Korea AI-server margin read-through — Alpha gathered Dell’s earnings numbers, and Lattice connected them into the Korean semiconductor and server value chain to build a view.
Marvell Q1 FY2027 results and the Korea semiconductor read-through — same flow: starting from Marvell’s custom-silicon numbers and carrying them into a Korea read-through.

Both posts separate “this is a Fact, this is an Inference, this is Speculation.” That habit is exactly the structure Lattice enforces, and the facts holding it up are the ones Alpha gathered.

Why publish this at all

When you do research long enough, the scariest thing is “not remembering what you said before." Good-looking theses are plentiful; going back to check whether they were actually right is tedious and uncomfortable. So most analysis is written once and forgotten.

Thesis OS deliberately builds that discomfort into the system. Every judgment gets evidence attached, every prediction gets logged, and everything gets graded later. Not because it is perfect, but because it is built so that when it’s wrong, you can see it.

The system is designed to run locally. You can try it with the bundled sample data — no API keys, broker logins, or paid feeds required. The license is MIT, and it needs Python 3.10 or newer.

And the three channels this system publishes through are exactly these: the blog (Korea Invest Insights) you’re reading now, Telegram @korea_invest_insights, and Substack.

Come take a look

The point of this post isn’t to brag — it’s an invitation. If you’ve ever wondered how to make investment research more honest and more testable, take a peek.

You don’t have to read all the code. Even skimming the README should give you a feel for “ah, so this is how these blog posts get made.”

👉 github.com/youngseongshin/thesis-investment-os

A star is welcome, but just browsing the structure is fine too. There is only one reason I opened the curtain: so you can see for yourself where and how this blog’s judgments come from.

Disclaimer: For research and information purposes only. Not personalized investment advice. The open-source system described is a research tool; readers are responsible for their own investment decisions and outcomes.

ADR Hits a Six-Year Low: Will the Megacap Chip Squeeze Tighten, or Is It Time for the Laggards?

Sat, 30 May 2026 10:30:00 +0900

This note follows up on Korea ADR at 67: The Index Holds, So Why Are Stocks Weak?, Korea Foreign Investor Flow Analysis, and Smart Money Returns to the KOSDAQ. Where the earlier pieces looked separately at market breadth (ADR), foreign flows into megacaps, and KOSDAQ liquidity, this one ties all three together to answer the practical question: “Will the concentration tighten, or will the laggards rebound?"

TL;DR

On May 29, 2026 the KOSPI set a fresh record at 8,476pt (+3.55%), but the same day’s ADR fell to roughly 52% — the lowest since March 2020, a six-year trough. A record-high index paired with nine of ten stocks falling is an extraordinarily narrow market.
The stars of this squeeze are Samsung Electronics and SK Hynix (the “megacap chip duo,” or what Korean traders call “Jeon-Nik”). The starting point isn’t a simple theme but a genuine memory earnings cycle — yet single-stock leverage ETFs piled a wave of speculative flow on top of it.
The core conclusion is singular. Don’t try to predict the unwind — confirm it with signals. The default right now is “concentration persists,” and laggards should be added only when ADR, turnover, foreign flows, and earnings estimates all turn together.
And the most important distinction: “riding the concentration” and “riding the concentration product (leverage ETFs)” are two completely different things.

1. What Is Actually Happening in the Korean Market

Start with the facts. ADR doesn’t measure how high the index is — it measures the breadth inside the market.

ADR = Advancing issues / Declining issues × 100

When ADR is 100, advancers and decliners are balanced. Well below 100 means “regardless of the index, most stocks are falling.” The roughly 52% reading on May 29 is a level not seen since the 2020 pandemic crash.

Item	As of May 29
KOSPI	8,476.15pt, +3.55%, all-time high
ADR	About 52%, a six-year low
Market breadth	A breadth collapse close to “nine of ten stocks falling”

The previous note showed ADR at 67; in the interval, the market’s internals narrowed further. In other words, the record-high index is not the result of a broadly strong market, but of a handful of megacaps doing the pulling.

Two Flows That Amplified the Squeeze

Two distinct flows magnified the concentration.

First, single-stock leverage ETFs. On May 27 alone, total turnover in Samsung Electronics and SK Hynix single-stock leverage ETFs was about 10.4 trillion won, against a combined market cap of roughly 4.9 trillion won. Dividing turnover by market cap gives a daily turnover ratio of about 207%. That signals short-term, event-driven flow rather than normal long-term capital. The combined turnover of the four KODEX and TIGER Samsung Electronics/SK Hynix leverage products alone came to about 9.42 trillion won.

Second, the MSCI rebalancing. MSCI applied its May review at the May 29 close, and domestic brokers estimated 1.2 to 1.4 trillion won of passive inflows from higher weightings for SK Hynix and Samsung Electronics. Money that simply tracks the index was structurally forced to buy more of the duo.

In short, today’s concentration is the product of real earnings + passive flows + leverage flows all aligning in one direction.

2. Two Interpretations

There are two broad ways to read this.

Interpretation	One-line summary	Rationale
A. The squeeze tightens further	Strong leaders run for a long time	Memory earnings cycle, passive and leverage flows, MSCI weight increases
B. It’s the laggards’ chance to rebound	A market this narrow snaps back	ADR at a six-year low; extreme breadth collapses don’t last

Both are partly right. A is correct in that “the strongest earnings momentum still sits with the duo,” and B is correct in that “extreme ADR lows have statistically been followed by rebounds.”

The real issue is what to do when both are true.

3. The Answer: Don’t Predict — Confirm

Bottom line first: no one can know in advance when the concentration unwinds. Strong leaders always run longer than expected. So abandoning the leaders too early on the reasoning that “rotation into laggards is coming soon” usually costs you.

Instead, judge the unwind not by “has the duo rolled over,” but by “has the rest of the market, ex-duo, gained the power to rise on its own?" Put plainly, these are the five things to watch:

Ex-duo relative strength: Does the KOSPI excluding the duo start to outperform the duo itself?
Broadening advancers: With the index holding, does the number of advancing stocks grow and does ADR recover above 70?
Broadening foreign flows: Do foreign net buys stop clustering in the duo and spread to other names?
Leverage ETF cooldown: Does leverage-ETF turnover cool while the underlying Samsung and SK Hynix shares hold firm?
Laggard estimate upgrades: Are the earnings estimates (EPS) of the neglected names being revised higher?

When three or four of these light up together, that’s when you can acknowledge a rotation. There’s no need to bet ahead of time with only one or two flashing.

The Most Important Distinction: “Riding the Squeeze” vs. “Riding the Leverage ETF”

Here’s a trap worth flagging. “I’ll ride the concentration in the duo” and “I’ll buy the duo’s leverage ETF” are entirely different statements.

A single-stock leverage ETF is not a stock — it’s a leverage product that resets daily. You’re buying not SK Hynix itself but “twice SK Hynix’s daily return,” and the longer you hold, the more the two become completely different assets.

These products carry negative compounding (volatility decay). Even before listing, Korea’s Financial Supervisory Service warned that “if a price repeatedly rises and falls, cumulative returns can diverge sharply from the underlying.” On the managers’ own simulations, holding a high-volatility name at 2x leverage for about three months often produces a loss of typically 8% to 12% on the leverage ETF even when the underlying is flat (0%).

The key is this: even if the squeeze doesn’t unwind and the duo merely trades sideways, the leverage ETF loses over time. On top of that, with no diversification the daily limit is a wide ±60%, and on June 28 the FSS opened an investigation into possible wash trades and overheating. In other words, it’s a product whose exit you can’t easily control.

So if you want to ride the concentration, the right approach is the underlying (cash) shares, not chasing the leverage ETF. Riding negative compounding — a “publicly disclosed inefficiency” — isn’t alpha; it’s just leverage.

4. Investment Idea Examples (Observation Points, Not Recommendations)

The below are not stock recommendations but examples showing “what conditions make a name a candidate.” This isn’t a call to buy now, but a map of where to look first when the five confirmation signals above light up.

Example 1 — The Next Bottleneck After GPU/HBM: Power, Substrates, Passives

As AI servers grow more complex, demand rises not only for GPUs and HBM but also for parts that stabilize power, substrates, and passive components. MLCCs, FC-BGA, and high-layer-count boards (MLB) are the prime examples. That said, some names here have already run a long way. One large component stock rose about 2.56x between late April and late May. So the point is not “what looks cheap” but “what is seeing fresh upward earnings revisions."

Example 2 — Traces of Foreigners’ Pre-Positioning

There’s one intriguing signal. On May 29, as retail dumped en masse and memory melted up, foreigners quietly net-bought some deeply beaten-down materials, parts, and equipment names. Wonik IPS, Leeno Industrial, and EO Technics were among them. Because domestic institutions mostly sold alongside retail, a “foreigners buying, domestic institutions selling” divergence emerged. Whether that divergence resolves with domestic institutions returning is the second confirmation point.

Example 3 — Where Does the Money the Leverage ETFs Bleed Go?

The most contrarian view is this: when everyone is buying the duo’s leverage ETFs, watch where the money those products bleed via negative compounding will flow. When leverage holders bleed in sideways and choppy ranges, that money ultimately has to flow back into more efficient exposure — index products, cash equities, and neglected quality names. This is a bet on direction rather than timing, which makes it relatively safe.

Common condition: all of the examples above need more than just “oversold.” A name only becomes a candidate when earnings aren’t impaired, estimates are being revised higher, and turnover and foreign flows are attaching for the first time — all at once.

5. The June Calendar Is the Real Variable for Rotation

The biggest enemy of a laggard rebound is, surprisingly, rates and liquidity. June is packed with heavyweight events.

Event	Date	Why it matters
9th nationwide local elections	June 3	Short-term political uncertainty
US May CPI	June 10	Whether inflation reignites
FOMC	June 16–17	Chair Kevin Warsh’s first meeting
SpaceX IPO (target)	June 12	Assumed to be a major liquidity-absorbing event

US April PCE inflation came in at +3.8% year over year, so the inflation burden is still not low. If June CPI accelerates further or the new Fed chair is read as hawkish, rates could rise and rotation into high-valuation laggards (especially high-PER component names) could be delayed.

And keep one statistical trap in mind. The claim that “the win rate of rebounds after extreme ADR lows is nearly 100%” omits that the sample is small and that most past cases occurred during rate-cutting regimes. We may instead be in the opposite regime, with inflation stirring again. You can’t mechanically transplant past win rates onto the present.

6. Wrap-Up — Fund Manager’s Comment

Back to the question: “Will the megacap chip squeeze tighten, or is it the laggards’ chance to rebound?” The honest answer is “both are possible, and right now we’re in the stage of waiting for the signal that separates them."

The two riskiest choices are:

Abandoning the leaders too early on the back of a low ADR alone — handing over strong earnings momentum at a discount.
Chasing an already-overheated single-stock leverage ETF too late — being dragged along with no exit by negative compounding and ±60% volatility.

So the most rational stance right now boils down to this:

Hold the duo via the underlying (cash) shares, do not chase the leverage ETF, and add laggards only after confirming that ADR, turnover, foreign flows, and earnings estimates all turn together.

Unwinding a concentration isn’t something you predict — it’s something you confirm. Not missing the moment that first confirmation signal lights up, using the five-point checklist above — that’s the most practical preparation you can make right now.

The market figures in this article are based on news and disclosures between May 27 and 29, 2026, and the named stocks are examples to illustrate the analytical flow, not investment recommendations. All actual investment decisions and their consequences rest with the investor.

Korea ADR at 67: Why the Index Can Hold While Most Stocks Are Weak

Wed, 27 May 2026 16:25:00 +0900

This note follows the complex risk-off and recovery-trigger framework, KOSPI foreign ownership versus Samsung and SK Hynix, Korea foreign-investor flow analysis and the National Growth Fund / KOSDAQ smart-money map. Those pieces looked at macro gates, KOSPI mega-cap flows and KOSDAQ policy capital. This one asks a simpler question: how healthy is the tape beneath the index?

TL;DR

Korea’s 20-day ADR fell from 113.1 to 67.3 over the last month. KOSPI fell from 116.3 to 68.5 and KOSDAQ from 111.6 to 66.7. The average stock is already in a correction even if the index still holds.
This is not broad risk-on. It is narrow leadership concentrated in AI infrastructure bottlenecks, MLCC / FC-BGA / SOCAMM / back-end semis, and selected shipbuilding / defense / power names.
The next trade is not chasing first-line leaders. It is watching for ADR recovery + rising turnover + first foreign / institutional flow into second-line candidates. The local screen points to HPSP, SFA Semicon, Hana Micron, Dongjin Semichem and KMW.

Source: Research OS local DB prices_daily + universe.csv. Data cut is the May 26, 2026 close. May 27 intraday data is not included.

1. What ADR Tells Us

ADR measures market breadth, not index level.

The formula is simple.

Daily ADR = advancing stocks / declining stocks × 100
20D ADR = sum of advancing stocks over 20 sessions / sum of declining stocks over 20 sessions × 100
Advance ratio = advancing stocks / total stocks × 100

An ADR of 100 means advancing and declining stocks are balanced. Below 80, declining stocks are clearly dominating. If ADR sits in the 60s or 70s while the index remains firm, the market is usually being carried by a few large caps or a narrow set of themes.

That is Korea now.

Market	2026-04-16 20D ADR	2026-05-26 20D ADR	Change	Read
KOSPI	116.3	68.5	-47.8p	Large-cap breadth has weakened
KOSDAQ	111.6	66.7	-44.9p	Small/mid-cap breadth has cooled sharply
All Korea	113.1	67.3	-45.8p	The average stock is already weak

The important point is that this is not only a KOSDAQ problem. KOSPI’s 20-day ADR is also down to 68.5. So the right read is not “KOSDAQ is weak but KOSPI is healthy.” It is this:

Korean market breadth has narrowed, and surviving capital is compressed into AI infrastructure, shipbuilding, defense and a few other leadership pockets.

2. Why the Late-May Bounce Is Not Enough Yet

On May 20, only 333 stocks rose while 2,082 fell. Daily ADR was 16.0 and the advance ratio was 13.5%. That was close to a short-term capitulation session.

The May 21-22 rebound was strong.

Date	Advancers	Decliners	Daily ADR	20D ADR	Advance Ratio	Read
2026-05-20	333	2,082	16.0	58.3	13.5%	Short-term capitulation
2026-05-21	1,695	720	235.5	61.9	68.6%	Technical rebound
2026-05-22	2,097	283	742.2	67.9	86.1%	Strong rebound
2026-05-26	749	1,660	45.1	67.3	30.2%	Breadth weakened again

The problem is May 26. After the strong rebound, decliners rose again to 1,660 and the advance ratio fell back to 30.2%.

So the current state is not confirmed breadth expansion. It is a bottoming attempt in breadth that is still being tested.

That links directly to the earlier macro risk-off framework. The condition there was that oil, long rates, the dollar, KRW, Chinese credit and foreign flows needed to stabilize together. The ADR data says that recovery has not yet spread through the market.

3. Current Regime: Narrow Leadership

The current Korean equity regime is Narrow Leadership / Selective Risk-On.

Item	Read
Market breadth	Weak. All-market 20D ADR is 67.3
Leadership	Very strong in AI infrastructure and selected shipbuilding / defense
Trading difficulty	High. Stock selection matters more than index direction
New exposure	Prefer pullbacks and second-line names over chasing first-line leaders
Portfolio stance	Hold relative-strength leaders, replace weak positions, avoid using all cash before breadth recovers

The mistake is to say “the index is holding, so the whole market is fine.” With ADR in the 60s, most stocks are not fine. The opposite mistake is to say “breadth is bad, so avoid everything.” That is also wrong, because capital is still very active in a narrow set of sectors.

The surviving groups are:

AI infrastructure bottlenecks: Samsung Electro-Mechanics, Jeju Semiconductor, Daeduck Electronics, Simmtech, Haesung DS, Hana Micron, HPSP
Memory mega-caps: SK Hynix, Samsung Electronics
Shipbuilding / defense / nuclear-SMR: HD Hyundai Heavy Industries, Hanwha Ocean, Doosan Fuel Cell
Power / optical / network: selected power cable, RF and optical names

The market is weak, but leadership is not. That is not a contradiction. When breadth collapses, capital often crowds even harder into the few themes that still work.

4. Where Leadership Actually Was

The last-month leadership list is compressed into AI infrastructure and shipbuilding / defense.

Returns are %, turnover and flows are KRW 100 million units.

Stock	1M	5D	Avg Turnover	Foreign 1M	Institution 1M	Retail 1M	Read
Jeju Semiconductor	+173.6	+28.4	3,692.8	+1,440.8	+595.8	-1,962.1	LPDDR second-order discovery. Hot
Samsung Electro-Mechanics	+146.0	+52.5	8,986.4	-9,036.0	+4,953.0	+3,922.2	MLCC + FC-BGA leader. Institution-led
Daeduck Electronics	+81.4	+20.5	1,198.1	+524.4	+827.0	-954.7	FC-BGA / MLB core name
Simmtech	+74.5	+32.3	822.6	+678.3	+1,376.0	-2,070.4	SOCAMM / substrate core
Haesung DS	+72.7	+19.2	297.3	+132.7	+656.4	-53.9	Heat spreader / substrate option
Hana Micron	+48.3	+5.0	1,200.6	+2,172.5	+158.3	+135.8	Foreign-led back-end recovery
HD Hyundai Heavy Industries	+51.0	+21.1	3,642.0	-5,344.5	+7,038.2	-1,987.9	Institution-led shipbuilding + nuclear option
Hanwha Ocean	+2.1	+16.3	2,369.4	+1,472.3	+1,393.3	+6,409.7	5D flow improvement, still a laggard

The key is that not all leadership has the same flow quality.

Samsung Electro-Mechanics rose 146.0% over one month, but foreigners sold KRW 903.6 billion. Institutions and retail absorbed the supply. That fits the SEMCO KRW 100T market-cap note: the AI passive-component re-rating is real, but chasing efficiency has fallen.

Hana Micron, by contrast, rose 48.3% with KRW 217.25 billion of one-month foreign net buying. That is why it looks more like a second-line back-end expansion candidate than an already over-owned first-line leader.

HD Hyundai Heavy Industries was sold by foreigners but bought heavily by institutions. That connects to the HD Hyundai Heavy Industries SMR option analysis: the shipbuilding / engine / SMR story is alive, but the rally is institution-led and price location matters.

5. Is KOSDAQ Weak, Or Selectively Coming Back?

KOSDAQ’s 20D ADR is 66.7. On the surface, that is weak. But it does not mean the whole KOSDAQ should be avoided.

The earlier KOSDAQ smart-money and Pearl Abyss rebound note argued that flows can turn before prices. The ADR data narrows that frame.

What matters is not buying KOSDAQ broadly. It is finding second-line names where turnover is just starting to accelerate, foreign/institutional flows are positive, and the 20-day moving-average extension is still manageable.

The local screen highlights:

Rank	Stock	Theme	5D	20D	5D Avg Turnover	Turnover Accel.	20D MA Gap	fi5	Read
1	HPSP	Semi equipment / AI infra	+12.7	+3.4	1,989.2	1.28x	+2.9	+509.3	Cleanest second-line candidate
2	SFA Semicon	Back-end	+19.9	-2.5	3,020.5	3.41x	+7.4	+259.1	Back-end expansion candidate
3	Hana Micron	Back-end	-1.2	+16.6	1,386.3	1.18x	+3.8	+129.4	Pullback candidate
4	Dongjin Semichem	Materials	+5.4	-4.0	642.5	1.15x	+3.1	+447.4	Materials-flow recovery candidate
5	KMW	RF / AI-RAN	+11.7	+20.8	259.0	1.30x	n/a	+226.8	AI-RAN event-confirmation candidate

fi5 is five-day foreign plus institutional net buying. Some local flow fields can be missing or incomplete, so Kiwoom / KRX flow validation is required before stock-level execution.

6. Breadth Expansion Triggers

This is not yet a broad-market buy zone. At an all-market ADR of 67.3, decliners still dominate.

The confirmation checklist is:

Trigger	Threshold	Meaning
20D ADR recovers to 80	All-market ADR above 80	Decliner dominance is easing
20D ADR recovers to 100	All-market ADR above 100	Advancers and decliners are balanced
Daily advance ratio above 55%	2-3 consecutive sessions	The rebound is not a one-day bounce
KOSDAQ turnover rises	Turnover up with more advancers	Small/mid-cap breadth can broaden
Foreign selling is absorbed	FX stable and index holds/rises	More absorption than true risk-off

By sector, the sequence to watch is:

Sector	Signal	Meaning
AI infra second-line	HPSP, SFA Semicon, Hana Micron, Dongjin Semichem turnover rising	Internal expansion within semis
Optical / RF / AI-RAN	KMW, RFHIC, Oi Solution flow turning	Marvell / NVIDIA AI-RAN linkage
FC-BGA / MLB	Daeduck, ISU Petasys, Korea Circuit re-accelerating	Custom ASIC / AI networking confirmation
Test sockets / back-end	ISC, Leeno, TSE, Doosan Tesna turnover rising	SOCAMM / ASIC test-infra expansion
Shipbuilding / defense second-line	Flows rotate into laggards during leader pullbacks	Rotation within an existing leadership theme

This links back to the Marvell / Broadcom Korea AI bottleneck preview. If the market is rotating from a single HBM trade into custom ASICs, AI networking, optical links and power integrity, the broad ADR can stay weak while those lower-stack bottlenecks keep attracting turnover.

7. Practical Read

The current Korean tape can be summarized in two lines:

The broad market is weak.
Leadership is not dead.

The action plan must reflect both.

Action	Condition	Targets
Hold existing leaders	Relative strength remains intact despite weak ADR	Samsung Electro-Mechanics, Daeduck and other AI infra names
Avoid chasing first-line leaders	20D ADR below 80	Hot first-line names
Watch second-line candidates	Turnover acceleration + positive fi5 + manageable MA gap	HPSP, SFA Semicon, Hana Micron
Replace weak positions	Holdings underperform the market and flows are weak	Non-leadership / non-core positions
Manage cash	Portfolio concentration is high before breadth recovers	Do not fully deploy cash while breadth is still weak

Chasing first-line leaders is inefficient here. Samsung Electro-Mechanics, Jeju Semiconductor and Simmtech have already moved sharply. But avoiding all equities because breadth is weak would miss the narrow leadership regime.

The cleaner stance is: hold leaders, watch second-line names, and require ADR above 80 before broadening exposure.

8. Conclusion

Korea is not a dead market. But it is not a broad market either.

As of the May 26, 2026 close, the all-market 20-day ADR is 67.3. KOSPI is 68.5 and KOSDAQ is 66.7. That is a poor backdrop for buying the average stock. At the same time, Samsung Electro-Mechanics, Jeju Semiconductor, Daeduck Electronics, Simmtech, Hana Micron and HD Hyundai Heavy Industries still show strong leadership.

So the regime is narrow leadership, not broad risk-on.

The three checks are:

Does all-market ADR recover above 80?
Do KOSDAQ turnover and the number of advancers rise together?
Do foreign and institutional flows begin to appear in AI-infra second-line names?

Until those conditions improve, chasing first-line leaders is less attractive than observing second-line candidates. If breadth recovers, the market can broaden. If it does not, only a few leaders will survive. The right stance is neither optimism nor pessimism. It is to accept that market breadth is narrow and follow where real turnover and quality flow are appearing.

Appendix. Evidence Classification

[Fact]

As of May 26, 2026, the all-market 20D ADR was 67.3.
On April 16, 2026, the all-market 20D ADR was 113.1.
On May 26, 2026, 749 Korean stocks advanced, 1,660 declined, and daily ADR was 45.1.
The last-month leadership list includes Samsung Electro-Mechanics, Jeju Semiconductor, Daeduck Electronics, Simmtech, Haesung DS, Hana Micron and HD Hyundai Heavy Industries.
HPSP, SFA Semicon, Hana Micron and Dongjin Semichem screened relatively well on turnover and flow conditions.

[Inference]

The current market is narrow leadership, not broad risk-on.
More important than simple foreign buying is whether a sector can hold price and attract turnover despite foreign selling.
Within AI infrastructure, rotation can move from SOCAMM / LPDDR into FC-BGA / MLB, back-end, optical and RF.
Second-line turnover and flow turns have better expected value than chasing first-line leaders after large moves.

[Speculation]

Marvell earnings may strengthen custom ASIC, optical and AI-RAN themes more than SOCAMM alone.
Broadcom earnings may re-ignite AI networking, FC-BGA / ABF and high-speed MLB names.
If ADR recovers above 80, the probability of second-line expansion rises.

[Blocked]

May 27, 2026 closing ADR is not included in this note.
Some five-day flow fields may be missing or incomplete in the local DB.
AI-RAN candidates such as RFHIC and Oi Solution require separate Kiwoom / KRX flow validation before firm stock-level judgment.

Data source: Research OS local DB prices_daily, universe.csv, korea_adr_recent_20260526.csv, korea_leaders_20260415_20260526.csv, second_line_theme_flow_candidates_20260527.csv. Data cut: May 26, 2026 close. This is research commentary, not investment advice.

Disclaimer: For research and information purposes only. Not investment advice. Names cited are for analytical illustration; readers should perform their own due diligence and consult licensed advisors before any investment decision.