Research Process 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.

Korea Invest Insights Thesis Check: What 207 Korean Articles Say About Our Edge

Tue, 26 May 2026 12:30:00 +0900

This is a meta follow-up to the Korea Daily Market Hub and Why Korea Part 5. The goal is not to claim portfolio performance. It is to ask which investment theses published on Korea Invest Insights lined up with subsequent price action, and where the process needs correction.

TL;DR

We collected all 207 Korean-language /ko/post/ articles from the Korea Invest Insights sitemap and mapped them to 593 article-stock pairs.
Using primary mentions only, 527 pairs had calculable performance. The win rate was 54.1%, +10% outcomes were 29.4%, +20% outcomes were 21.1%, and -10% outcomes were 24.3%.
The strongest cluster was AI infrastructure, semiconductors, HBM and substrates. Across 305 primary pairs, the average return was +15.5%, the median was +7.8%, and the win rate was 67.9%.
A separate thesis map compresses the corpus into 14 investment theses plus 53 operating logs. Roughly 39% of non-operating posts sat inside a broad Korea growth / AI-semiconductor risk-on cluster. That was a source of performance, but also a correlation risk.
The weakest areas were gaming, parts of biotech and medtech, K-beauty / consumer, listed VC proxies and some idiosyncratic deep dives. These cases often had good narratives but weaker price bridges.
The core lesson: KII’s edge was strongest when theme diffusion + flows + bottleneck logic aligned. It was weaker when product narratives, binary events or low-liquidity themes were stretched into structural equity theses too early.

This is not investment advice and not an actual portfolio return. It does not include position sizing, execution price, stops, taxes, transaction costs, publication-time effects or realized trading decisions.

What Was Measured

The purpose was to test research signals, not writing quality. The question was: which types of thesis were followed by positive price action?

Item	Basis
Article source	Korean sitemap, Korean index.json, LLM guide
Articles	207
Article-stock pairs	593
Successful performance calculations	557
Primary-mention successful calculations	527
Primary-mention ticker count	166
Price source	Research OS local DB, KR prices_daily + Kiwoom open fallback, US_Crawler prices_daily
Latest price date	2026-05-22
Entry price	Publication-date open, or next trading-day open if the publication date was a market holiday
Latest price	2026-05-22 close

Primary mentions were based on tag_code, tag_name and tag_alias. Title or description aliases were excluded from the core interpretation because they create more false positives.

Overall Scorecard

Metric	Result
Positive-return share	54.1%
+10% or better	29.4%
+20% or better	21.1%
-10% or worse	24.3%
Average return	+7.9%
Median return	+0.6%
25th percentile	-9.3%
75th percentile	+13.9%

The result is constructive but not clean. The mean is positive, but the median is only slightly above zero. That means a relatively small number of large winners pulled up the average. The research process produced right-tail outcomes, but it also carried meaningful left-tail risk.

Thesis Map: The Bigger Risk Is Correlation

The performance workbook answers one question: what went up after publication? The thesis map answers a different question: what were we repeatedly exposed to?

On that lens, 206 posts compress into 14 investment theses and 53 operating logs. The count differs by one from the 207-article sitemap performance run because the two exercises used slightly different grouping rules.

Item	Value	Read
Thesis-map posts	206	Comparable to, but not identical with, the 207-article performance dataset
Investment theses	14	Large idea buckets: stocks, sectors, macro, policy and process
Operating logs	53	Daily Wraps, weeklies and screeners; roughly 26% of the map
Korea growth / AI-semi risk-on cluster	59	Pearl Abyss 25, AI infra 18, Samsung Electro-Mechanics 8, Samsung Electronics 8
Share of non-operating posts	Roughly 39%	59 divided by roughly 150 non-operating posts

Pearl Abyss is not a semiconductor company. The reason it appears in this correlation bucket is that the stock behaved more like a Korea growth / domestic risk-on exposure than a defensive standalone asset. In other words, the map is not a sector taxonomy. It is an exposure map.

The mid-review by thesis looks like this:

Thesis	Posts	Residual Alpha	Mid-Review
Pearl Abyss platform / franchise re-rating	25	Low	Deep tracking, but -28.1% price outcome. The market still wants a 2027 earnings bridge and next-title visibility.
Samsung Electro-Mechanics AI package components	8	Low	Thesis strongly validated by price. The remaining question is execution toward a 2028 OP path, not first-time reclassification.
AI semiconductor infrastructure beyond HBM	18	Medium	Best-performing cluster, but still exposed to an AI CapEx peak-out.
Samsung Electronics memory-cycle to AI-platform reclassification	8	Medium	Re-rating logic remains plausible, but foundry and memory-cycle risks must stay inside the thesis.
Why Korea / Korea discount compression	5	Medium	Structural frame worked, but the actual market was heavily concentrated in Samsung and SK Hynix.
KOSDAQ policy / smart money	8	Medium	Policy direction looks right; execution delays and fund-deployment risk need tracking.
Pamicell perception-change series	4	High	Clean recognition-gap thesis, but still needs direct evidence for the Doosan Electronics BG proxy logic.
Korean AI / sovereign AI / listed VC proxies	6	High	Potential information edge, but the bridge from private-market exposure to listed-stock re-rating is slow.
Humanoid robotics valuation skepticism	4	Medium	One of the few clusters with built-in valuation skepticism. That tone should be used more broadly.
Operating logs	53	System	Trust infrastructure rather than a single alpha claim; weak-side and short/risk screeners are still underbuilt.

Combining the thesis map with the performance table produces a more balanced conclusion. The good outcomes were not random. But many of the good outcomes were also not independent. A large share of the winners sat in the same AI-memory / AI-infrastructure / Korea growth factor complex.

What Worked

First-mention winners were concentrated in AI infrastructure and semiconductor bottlenecks.

Stock	First Mention	Return From First-Mention Open	Read
Samsung Electro-Mechanics	2026-04-09	+161.7%	AI MLCC, FC-BGA, silicon capacitors, power-integrity re-rating
Jeju Semiconductor	2026-05-13	+121.0%	Memory / SoCAMM / earnings-surprise diffusion
Opticore	2026-05-09	+100.0%	High-beta optical / CPO theme
SK Square	2026-04-14	+96.8%	SK Hynix NAV and holding-company discount compression
TES	2026-04-09	+89.7%	Semiconductor equipment, PEAD, smart money
SK Hynix	2026-04-14	+77.8%	HBM leader and AI-memory earnings leverage
Daeduck Electronics	2026-04-20	+61.7%	AI substrates and FC-BGA diffusion
Marvell	2026-04-10	+58.8%	Custom AI silicon and interconnect bottleneck

The common pattern was:

A large theme existed: AI infrastructure, HBM, substrates, power integrity, networking, optics.
Flows confirmed the story: foreigners, institutions or smart-money screens showed demand.
The logic was bottleneck-based: these were not optional products, but components needed for the system to work.
The theme moved down the value chain: from Samsung Electronics and SK Hynix into Samsung Electro-Mechanics, Daeduck, Simtech, TES and Jeju Semiconductor.

The Samsung Electro-Mechanics case is the clearest example. The thesis began with AI substrates and MLCCs, then expanded into silicon capacitors, Murata comparisons and AI-server passive-component bottlenecks. That sequence worked. But it also means the easy part of the thesis may already be priced.

What Did Not Work

The weaker cases were more idiosyncratic.

Stock	First Mention	Return From First-Mention Open	Read
Pearl Abyss	2026-04-04	-28.1%	Gap between game data and equity re-rating
GigaLane	2026-04-17	-27.5%	Low-liquidity neglected-stock volatility
Next Biomedical	2026-04-14	-26.3%	Medtech event risk and limited earnings bridge
OE Solutions	2026-05-09	-25.1%	CPO theme proximity and valuation risk
Openedges Technology	2026-04-25	-23.9%	Long-duration IP thesis versus short-term multiple compression
Hyundai Rotem	2026-05-01	-23.7%	Late entry into a strong defense / rail theme
DSC Investment	2026-04-29	-23.2%	Listed VC exposure not yet translating into equity demand

These cases had different stories, but similar failure modes. Product or industry narratives were often better than the near-term equity bridge. Some themes were directionally right but not directly monetized by the listed stock. Several names were small or high-multiple stocks with weak liquidity, making them more vulnerable when risk appetite softened.

Pearl Abyss is the most important negative case. The game data did not collapse. The issue was that the market wanted a bridge from Crimson Desert success to repeatable earnings: DLC, the next title, capital allocation and the 2027 earnings path. Product data can support a hold thesis. It does not automatically create a stock re-rating.

Cluster Results

Keyword-based clusters sharpen the lesson.

Cluster	Pairs	Tickers	Average	Median	Win Rate	+10% or Better	-10% or Worse
AI infra / semis / HBM / substrates	305	92	+15.5%	+7.8%	67.9%	40.7%	18.7%
Smart money / PEAD / quality	138	59	+11.1%	+4.6%	60.1%	41.3%	23.9%
Power / nuclear / energy	67	33	+9.1%	+2.4%	56.7%	35.8%	31.3%
Robotics / physical AI	31	19	+7.9%	+2.1%	58.1%	29.0%	22.6%
Biotech / healthcare	34	29	+2.7%	-4.2%	35.3%	26.5%	35.3%
Gaming / Pearl Abyss	49	15	+2.0%	-13.9%	34.7%	20.4%	57.1%
Listed VC / venture proxies	18	13	-5.5%	-4.6%	22.2%	11.1%	38.9%
K-beauty / consumer	22	11	-8.7%	-8.5%	22.7%	4.5%	31.8%

The clearest edge was in the AI infrastructure value chain. The second-best signal was the screener process: PEAD, quality and smart-money intersection notes were more robust than pure narrative writeups.

Power and energy were positive but volatile. K-beauty, listed VC, gaming and parts of biotech were weaker because the bridge from story to earnings, flows or timing was less direct.

Process Lessons

1. Separate good thesis from good entry price

A correct thesis can become a poor entry after the market prices it. Samsung Electro-Mechanics is now a valuation question, not just a thesis question. Pearl Abyss is the reverse: product evidence may be strong, but the equity needs a clearer earnings bridge.

2. Make primary tags stricter

Performance attribution depends heavily on tag quality. Comparison names, risk examples and passing mentions should not become primary tags. Core thesis names should always have ticker tags.

3. Separate article types

A Daily Wrap, a screener note, a deep dive and an explainer are not the same product.

Article Type	Correct Performance Lens
Daily Wrap / Screener	Short-term flow and earnings-drift validation
Deep Dive	One-to-three-month thesis validation
Hub / Explainer	Discovery, SEO/GEO and reader navigation value

4. Update failures faster

Losers should not simply be called wrong. They need classification: broken thesis, delayed thesis, overpaid entry, weak liquidity, or missing catalyst. Pearl Abyss, Openedges, K-beauty and listed VC proxies deserve follow-up reviews under that framework.

Current Judgment

The interim read is straightforward:

KII has a research edge, but it is not evenly distributed across all article types or sectors.

The edge is strongest where AI infrastructure bottlenecks, flows and earnings drift overlap. It is weaker where the analysis relies on product quality, binary events, long-duration narratives or listed proxies for private-market exposure.

The thesis map adds one more point: the best-performing posts were not independent exposures. AI memory, AI infrastructure, Samsung Electro-Mechanics, Samsung Electronics, KOSDAQ growth and parts of Pearl Abyss were all tied to Korea risk-on conditions. The next review should therefore track not just returns, but factor concentration and thesis correlation.

The operating change is therefore:

Keep AI infrastructure and semiconductor bottlenecks as core coverage.
Keep Daily Wrap and screener-driven notes, but tighten tagging and tracking.
Apply more price discipline to gaming, biotech, consumer and listed VC stories.
Repeat this performance review monthly, not as a scorecard for bragging, but as a feedback loop for the research process.
Track correlation risk separately. Many posts can still be one large position if they share the same factor exposure.

Data Notes

Facts: 207 Korean articles, 593 article-stock pairs, 557 successful performance calculations, 527 successful primary-mention calculations, 166 primary-mention tickers. Latest price date: 2026-05-22.

Inference: KII’s strongest historical signals came from AI infrastructure, semiconductors, power and smart-money screening. The weaker areas were more idiosyncratic and less directly tied to earnings or flows.

Limitations: This is not actual portfolio performance. It excludes sizing, execution, costs, taxes, stops, publication-time effects and post-publication judgment changes. Posts published after the latest price date are blocked from performance calculation.

Sources

Korea Invest Insights Korean sitemap
Korea Invest Insights Korean index.json
Korea Invest Insights LLM guide
Research OS local DB: KR prices_daily + Kiwoom open fallback, US_Crawler prices_daily
Internal outputs: kii_article_performance_workbook.xlsx, ticker_summary_primary_mentions.csv, article_stock_performance.csv, summary_report.md, generated 2026-05-26 KST

For research and information purposes only. Not investment advice. Past performance does not guarantee future results.