Record Earnings.
So Why Has the
KOSPI Nearly Halved?

Earnings forecasts rose 33%,
while the index fell 45%.

An index level can be expressed as earnings multiplied by the PER P/E multiple investors are willing to pay. If earnings rise while prices fall, the multiple must contract. Since June 22, multiple compression accounts for the entire decline in this arithmetic breakdown.²

Multiple compression describes the arithmetic of the decline; it does not explain why investors lowered the multiple. They may doubt the durability of earnings, demand a higher risk premium, react to rates or positioning, or respond to market flows. Sections 02–05 examine those possible drivers.

What Matters

• A forward P/E of 3.54x is the lowest since 2001. PBR P/B (price divided by book value) is also below 1, at 0.95x.²
• Earnings estimates rose even during the selloff. Forward 12-month EPS increased 9.0%, from 1,310 on October 28 to 1,429 on November 5. About 3 percentage points of that gain came from the monthly roll of the forward window toward the more profitable year 2027. The 32.9% gain since June also includes this mechanical effect.²
• The discount is deeper than three weeks ago. NO.12 examined the earnings cuts implied by a KOSPI level of 5,554 on October 15. By November 6, the index had fallen another 8.9%, even as earnings forecasts climbed further.

Yuanta Securities applied the median P/E since 2004 to today's index level and backed out implied earnings of roughly ₩400tn—63.4% below the ₩1,103.7tn analyst consensus. Since 2007, this gap has generally remained within 20%, widening to around 30% even in extreme periods such as the global financial crisis.⁵

This is a hypothetical calculation that holds P/E at its historical level, not a direct measure of what investors actually expect earnings to be. Yuanta interprets the gap as a sign that the market distrusts the stepwise profit growth assumed in consensus forecasts. The calculation alone cannot isolate how much reflects rates or required risk premia.⁵

What, then, is the market questioning? Broker research points to two fundamental doubts, both sharpened by U.S. regulation.

The first doubt:
How long will these earnings last?

The market doubts the longevity of profits more than their current level. Semiconductor PER P/E multiples of 2–3x have historically appeared near cyclical earnings peaks.⁶

Why Investors Are Skeptical

• Memories of cyclical peaks. Domestic semiconductor forward P/E was similarly low in Q3 2010, Q2 2014 and Q3 2018—periods when semiconductor operating earnings were at near-term peaks.⁶
• Concentration risk. A shift in chip-sector forecasts moves the entire index's earnings outlook. Excluding Samsung Electronics and SK hynix, Q3 earnings missed expectations by 2.7%; 2026 operating-profit forecasts were cut in 19 of 26 sectors.² Semiconductors also represent 52.8% of KOSPI market capitalization.⁶
• Upward revisions are slowing. Weekly earnings-estimate upgrades decelerated from +10.07% to +1.70% for SK hynix and from +6.37% to +2.62% for Samsung Electronics. Upgrades have shifted toward AI compute and networking names such as AMD and Arista.⁷
• Stocks have led earnings down before. After Samsung Electronics shares peaked in November 2017, January 2021 and July 2024, operating profit still rose another 9.8%, 43.5% and 33.2%, respectively. Hanwha Investment & Securities argues that equities price changes in expected future growth, rather than simply reflecting current profits.⁸

Hana Securities notes that stocks with rising earnings estimates have underperformed for seven consecutive weeks. Its interpretation: investors are anticipating when those estimates will eventually turn down, rather than rewarding today's revisions.⁷

Evidence Against a Downturn

• Customers have not canceled orders. SK Securities says past memory downturns typically followed this pattern: share-price peak → order cuts about three months later → memory-price declines another three months later. No order cuts have been identified since the June peak.⁹
• Contract prices are still rising. Q4 memory contract prices are projected to rise 12.5% for DRAM and 17.5% for NAND.²
• Supply agreements extend further out. Long-term agreements (LTAs)now extend beyond 2030. SK Securities expects prices for 2027 high-bandwidth memory (HBM) to rise as well.⁹

Historically, when the Philadelphia Semiconductor Index fell more than 30% from a peak, the S&P 500 semiconductor forward EPS estimate was cut an average 16%. When Samsung Electronics shares lost over 40%, its forward net-income estimate was cut an average 34%.¹⁰ This time, estimates have not been cut yet. For more on the pace of earnings, NO.11 see

The second doubt:
Can AI financing hold up?

Much of Korean semiconductor profitability depends on investment in U.S. AI data centers. If builders cannot secure funding, GPU orders may be delayed—and demand for the HBM memory installed alongside those GPUs could weaken. For now, the stress is concentrated not in U.S. Big Tech generally but in newer cloud operators building data centers with borrowed money (neoclouds), notably CoreWeave.²

What the Numbers Show

• CoreWeave's five-year CDS (default protection premium)rose from 475 bp in late May to a record 1,268bp on Nov 5, then eased to roughly 1,155bp on Nov 6. A spread of 1,000bp implies a cumulative five-year default probability of around 60%, depending on the assumed recovery rate (Daishin's conversion: roughly 63–64%).²
• CDS spreads widened among AI suppliers. Over 30 days, Oracle rose 59bp, Broadcom 62bp and Nvidia 49bp, versus just 1bp for Microsoft.²
• U.S. technology-sector investment-grade credit spreads rose from 117bp in early January to a year-to-date high of 159bp. Since late September, investment-grade spreads have widened faster than high-yield spreads.¹⁰
• Operating cash cannot keep pace with capex. This year, hyperscaler capital spending is expected to increase 82%, versus only 25% growth in operating cash flow. The respective growth rates begin to converge in 2027, at 39% and 33%.¹⁰
• CoreWeave's repayment schedule. Scheduled principal repayments for 2027 total roughly $6.2bn across several borrowings, based on Q2 filings. This is not a single bond maturity: the issue is whether cash receipts under long-term contracts align with debt repayment dates.¹¹

But why did bond yields fall for Big Tech companies with negative free cash flow? Negative FCF does not automatically imply an inability to service debt. Big Tech has cash reserves, operating cash from established businesses, access to bond markets, and discretion over the pace of investment. Neocloud operators finance more of their infrastructure with debt and rely heavily on external funding. That distinction underpins Daishin Securities' view that hyperscalers' financing conditions have actually improved.²

Why Was Korea Hit Harder?

Since November 3, hyperscaler shares have edged higher. Semiconductor suppliers and the major chip-supply economies—South Korea, Japan and Taiwan—have suffered much steeper declines.¹² The 2027 capex consensus for five major tech companies rose another 1.2% after the election.² iM Securities argues that the market is pricing a slowdown in the growth rate of AI investment, not a halt to investment itself.¹²

The Next Test: Anthropic's IPO

• A Nasdaq listing before Thanksgiving (November 26)is reportedly the target, with a possible offering size around $100bn.^{12, 13}
• Valuation expectations have been coming down. Daishin Securities estimates a revised range of $1.0tn–$1.3tn. Kyobo Securities reports that some investors are discussing valuations near $1.3tn, down from earlier expectations as high as $2tn.^{2, 14}
• As of November 5, the IPO price range had not been disclosed. It was still absent from the amended registration statement filed November 2.¹³

Kyobo Securities sees Anthropic's eventual offering price and proceeds as a test of investors' willingness to fund—and value—the AI industry.¹⁴ Hana Securities presents a counterpoint: hyperscalers' current return on invested capital (ROIC) of 18% differs from the late-1990s downturn in telecom and equipment makers, suggesting technology credit spreads could begin narrowing rather than widening further.¹⁰

Regulation amplified both doubts;
the exposure is concentrated in 2028

The November 3 midterm elections brought new governors to several states with sizable data-center pipelines. Some governors-elect pledged moratoriums on new permits or conditional pauses. Daishin Securities views the prospect of restrictions as a trigger for the selloff, sharpening concerns about both earnings durability and AI financing. But state restrictions primarily affect new approvals: most capacity due online next year is already under construction.^{2, 15}

• Other brokers cite similar figures. Hanwha Investment & Securities also estimates that around 75% of capacity scheduled for completion in 2027 has already broken ground.¹⁶
• 57% of unstarted projects are in high-risk states. Kyobo Securities estimates that, among hyperscalers'approximately 23GW of unstarted projects, 13GW is in states facing elevated regulatory risk: Texas 5.1GW, Georgia 3.1GW and Pennsylvania 2.6GW.¹⁴
• Construction is a relatively small slice of capex. Data-center construction accounts for roughly 11% of hyperscaler capital spending. iM Securities argues that construction delays may therefore have a limited impact on total capex.¹²

The two figures labeled “23GW” mean different things: Daishin's estimate is the portion of 2028 net additions exposed to possible delays, while Kyobo's figure is the entire hyperscaler pipeline that has not broken ground. The 33.8GW for 2027 (20 major states) and 62.2GW for 2028 (nationwide) also cover different universes and cannot be used to calculate a growth rate. For the state-by-state regulatory map, see NO.27; for data centers and ESS demand, see NO.28.

The actual operational impact of data-center delays may be a 2028 issue. Stocks, however, discount that risk today—and foreign investors have accounted for much of the selling.

Foreign investors were the sellers;
their flows moved alongside U.S. credit

Foreign investors have been net sellers for seven consecutive months. The direction of the correlation is striking: this year, foreigners tended to sell as Korean earnings forecasts rose (correlation coefficient −0.32). Earnings forecasts alone do not explain their trading.^{2, 17}

Since 2019, U.S. high-yield spreads have had the strongest relationship with foreign investor flows (−0.46). In 2026, the U.S. Dollar Index (−0.65) and USD/KRW (−0.60) ranked first. The KOSPI itself showed only +0.28.¹⁷ These are correlations in contemporaneous changes; they do not tell us which variable moved first. They do, however, support watching credit and FX alongside the KOSPI for evidence that foreign investors might return.

How Heavy Was the Selling—and Why So Fast?

• Much of the overweight has already unwound. Daishin calculates that Korea's weight in a hypothetical global portfolio rose from 6% at the beginning of the year to 10.7% on June 19 as Korean stocks surged, before falling back to 6.6% by November 6.²
• Foreign ownership of semiconductors: 46.6%. As of November 5, the lowest level since 2013.²
• Rules-based selling amplified the fall. Volatility-targeting funds (up to $135bn, Barclays estimate) and leveraged ETFs (about $47bn, Nomura estimate) sold mechanically. The 44.5% decline from peak to trough took 93 trading sessions, faster than in 1997 (109) or 2008 (240).²

That does not necessarily mean foreign buyers will return quickly. In past episodes when the index fell more than 25% and heavy foreign selling persisted beyond 100 sessions, foreign net selling continued over the next one to three months in 70–74% of observations. Hana Securities looks for a return no earlier than three months out; the statistic is based on qualifying trading-day observations.¹⁷

Where brokers see downside support:
Broadly 4,800–5,300

Does that make the market cheap? The November 9 close of 5,018.85 was below Daishin's undershooting threshold of 5,330 (12-month forward PBR P/B of 1.0x), and just above Hanwha's lower bound of 5,000. All price levels and recommendations below are the respective brokers' views.^{2, 3, 16}

Most brokers agree that stocks look cheap, but Kyobo and Eugene caution against chasing a rebound or crowding into the same positions. The historical statistics on rebounds after drawdowns exceeding 30%, including sample conditions, are in the expandable “Historical Statistics” section.^{2, 6, 14}

To see whether confidence is returning,
watch more than the KOSPI

What needs to hold for the valuation case to work? Chip orders, AI funding, rates, the dollar and foreign flows all matter. The numbers below are publicly observable. Statuses are as of the evening of November 9.

The earnings-related signals are broadly holding. Most unmet criteria concern the other two areas: U.S. credit and the conditions for foreign investors to return. This week's first tests are CPI on November 10 and CoreWeave on November 11. As the results arrive, compare them with the thresholds in the table.²

A completed scorecard does not guarantee the KOSPI will rise. This is not a market-bottom timing model; it tracks whether evidence is accumulating for a lower required risk premium. The U.S. high-yield spread of 3.48% is this year's highest but remains below the 4.61% reached during the April 2025 reciprocal-tariff shock. Stress is concentrated more in AI infrastructure than across U.S. credit as a whole.^{2, 19}

Most year-over-year figures for U.S. October CPI will not be available because the October 2025 survey was missed during the federal government shutdown.^{2, 20} That makes the month-over-month change in core inflation the focus this time. U.S. bond markets close November 11 for Veterans Day, so CDS and bond-market reactions to CoreWeave's earnings will not be observable until November 12.⁷ For the Federal Reserve's December meeting, see NO.20 .

This Month's Catalysts:
The First Tests Come Tuesday and Wednesday

Times in Korea Standard Time. CoreWeave/Nvidia timings reflect post-results conference calls (Daishin and Hana event calendars). MSCI announcement: 11 p.m. Central European Time on November 11.^{2, 7, 21, 22}

Risks That Could Delay or Reverse the Thesis

The first risk would erode earnings themselves. The second and third would raise investors' required discount even if earnings forecasts held steady. If all three are avoided, the gap between earnings and prices could narrow as confidence returns.

Next Checkpoints

• 11.10 22:30 U.S. core CPI ≤0.2% m/m; then whether the 10-year Treasury yield holds below 4.8%
• 11.11 07:00 CoreWeave's Q3 funding volume, interest rates and collateral terms (Daishin considers rates below the current ~16% reassuring, though bonds with different collateral and maturities are not directly comparable); whether its five-year CDS 1,000 bp moves below the threshold can be assessed from November 12, when U.S. bond markets reopen
• 11.11 Nov 1–10 semiconductor exports; whether foreign investors turn net buyers of Korean chip shares that day
• Daily FRED U.S. high-yield spread (3.48%) and net foreign buying of cash equities and futures
• Late November Anthropic's IPO price range and actual proceeds

What Is Known—and What Is Not