South Korea’s KOSPI rose 4.61 percent on Monday, 7 September 2026, closing at 6,995.39 — up 308.18 points, and exactly 4.61 points short of reclaiming 7,000. It had already opened 3.34 percent higher at 6,910.78. Samsung Electronics gained 5.68 percent to 270,000 won; SK hynix jumped 8.26 percent to 1,783,000 won. In Tokyo the Nikkei 225 added 1,378 points, or 2.12 percent, to 66,399, led by SoftBank Group at 11.78 percent and memory maker Kioxia at 5.4 percent. Wall Street was shut for Labor Day.
The trigger was a model. OpenAI released GPT-6 Astra on 3 September as a preview for selected partners and opened it to paying users the following day. Asian semiconductor desks read into it what they have read into every model announcement for eighteen months: more data centres, more accelerators, more high-bandwidth memory. That is the story every wrap-up will carry.
It is not the interesting number of the day. The interesting number sits in a footnote to a strategy note, and it is 5.3.
The number that matters is not 4.61 — it is 5.3
Timothy Moe, chief Asia-Pacific equity strategist at Goldman Sachs, reaffirmed his KOSPI target over the weekend: 12,000. He raised it from 9,000 in early June and has not touched it through one of the most brutal summers in Korean market history. What matters is not the target but its derivation. Goldman applies a forward price-to-earnings multiple of 7.5 to get to 12,000. And it justifies that by pointing out that the market currently trades at roughly 5.3 times.
5.3 times earnings. For the best-performing equity market in the world in 2026. For an index whose constituents, on Goldman’s own numbers, will grow net profit by 360 percent this year. The S&P 500 was priced at a little over twenty times forward earnings in late August. That is an earnings yield of 18.9 percent in Seoul against roughly 5 percent in New York — a ratio of almost four to one between two markets serving the same wave of demand.
A discount of that size stops being a discount. It becomes a statement. And the statement is not that the market doubts the profits. It is that the market gives them no future.
How an index can rise 66 percent and get 64 percent cheaper
The KOSPI closed 2025 at 4,214.17 — a 76 percent annual gain, its best year since 1999 and the highest year-end close in what was then a 43-year history. From there to Monday’s 6,995.39 is 66.0 percent. No market on earth has delivered more in 2026.
And the same index is nonetheless far cheaper than it was in January. That is not an opinion, it is division. If price is 1.66 times and earnings are 4.6 times — which is what a 360 percent increase means — then the multiple is 0.36 times. Working backwards from today’s 5.3, the year opened at roughly 14.7 times. The price rose 66 percent, the multiple fell 64 percent. Both at once, in the same index, over the same eight months.
That is the real event of East Asia’s 2026, and daily coverage misses it consistently, because daily coverage reports percentage moves in index levels and not percentage moves in valuation. The market has accepted the entire earnings explosion and refused to capitalise a single won of it. It is paying for about five years of these profits. After that, its price says, the thing is over.
This is exactly where Goldman plants its flag, and it uses a word that analysis rarely deploys so precisely: duration. It is not the level of earnings that is underestimated, but the length of the cycle. The argument in Seoul is not an argument about the profit number. Everyone agrees on the profit number. It is an argument about how many years it lasts.
July was not an earnings collapse — it was a terminal-value collapse
To understand why the multiple melted, look at the summer. On 22 June the KOSPI closed at 9,114.55, the highest level in its history. On 8 July it fell 5.35 percent to 7,246.79, more than 20 percent below that high and therefore into a bear market — three weeks after the record. July ended down 28.9 percent: the worst month this index has ever had. Worse than October 1997, at 27 percent, when Korea fell into the Asian crisis. Worse than the worst month of the 2008 financial crisis, at 23 percent. Roughly 250 trillion won of market value was gone. On 28 July alone the index lost 10.84 percent, SK hynix 14.65 percent and Samsung more than 13 percent.
The trigger was not an earnings report. It was reports that China had begun mass-producing its own deep-ultraviolet lithography machines — the tool class the Dutch firm ASML has dominated for decades and without which modern memory cannot be built. That is not news about the current quarter. It is news about terminal value. If China can make the tools, the question is no longer how much Samsung and SK hynix earn in 2026, but for how many more years they earn it.
And that is precisely how the market responded: it deleted the out-years, not the estimate. The cross-check is unambiguous, and it is the single strongest figure in this piece. Over the same stretch in which the index fell 23 percent below its high, the consensus for Samsung’s third-quarter operating profit was raised to 104.31 trillion won — from 55.52 trillion won three months earlier. An 88 percent increase in the estimate while the share price was collapsing.
Anyone who wants to see pure multiple compression does not need a textbook. They need those two numbers side by side.
What the physical market says: contracts, wafers, scarcity
Valuations are opinions. Contract prices are signed agreements. And so far the agreements say nothing about a cycle ending. Research house TrendForce expects server DRAM contract prices to rise 13 to 18 percent quarter on quarter in the third quarter of 2026, with individual negotiations running as high as 20 to 30 percent. What is capping the increase is not softening demand but long-term supply agreements that limit how fast a price can move.
Behind it sits a physical reallocation. High-bandwidth memory will consume around 23 percent of total DRAM wafer output in 2026, up from about 19 percent the year before. Because an HBM stack eats far more wafer area per usable device than conventional memory, every point of HBM share removes a multiple of ordinary bits from the market. The result is supply-demand gaps of 4.9 percent in DRAM, 4.2 percent in NAND and 5.1 percent in HBM — the widest since 2011. SK hynix has publicly floated that the shortage could run past 2030.
The earnings are correspondingly absurd. SK hynix reported an all-time-high operating profit of 64.1 trillion won for the second quarter of 2026, about 43.7 billion dollars. Samsung guided to 89.4 trillion won. Two companies, 150 trillion won or roughly 104 billion dollars of operating profit in a single quarter. And the market pays 5.3 times for it.
Two stocks are the index
Before anyone reads Monday’s gain as a broad Korean upswing, do the weighting arithmetic. Samsung Electronics most recently accounted for 26.04 percent and SK hynix for 20.90 percent of KOSPI market capitalisation — 46.94 percent combined. At the end of the prior year the pair was 22.63 percent. In late May they briefly crossed 50 percent. Two order books are half the market.
Run Monday’s moves through those weights and Samsung contributed about 1.48 percentage points, SK hynix about 1.73 — together roughly 3.2 of the 4.61 points, or about 70 percent of the day from two names. Reports from the session counted 441 decliners while the index was comfortably higher. This was not Korea rallying. It was memory rallying, measured against an index that is half memory.
Tokyo showed the same pattern in a milder form. The price-weighted Nikkei gained 2.12 percent; the broad Topix rose 22.57 points to 4,125.80, or 0.55 percent. The Nikkei rose almost four times as much as the market it supposedly represents. On days like this, reading index levels as an economic indicator means reading a weighting methodology, not an economy.
Who bought on Monday — and who sold
The exchange’s flow data are more revealing than the prices. Foreign investors bought a net 2.5525 trillion won. Domestic institutions bought a net 2.6398 trillion won. And Korean retail investors sold a net 6.8274 trillion won.
That is a reversal. Korean private investors — known at home as the ants — were the load-bearing force of the 2025 rally. They rode the recovery out of the July crash and are now using it to leave, at a level still 23 percent below the June high. An investor who watched their account at 9,114 in June and then lost 28.9 percent in four weeks does not sell a 4.6 percent bounce out of valuation discipline. They sell out of relief. The other side of that trade is institutional and foreign, and it is paying 5.3 times.
The third party sits in America and was closed on Monday. The Friday before, the S&P 500 lost 0.38 percent to 7,718.60, the Nasdaq Composite 0.29 percent to 26,506.99 and the Dow Jones 0.51 percent to 53,414.25, after non-farm payrolls came in at 162,000 against an expected 53,000 and pushed the odds of a 16 September rate hike back up. The declines were led by Apple at 2.55 percent, Alphabet at 2.10 percent and Microsoft at 2.05 percent. That is the buy side of memory. The bill for 13 to 18 percent more expensive server DRAM does not land on Samsung; it lands in hyperscaler capital budgets and depreciation schedules. On the same weekend Seoul celebrated the seller, the buyer lost value.
How US investors are actually exposed
The cleanest domestic read-across is Micron Technology, the only American memory pure-play of scale and effectively the third node of the DRAM oligopoly alongside Samsung and SK hynix. Micron gets the same contract prices, the same HBM allocation economics and — importantly — the same multiple problem: cyclicals earning records trade at trough multiples for a reason. SanDisk and Seagate sit on the NAND and storage side, where the 4.2 percent supply gap is narrower than DRAM’s but where AI-era data retention is a genuine second demand leg.
One rung upstream, Lam Research and Applied Materials are the most memory-levered names in wafer fab equipment; etch and deposition intensity is where HBM stacking spends money. Note what this implies: equipment orders follow capacity decisions, not spot prices, so this is a bet that the shortage persists long enough to be answered with new fabs. That is the same duration bet as the KOSPI multiple, taken from the other end of the chain.
The counterparty exposure is worth naming explicitly, because most American portfolios already hold it. Nvidia buys HBM by the stack; every dollar of memory inflation is a dollar of its bill of materials. The hyperscalers — Microsoft, Alphabet, Amazon and Meta — are converting memory scarcity into higher capital expenditure and longer depreciation tails. If you own a market-cap-weighted US index fund, you are already long the buyer of Korean memory. The question this article poses is whether you also want to own the seller, which currently costs about a quarter of the multiple.
For direct Korea exposure the practical route is a single-country index fund tracking MSCI Korea. Know two things. First, concentration: such a fund is nearly half Samsung and SK hynix, which makes it a two-stock bet with a country label. Second, currency. The won stood at 1,343.77 per dollar on 7 September, up 5.23 percent in a month and its strongest since June 2025, after trading far weaker earlier in the year. Over short horizons the currency can double or erase the price move.
On tax, nothing exotic: US long-term capital gains at 0, 15 or 20 percent plus the 3.8 percent net investment income tax where applicable, short-term at ordinary rates, and the wash-sale rule if you are harvesting losses in a sector this volatile. Korean withholding on dividends is generally reduced by treaty and, for a taxable account, typically recoverable via the foreign tax credit; an exchange-traded fund handles it at the fund level.
What argues against it
The counter-arguments deserve respect, and the strongest is also the oldest. A low price-to-earnings ratio on a record earnings number is not a buy signal; it is the classic marker of a cycle top. Cyclicals look cheap at the peak and expensive at the trough because the denominator moves faster than the numerator. If memory prices fall 30 percent in 2027, today’s 5.3 times will retrospectively have been 15 times. The market is pricing exactly that scenario, and doing so is not irrational — it is the base case from four decades of semiconductor history.
Second, July’s China argument has not been refuted. It has merely left the headlines. Whether and how quickly Chinese DUV tools reach production-grade yields is the real question sitting under the multiple, and it will not be settled by a quarterly report. It will be settled over years.
Third, capacity responds. Record margins are the most reliable invitation to invest that exists. Both Samsung and SK hynix are expanding, and memory cycles have historically died because everyone finished building at the same time. Fourth, the concentration itself: an index that is half two stocks does not have diversification, it has a word for it. And fifth, the macro. US consumer prices for August land on 11 September and the Fed meets on 15 and 16 September; after the payrolls surprise, the odds of a hike are back around even. For a market at 5.3 times earnings, higher US rates are less a valuation problem than a currency and capital-flow problem.
The test that settles it — and the calendar
This piece’s thesis can be falsified cleanly, and it should be. If server DRAM contract prices fall in the fourth quarter of 2026 for the first time in eighteen months while earnings estimates are still rising, then the market was right at 5.3 times and the contracts were the lagging signal. If prices do not fall and the multiple still does not expand, then this is no longer a cycle forecast — it is a permanent risk premium on East Asian geopolitics, and good quarterly numbers will not remove it.
Four dates structure the next few weeks. The US consumer price index on 11 September and the Fed decision on 16 September. In early October, Samsung and SK hynix publish preliminary third-quarter figures — the moment when we learn whether the 104.31 trillion won Samsung estimate held or whether analysts went too far. And through the fourth quarter, the next round of contract negotiations.
Two chart levels are unusually meaningful here rather than merely technical: 7,096.89, the 23 July close and the ceiling of the post-crash recovery, and 9,114.55, the June high. Roughly 28 percent separates them — and that distance is precisely the question at issue.
Monday in Seoul was a good session with a bad explanation. An AI model did not move the index 4.61 percent; two stocks that together are half of it did, in a market that refuses to capitalise its own profits. The real news is the valuation. An equity market whose corporate earnings are nearly quintupling in a single year trades at 5.3 times, while the American market serving the identical demand wave costs around twenty. Either Seoul is right and these profits are a two-year event to be harvested rather than capitalised, or New York is right and demand for compute is structural — in which case the discount the memory makers carry against their own customers is one of the largest valuation anomalies this cycle has produced. Both cannot be true.
Try TradingView Free for 30 Days
Plus get a $15 discount on your first subscription through this link.
Read more in our topic hub: Topic Hub: Quarterly Earnings Tracker 2026
