Everything About AI
NextFin News - South Korea's benchmark KOSPI index closed the third quarter of 2026 down 18.9% from the end of June, turning the world's best-performing major equity market of the first half into its worst performer of the third quarter, as the artificial-intelligence chip rally that propelled it to record highs reversed with almost equal violence.
The index fell from 8,476.48 at the end of June to 6,870.81 at the September 30 close, and sits 24.6% below its record closing high of 9,114.55 set on June 22. The reversal came with record-breaking volatility: on July 28 the KOSPI plunged 10.84% to 6,023.66, one of its worst single sessions on record, before rebounding 17.9% three days later - the largest one-day gain in the index's history. The whipsaw is the market's verdict on a simple structural fact: Samsung Electronics and SK Hynix together account for more than half of the KOSPI's weighting, so a wobble in the global AI buildout becomes a national market event.
The Setup: A Market Built on Two Chips
The story of the Korean market in 2026 is a story of concentration. The KOSPI more than doubled in the first half of the year, crossing 9,000 for the first time on June 18, driven almost entirely by the high-bandwidth memory boom that feeds AI data centers. SK Hynix, which held an estimated 57% revenue share of the HBM market in the third quarter of 2025 compared with Samsung's 22%, according to Counterpoint Research, surged more than 250% year-to-date at its peak. Samsung, racing to close the HBM gap, gained more than 160%.
That concentration is not an accident of weighting rules alone. Foreign investors, who own roughly 32% of the KOSPI's market capitalization, bought Korea as a pure, liquid proxy for the AI memory cycle. Domestic retail investors piled in through newly approved single-stock leveraged exchange-traded funds tied to the two chipmakers. The result was a market structure in which the AI trade was not one sector among many - it was the index.
The reversal began in early July as doubts surfaced over whether hyperscaler capital expenditure could sustain the pace implied by chip valuations. The KOSPI fell 4.9% on July 7 and another 5.35% on July 8, closing at 7,246.79 and confirming bear-market territory - a drop of more than 20% from the June 22 record. Foreign investors sold more than 13 trillion won across nine consecutive sessions. By July 28 the panic peaked: the index fell 10.84%, triggering a market-wide circuit breaker - the eighth such event on the Korean exchange that year - as SK Hynix and Samsung each dropped more than 10%.
July closed with Samsung down 21% for the month and SK Hynix down 35%, according to contemporaneous market reporting. The rout erased nearly $2 trillion in market value at its trough. Yet the rebound on July 31 - a record 17.9% single-day surge to 6,595.45, sparked by Amazon's confirmation that AI infrastructure demand remains undersupplied through 2027 - showed the other side of the same coin. A market this concentrated does not drift; it gaps.
Why the Fall Was So Fast: Three Transmission Channels
The first channel is mechanical. With Samsung and SK Hynix comprising more than 50% of the KOSPI's weighting - data provided by Emmer Capital showed the pair accounted for more than half of the index's weighting as of June - a 5% move in the two names translates into roughly a 2.5% move in the index before any contagion. On July 28, when both fell more than 10%, the index math alone dictated a historic session. Three semiconductor stocks accounted for 87% of the KOSPI's entire market-capitalization gain in the rally, so when the gain reversed, there was no diversified corner of the market to absorb it.
The second channel is the investor base. Foreign holders of about a third of the index are the marginal price-setters and the fastest to exit when the global AI narrative wobbles. On September 3, foreign and institutional investors net sold a combined 13.9 trillion won in a single session as the index sank 3.99% to 6,562.72. Earlier in the summer, foreign investors had sold roughly 28 trillion won over eight consecutive sessions before flipping to net buying on July 8 - the same day retail investors finally capitulated into net selling after buying the dip all the way down. The two groups were on opposite sides of the trade, and both were using leverage.
The third channel is the currency. Even when Korean shares held up in won terms, the won weakened against the dollar through much of the period, eroding dollar-denominated returns for offshore holders. That created a feedback loop: a weaker won made Korean assets look cheaper to some buyers, but for foreign holders it amplified losses and accelerated the exit. The KOSPI's year-to-date gain of 41.5% in dollar terms, as of late July, was far smaller than its local-currency advance.
"Korea is a bit of an outlier, if you look at the other stock markets' reaction," said Jason Hsu, chief executive officer at Rayliant Global Advisors, noting earlier in the year that the KOSPI's heavy concentration in a handful of technology stocks meant market moves tend to be magnified relative to more diversified indices.
The Fundamental Defense: Earnings Are Real
The bear case for Korean equities runs into an inconvenient fact: the earnings are not imaginary. SK Hynix reported record second-quarter results on July 29 - revenue of 79.3 trillion won, operating profit of 60.5 trillion won, an operating margin of 76%, and net profit of 93.9 trillion won, all all-time highs. Samsung posted a record quarterly operating profit in the same period. Industry estimates compiled ahead of the late-October earnings releases point to a combined third-quarter operating profit of nearly 190 trillion won for the two companies, with Samsung around 111.8 trillion won and SK Hynix around 78.1 trillion won.
Memory pricing remains tight. Micron, SK Hynix, and Samsung have reported being sold out of DRAM and high-bandwidth memory through 2027. SK Hynix's lead in HBM is not easily dislodged: it derives nearly all of its revenue from memory chips, while Samsung spreads across smartphones, displays, and foundry work. Within Samsung, the memory division generated operating profit of about 24.9 trillion won in 2025, but the group's AI push is reorienting capacity toward server customers at the expense of conventional memory output.
Management has responded to the share-price collapse with the one tool that cannot be faked: cash. SK Hynix announced a 40 trillion won share buyback and raised its shareholder-return target to more than 50% of cumulative free cash flow generated between 2025 and 2027. Samsung followed with a package of 90 trillion to 110 trillion won in shareholder returns for 2026 - the largest ever by a Korean company - including roughly 30 trillion won in cash dividends in the third quarter. The buyback shield is real, but it has limits: corporate buying of Samsung and SK Hynix shares was being used by foreign and institutional investors as an opportunity to take profits in early September.
Janet Mui, head of market analysis at RBC Brewin Dolphin, framed the pullback as a healthy pause rather than a loss of faith in the AI trade: "Given that the South Korean chipmakers have basically rallied multiple times over the past years, this brief pause or break is actually quite healthy."
Cyclical or Structural: The Call That Determines the Conclusion
This is the question the market is really arguing about, and the answer has to be split in two. The memory cycle itself is cyclical - it always has been. DRAM and NAND are commodity products with boom-bust pricing that has mean-reverted through every cycle since the 1980s. The valuation de-rating we saw in the third quarter is a cyclical event: investors priced a supercycle, then repriced to a cycle, and the multiple compressed accordingly. History says this leg will revert when pricing tightens again.
But the market structure is structural, and it will not fix itself. A benchmark index in which two memory companies carry more than half the weight is a structural vulnerability, not a cyclical condition. The Bank of Korea reported that daily KOSPI volatility reached 4.1% from January to July - triple the average of 30 major economies and higher than during the 2008 financial crisis or the pandemic. The exchange triggered market-wide circuit breakers eight times through July, including on back-to-back sessions on July 28 and 29. That level of fragility is a feature of construction, and it persists regardless of where memory prices trade.
The same duality applies to the AI demand underneath. The demand is secular - data-center buildout is a multi-year phenomenon, and HBM shortages are projected to extend through 2027. But the pricing power that demand confers is cyclical, because memory is a commodity and every boom invites capacity expansion, including from Chinese competitors such as CXMT, whose strong market debut deepened competition worries in July. So the correct read is: a structural demand tailwind running through a cyclical pricing mechanism, amplified by a structurally fragile index.
There is also a seasonal pattern worth noting. According to Korea Exchange data reviewed over 27 years, the third quarter - July through September - has averaged a 0% return, the weakest of the four quarters, while the first, second, and fourth quarters averaged 2.4%, 5.2%, and 3.6% respectively. Narrowing to data since 2020, the third quarter was the only one with a negative average return, -0.6%. The 2026 third quarter was not an aberration in direction; it was an aberration in magnitude.
The Counter-Thesis: This Time the Fundamentals Hold
The strongest case against the bearish read is straightforward and well-supported. Earnings are at record levels, not inflated by accounting. SK Hynix's 76% operating margin in the second quarter is not a bubble multiple - it is a cash margin earned on chips that are genuinely scarce. The HBM shortage is expected to persist through 2027, which gives the two Korean champions a multi-year revenue runway that did not exist in previous memory cycles. Samsung and SK Hynix are returning 130 trillion to 150 trillion won to shareholders across 2026-2027, a floor under the share price that did not exist in 2008 or 2018. And despite the third-quarter collapse, the KOSPI is still up roughly 60% year-to-date as of mid-September - the best-performing major market of 2026 as a whole.
This argument is serious, and it is why a blanket "the AI bubble has burst" verdict would be wrong. The bulls are correct that the demand is real and the cash is real. But they are making a company-level argument about a market-level problem. SK Hynix can earn record profits and still see its share price fall 35% in a month, because the share price embeds not just this year's earnings but the entire future path of memory pricing, and that path is a commodity path. The buybacks provide support, not immunity: foreign and institutional investors used corporate buying as an exit door in early September. And the year-to-date gain, while impressive, is exactly what makes the market vulnerable - a 60% advance leaves little margin for disappointment.
What Comes Next: Three Horizons and the Signal That Would Break the View
In the short term - the next one to three months - the market is a function of sentiment and liquidity, and both remain fragile. The key events are the third-quarter earnings releases from Samsung and SK Hynix in late October, and the guidance they give for the fourth quarter. Retail investors, who bought the dip into the July collapse, have been net sellers on the way back up, which suggests the leveraged cohort has not fully unwound. The VKOSPI fear index spiked above 70 at the July trough; it can spike again on any negative surprise from U.S. hyperscalers. August net inflows into EWY, the flagship Korea ETF, fell 92.3% from July, and 1.53 trillion won was pulled from a DRAM-focused exchange-traded product in the second half of August alone.
In the medium term - six to twelve months - fundamentals take over, and here the picture is more constructive. If HBM pricing holds through 2027 as industry forecasts suggest, earnings will remain at elevated levels and the buyback programs will continue to absorb supply. The base case is a range-bound market, roughly 6,500 to 8,000 on the KOSPI, as record earnings offset a lower valuation multiple. The upside case requires memory pricing to re-accelerate on stronger-than-expected AI capital expenditure, which would push the index back toward its June highs. The downside case is a genuine AI capital-expenditure slowdown, which would compress earnings estimates and take the KOSPI below 6,000.
In the long term - beyond a year - the structural question dominates: can Korea diversify its market away from two memory names? The Value-Up governance reform program was supposed to narrow the "Korea discount" by improving shareholder returns across the chaebol, but it has not changed the index's composition. Until it does, Korea will remain the world's most liquid, most concentrated bet on the AI memory cycle - the highest-beta way to own the trade, up and down.
The falsifying signal for the view laid out here is specific and observable: if Samsung and SK Hynix report third-quarter HBM average selling prices that are flat or declining, or guide fourth-quarter memory revenue lower, the "secular demand supports valuations" argument breaks and the structural de-rating resumes. A secondary signal is foreign ownership of Samsung Electronics: it fell to 46.63% in early August from a high of 52.40% at the start of the year, and a break below roughly 45% would confirm that offshore holders are treating the Korean market as a trading vehicle rather than a long-term holding.
The final irony of South Korea's third quarter is that the market did the right thing for the wrong reason in the first half, and the wrong thing for the right reason in the second. The rally was justified by real earnings, but priced as if they would last forever. The selloff was justified by concentration risk, but executed as if the earnings had vanished. The truth sits between the two: this is a cyclical de-rating of a structurally concentrated market, and until the index stops being a two-stock market, every AI tremor anywhere in the world will register as a Korean earthquake.
Explore more exclusive insights at nextfin.ai.
