H1 2026 Review: AI Infrastructure Bottlenecks Broadened, but the Market Did Not

Related context This is a half-year follow-up to Samsung 2Q26 preview, Micron FY3Q26 earnings, the NVIDIA elasticity framework for Samsung and SK Hynix, how hard it was to beat the KOSPI benchmark, the KOSPI 60-day disparity framework, the Real Money Flow framework, and the AI supercycle midgame note.

TL;DR

H1 2026 looked like a broad risk-on period from the index level. Under the surface, it was much narrower. Korea was a KOSPI megacap crowding market. The U.S. was an AI infrastructure market where leadership moved from megacap AI platforms into memory, storage, servers, semiconductor equipment and power. Neither market was a broad “everything went up” market.

The headline Korean index was extraordinary. KOSPI rose from 4,214.17 on December 30, 2025 to 8,476.48 on June 30, 2026, up 101.1%. KOSDAQ fell from 925.47 to 916.18, down 1.0%. Across roughly 2,700 Korean stocks, 856 rose and 1,788 fell. The median stock return was -12.6%. The index doubled, but the average stock was in a weak market.

The U.S. story was also more specific than the index suggests. The S&P 500 gained 9.2% and Nasdaq gained 11.1%, but the real story was AI infrastructure bottlenecks. SanDisk, Micron, Western Digital, Dell, Marvell, ARM, Applied Materials and Lam Research were much stronger than NVIDIA. The market was buying the places where the AI factory actually bottlenecks.

The H2 strategy is not simply “find cheap stocks.” The better rule is to buy proven bottlenecks with real revenue and contracts on pullbacks. Option-like equipment, gaming and biotech event names should be sized smaller. Continuous-revenue bottlenecks in memory, eSSD, MLCC, substrates, power, shipbuilding, defense and large biotech platforms deserve more attention.

0. Data Basis and Reconciliation

The data basis is:

Some figures differ slightly by dataset and reference date. For example, the final KOSPI return is +101.1% using the June 30 close, while some local basket snapshots show +96.7%, and the June 23 interim review showed +90.4%. U.S. AI stock returns also differ depending on adjusted-close treatment and vendor handling.

This post reconciles those differences as follows.

The point is not false precision. The conclusion is stable: indices were strong, breadth was weak, and AI infrastructure returns moved from GPUs into memory, storage, servers, equipment and power.

1. Index Performance: KOSPI Doubled, KOSDAQ Fell

At the index level, Korea was spectacular. KOSPI doubled in six months. But KOSDAQ fell. Within the same Korean market, large-cap AI infrastructure and growth stocks lived in different worlds.

The U.S. index story also understates the internal rotation. The S&P 500 and Nasdaq were up around low double digits, while semiconductor ETFs were up far more. SOXX gained 104.2%, close to the KOSPI’s gain, showing how concentrated the global AI hardware trade became.

2. Korea: The Index Doubled, the Median Stock Fell

This is an unusual combination. The index was in a historic bull market, while the median stock was weak. That is why the lived investor experience was so different from the index headline.

The main H1 leaders were:

The Korean market’s structure was:

Samsung Electronics and SK Hynix
→ Samsung Electro-Mechanics, MLCC and FC-BGA
→ substrates, PCB and semi equipment
→ power equipment, industrials and some financials

KOSDAQ, many biotech names, platforms and parts of secondary batteries lagged. Korea was strong, but most Korean stocks were not.

3. Korea Was Not a Foreign-Buying Market

YTD flow data on the local Kiwoom amount basis:

The key point is that Korea was not a simple foreign-buying market. Foreigners sold Samsung Electronics and SK Hynix heavily. Domestic retail, financial investment accounts and institutions absorbed the flow.

Foreigners sold the core megacaps while those stocks rallied. That is not a normal foreign-led rally. It was a domestic absorption market.

Foreigners did not sell everything. They rotated into select bottlenecks and global-fund-friendly assets such as Samsung Electro-Mechanics, Doosan Enerbility, Samsung SDI, FADU, Hanwha Ocean and Sanil Electric. The right description is not “foreigners bought Korea.” It is “foreigners sold the largest memory megacaps, domestic liquidity absorbed them, and foreigners reallocated into select second-layer bottlenecks.”

4. The Hidden Buyer Was Financial Investment, Not Long-Only Real Money

Institutional buying was not the same as long-only accumulation. The strongest institutional subcategory was financial investment, not investment trusts, private funds or pension funds.

This matters. Real money can sustain a trend. Financial-investment and program flow can be powerful, but it is more vulnerable to futures expiry, ETF caps, derivative positioning and leverage mechanics.

H1 Korea was both a fundamental rerating market and a derivatives/ETF market. Confusing the two leads to repeated mistakes.

5. KOSDAQ Lacked Buyers, Not Stories

KOSDAQ had plenty of stories: biotech, robotics, CPO, SOCAMM, semiconductor equipment, gaming and AI software. The index still lagged because the buyer base was weak.

The June 23 interim snapshot was already clear:

The final June 30 structure was the same: KOSPI +101.1%, KOSDAQ -1.0%.

KOSDAQ struggled because:

1. Retail capital was tied up absorbing Samsung Electronics, SK Hynix and large AI infrastructure names.
2. Institutional real money did not buy KOSDAQ broadly.
3. Foreign buying was narrow, focused on names such as FADU, Jeju Semiconductor and Simmtech.
4. Rates and FX pressure weighed on biotech and unprofitable growth multiples.

The condition for KOSDAQ recovery is:

KOSDAQ recovery =
retail capital returning
+ institutional real-money buying
+ Samsung Electronics and SK Hynix moving sideways rather than falling
+ KOSDAQ earnings revisions stabilizing

If Samsung and SK Hynix crash, money does not automatically rotate into KOSDAQ. Risk-off comes first. The best environment for KOSDAQ is not megacap collapse. It is megacap consolidation.

6. The U.S. Market Was Not Just NVIDIA

The real U.S. signal was not “AI went up.” It was:

The physical bottlenecks after the GPU outperformed the GPU leader.

Representative H1 return ranges:

Big Tech was mixed:

The market bought the points where the AI factory bottlenecks:

GPU
→ HBM and DRAM
→ eSSD and storage
→ servers and racks
→ optical and interconnect
→ semiconductor equipment
→ power and cooling

This is directly relevant to Korea. HBM maps to Samsung and SK Hynix. eSSD and NAND map to Samsung, FADU and Jeju Semiconductor. MLCC and FC-BGA map to Samsung Electro-Mechanics and substrate names. Power maps to Doosan Enerbility and electrical equipment. Equipment maps to HPSP, TES, VM and Jusung.

7. Macro Was Not a Pure Tailwind

This was not a clean multiple-expansion macro environment. The dollar rose, the won weakened, U.S. yields rose and oil rose. Gold and Bitcoin fell.

Stocks rose anyway because the AI infrastructure earnings cycle overwhelmed macro pressure. That is important for H2. If rates, the dollar and oil all rise together again, the most crowded bottleneck stocks can suffer multiple compression even if demand remains strong.

The realistic risk is:

AI demand stays strong
+ earnings remain good
+ rates, dollar and oil remain firm
+ valuation and leveraged flows crack first

That is the same stock-elasticity problem discussed in the NVIDIA/Samsung/SK Hynix framework.

8. Non-Consensus Truth 1: Leadership Moved From Compute to Memory

The first non-consensus point is U.S. AI leadership. NVIDIA is still the center of AI. But incremental stock performance moved into Micron, SanDisk, Western Digital, Dell, Marvell and ARM.

That does not mean compute no longer matters. It means the marginal return moved down the stack.

Cost items get pressured. Bottlenecks rerate. Narratives require timing discipline.

9. Non-Consensus Truth 2: The Real Top Came Before the Index High

Breadth weakened before the index did. The June index high was driven by Samsung Electronics, SK Hynix and a small set of AI infrastructure names. KOSDAQ and the median stock had already weakened earlier.

Weak median stock
+ weak KOSDAQ
+ strong KOSPI memory megacaps
= index high and weak lived market at the same time

For H2, the key indicators are breadth, ADR, KOSDAQ/KOSPI relative strength, megacap memory trading-value share, and KOSDAQ retail/institutional joint buying.

10. Non-Consensus Truth 3: Multiples Moved Before Fundamentals Fully Arrived

Samsung, SK Hynix, Micron and Samsung Electro-Mechanics all had strong fundamentals. But stock returns moved faster than near-term earnings. The market changed the multiple.

Old frame:
memory = commodity cycle = peak-earnings discount

H1 frame:
memory = AI infrastructure bottleneck = supply shortage plus contract premium

The H2 risk is not just bad earnings. It is good earnings with lower multiples if the market starts treating memory as commodity again. Investors should watch EPS revision speed, beat magnitude and FY2027 visibility, not just the next quarterly beat.

11. Non-Consensus Truth 4: Leverage Plumbing Amplified Tail Risk

Korean memory megacaps were affected not only by fundamentals, but also by leverage plumbing: 2x and 3x Samsung/SK Hynix ETFs listed offshore, domestic leverage and inverse products, futures and program flow.

Simplified upside loop:

stock up
→ leveraged ETF rebalancing buy
→ futures and program support
→ more megacap upside
→ more momentum capital

Downside loop:

stock down
→ leveraged ETF rebalancing sell
→ futures and program pressure
→ retail margin and forced-selling risk
→ larger drawdown

Korea can therefore move more violently than Micron even when the fundamental direction is similar.

12. Non-Consensus Truth 5: Supply Moved From Strong Hands to Weaker Hands

The important question is not only who sold. It is who bought.

H1 saw foreigners and some institutional accounts reduce exposure in memory megacaps while retail and financial-investment accounts absorbed the flow.

H2 monitoring should combine foreign cash flow, retail margin, deposits, ETF balances, and program flow.

13. Non-Consensus Truth 6: Foreigners Bought Globally Explainable Korean Assets

Foreigners did not buy every Korean theme. The stocks they bought were globally explainable.

Common features:

power, nuclear, shipbuilding and defense
global quality industrials
large battery names
large biotech platforms
direct AI infrastructure bottlenecks

For H2, foreign long-only capital is more likely to buy assets it can explain in an investment committee than obscure small-cap stories.

14. Non-Consensus Truth 7: Revenue Pattern Beat Company Quality

In H1, the quality of the revenue pattern mattered more than abstract company quality.

Weak revenue-pattern names:

Strong revenue-pattern names:

The H2 rule is simple:

Size option-like names smaller. Size recurring-revenue bottlenecks larger.

15. Bonus Hypothesis: Korea Became a Leveraged U.S. AI Proxy

Korea’s market identity changed in H1. Historically, Samsung and SK Hynix were proxies for EM, China and global IT inventory cycles. In H1 2026, the linkage to U.S. AI capex became much stronger.

Session analysis suggested a stronger relationship between U.S. prior-day AI/semi moves and Korea next-day memory megacaps, including an interim correlation estimate of 0.42. That figure needs formal revalidation, but the price action is consistent with the direction.

The issue is diversification. If Korea behaves like a leveraged U.S. AI capex proxy, domestic buffers weaken when the global AI factor cools.

Another watch item is capex. Session analysis classified large SK Hynix capex commitment as a peak-confidence checkpoint. The exact amount and scope should be rechecked against filings before it becomes a core argument. But historically, peak capex announcements often arrive when demand looks strongest.

Key H2 capex questions:

16. H2 Checklist

17. H2 Strategy

The H2 strategy in one sentence:

Keep AI infrastructure core exposure, reduce option-like positions, and shift toward recurring-revenue bottlenecks and Korean quality cyclicals that global capital can explain.

18. Final View

The crowd may remember H1 2026 as the best Korea/semiconductor supercycle year. That is not wrong. KOSPI doubled and memory, substrates, equipment and power leaders exploded.

But the more important story is:

leadership became extremely narrow
multiples moved before fundamentals fully arrived
foreigners sold the core megacaps while domestic liquidity absorbed them
financial-investment and ETF/program flows amplified the market
KOSDAQ lacked buyers, not stories
U.S. AI leadership moved from GPU into memory, storage, servers and power

The H2 question is not only “how much more upside is left?” The better questions are:

1. Does stock elasticity remain positive even after strong earnings?
2. Can domestic liquidity keep absorbing foreign selling?
3. Do KOSDAQ buyers return?
4. Do AI bottlenecks convert from narrative into revenue and contracts?

H1 was risk-on, but it was not broad. It was a compressed leadership market. H2 requires more precision: real bottlenecks over cheap stocks, recurring revenue over one-off catalysts, and Core OP plus FCF over headline earnings.

Data Note

• Korea data: KR local DB, Kiwoom investor flow data and Naver Finance support as of the June 30, 2026 close.
• U.S. data: US local DB and yfinance adjusted close as of the June 29, 2026 close.
• Interim snapshots: June 23 KOSPI/KOSDAQ relative performance and session flow analysis.
• U.S. June 30 regular-session results are not included.
• Less-verified capex single figures are treated as H2 monitoring items rather than core proof.

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.