The AI Correction Landed in Seoul First

In the week of June 16, 2026, two markets looked at the same boom and saw opposite things. US options trading had analysts reaching for a word they rarely use: euphoric ^[1]. That same week, Seoul's Kospi peaked and turned ^[2]. South Korea and Taiwan are Asia's hardware hubs, the two economies that build most of the world's AI chips. The correction has a start date and an address, and the address is Seoul. From that June peak to the end of July, the Kospi fell 39 percent, dragged by Samsung Electronics and SK Hynix, while US indexes sat in a holding pattern ^[2]. Its only interruption was a mood, not a fundamental: an investment firm named a small solar supplier its top under-the-radar AI pick on data-center power demand, and Seoul briefly chased it back up — a fear-of-missing-out rebound that has since been wiped out ^[3]. On Tuesday, October 13, the index dropped 7.4% to 5,914.35 ^[4], Samsung down 9% and SK Hynix down 10% ^[5]. The day had co-drivers — Russian strikes on Ukrainian energy and a Houthi attack on Riyadh that pushed Brent toward $105.9 ^[4] — and the proximate spark was Chinese memory maker ChangXin's announcement of a denser server chip, a trigger analysts called disproportionate, since CXMT trails the leaders by two or three generations and does not compete in high-bandwidth memory at all ^[6]. None of that changes the arithmetic: this is the second AI-anxiety crash in a year for a market that first fell below 4,000 in November 2025 on the same fear ^[7]. The American ledger reads almost serene by comparison. The S&P 500 ended July down a tenth of a percent ^[8], and UBS strategists had a name for what had just happened.

We see it as a healthy correction in an AI spend cycle that is just 2 yrs in (vs ~10yrs for prior Tech cycles). — De'Mond Parker

To be fair, the US did take a hit in that window. A four-day selloff in late July erased about $1.2 trillion of tech value on rising capex and missed earnings, but it was rotation-scale damage, Nvidia down just 5 percent ^[9]. That is exactly what makes the divide geographic rather than a claim that America escaped. By October 4, nine days before Tuesday's rout, US indexes were pressing record highs, the Nasdaq 100 up 22 percent on the year ^[10]. So what has actually been repriced? Apollo's chief economist Torsten Slok put the mechanism in one sentence.

What the market is repricing is hyperscaler credit fundamentals, namely a debt-financed AI capex cycle with rising leverage, negative free cash flow and uncertain payback on depreciating assets. — Torsten Slok

Not demand, then: the debt beneath the demand ^[11]. And that repricing has already run on the buyers — the US companies that buy what the Asian hubs supply — inside the US complex itself. Oracle is down more than 40 percent and Broadcom about 25 percent since early June, while Nvidia held because a 30 percent rise in forward earnings estimates offset a shrinking multiple ^[12]. S&P cut Oracle to BBB-, the lowest investment-grade rung, calling it a "show me" story: revenue projected up 239 percent against debt up 412 percent, with 2027 capex guidance raised from $60 billion to $95 billion ^[13]. Its Jupiter loans trade at 89 to 91 cents on the dollar ^[14]; its credit-default swaps sit at a record 261 basis points, and Abilene — one of its AI data-center campuses — carries $9.4 billion of construction loans due in 2028 ^[15]. That is the seam, not the whole story, but it explains the geography. Meanwhile the physical boom never paused. Memory contract prices jumped nearly 20 percent in the very quarter the Kospi lost 39 percent ^[16], and the shortage is locked in through 2027 ^[17]. The same IMF that on Tuesday named an AI share-price correction a top-three threat to global financial stability ^[18] also raised South Korea's 2026 growth forecast to 3.3 percent and Taiwan's to 10.7 percent, the two hubs that together source 62 percent of net trade in AI computing hardware ^[19]. The repricing has been exact about where it has landed: the suppliers, and the most indebted buyers. The one ledger it has not reached is the broad US index itself, where Alphabet, Amazon and Meta's free cash flow has just turned negative ^[10].