{"id":448,"date":"2026-07-30T08:06:29","date_gmt":"2026-07-30T08:06:29","guid":{"rendered":"https:\/\/karolpelc.com\/blog\/?p=448"},"modified":"2026-07-30T22:43:46","modified_gmt":"2026-07-30T22:43:46","slug":"it-wasnt-ai-that-broke-it-was-the-markets-machinery-for-pricing-it","status":"publish","type":"post","link":"https:\/\/karolpelc.com\/blog\/it-wasnt-ai-that-broke-it-was-the-markets-machinery-for-pricing-it\/","title":{"rendered":"It Wasn’t AI That Broke. It Was the Market’s Machinery for Pricing It."},"content":{"rendered":"\n

It Wasn’t AI That Broke. It Was the Market’s Machinery for Pricing It.<\/p>\n\n\n\n

How South Korea erased roughly $2 trillion in forty days, while its two biggest companies reported record earnings.<\/p>\n\n\n\n

This is being written mid-crash, with the tape still moving. In July 2026 the KOSPI suffered a record 33% monthly decline, worse than October 1997 at the depth of the Asian financial crisis. Circuit breakers triggered on consecutive days for the first time in the index’s history. In one month the index sliced through its 50-, 100- and 200-day moving averages, all of which had been bullishly aligned when the month began. Drawdowns of this magnitude have happened to Korea before, but they took a year or more to play out. This one took forty days, and it is not obviously finished.<\/p>\n\n\n\n

And here is the statistic that tells you this is not a normal crash: after all of that, the KOSPI closed July 30 at 5,593.56, still up 29.8% year-to-date. SK Hynix, the epicentre of the collapse, is still up 92% on the year. Samsung Electronics, up 60%. The same week the market was halting trading, Samsung reported a 250-fold jump in quarterly semiconductor net income.<\/p>\n\n\n\n

The businesses didn’t break. The market did. Understanding how is the most useful lesson this episode offers, because the machinery that broke in Seoul is running, right now, in most major markets.<\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n

The AI memory supercycle: why the boom was real<\/h2>\n\n\n\n

Every bubble worth studying begins with something true. To understand why Korea, of all markets, went vertical, you need to understand why memory became the bottleneck of the entire AI buildout.<\/p>\n\n\n\n

Modern AI workloads, from training frontier models to inference at scale and increasingly agentic AI, are memory-constrained rather than compute-constrained. A GPU can only process data as fast as memory feeds it, and standard DRAM is too slow. The answer is High Bandwidth Memory: DRAM dies stacked vertically and connected through the silicon itself, delivering data up to 10x faster. Every NVIDIA accelerator ships with HBM bonded directly onto the package. It is a non-negotiable, inelastic input. The HBM market, roughly $35 billion in 2025, is projected to reach $132 billion by 2028, a ~75% annual growth rate. Samsung and SK Hynix together control about 70% of global DRAM. There is no AI buildout without Korea.<\/p>\n\n\n\n

Within that duopoly, SK Hynix became the kingpin. It has led every HBM generation since HBM2E and sits as NVIDIA’s primary supplier, exclusive on the most advanced stacks for the B300 platform, with a cost structure estimated 25-30% better than competitors. The economics are extreme: HBM generated over 40% of its DRAM revenue in Q1 2026 on just ~14% of bit shipments, at operating margins above 70%. Roughly ten long-term agreements gave multi-year demand visibility; the entire 2026 output of HBM, DRAM and NAND was sold out before the year began. In May 2026 SK Hynix crossed $1 trillion in market cap, only the third Asian company ever to do so, after a >900% one-year run.<\/p>\n\n\n\n

Samsung was the catch-up trade. Having missed the early HBM cycle, its 2026 re-rating thesis rested on closing the gap, and it did: NVIDIA certified Samsung’s HBM4 for the Vera Rubin platform in June, with its HBM share projected to nearly double. Meanwhile its scale in conventional memory became an underappreciated earnings machine. As HBM ate wafer capacity (each HBM bit consumes 3-5x the wafer area of standard DRAM), commodity DRAM went into structural shortage. Samsung’s DRAM margins hit a historic 78%, and in July it signed a memory MOU with Broadcom worth over $200 billion through 2030.<\/p>\n\n\n\n

The result was an earnings inflection with few precedents in large-cap history. SK Hynix’s 2026 consensus revenue of \u20a9350 trillion implied +267% versus 2025, against a \u20a933 trillion trough in 2023. Samsung’s 2026 operating profit was forecast at \u20a9370 trillion, up 750%, with the chip division contributing 98% of it. This was not a story stock. It was arguably the fastest profit acceleration ever recorded at this scale.<\/p>\n\n\n\n

The price action, however, went further still. The run-up into the June peak:<\/p>\n\n\n\n

\n \n \n \n \n \n \n \n
<\/th>\n \n SK Hynix
\n Peak: 22 June 2026<\/span>\n <\/th>\n 		\n Samsung Electronics
\n Peak: 18 June 2026<\/span>\n <\/th>\n 		\n KOSPI
\n Peak: mid-June 2026<\/span>\n <\/th>\n <\/tr>\n <\/thead>\n
YTD to peak<\/td>\n +336%<\/td>\n +181%<\/td>\n ~+112%<\/td>\n <\/tr>\n
1 year to peak<\/td>\n +1,051%<\/td>\n +508%<\/td>\n ~+205%<\/td>\n <\/tr>\n
3 years to peak<\/td>\n ~25\u00d7<\/td>\n +406%<\/td>\n ~+255%<\/td>\n <\/tr>\n <\/tbody>\n <\/table>\n<\/div>\n\n\n\n

SK Hynix returned roughly twenty-five times over three years, with a 336% gain packed into the first six months of 2026 alone. Samsung quintupled in twelve months. Both stocks topped within four days of each other, June 18-22: one narrative, one trade, one top. Days after the peak, SK Hynix announced a US ADR listing seeking ~$29.4 billion, one of the largest offerings of all time, timed at the exact top.<\/p>\n\n\n\n

The peak looked like certainty<\/h2>\n\n\n\n

What did the sell side think at the exact top? As of June 30, SK Hynix carried 45 Buys, 1 Hold, 0 Sells; Samsung 44 Buys, 0 Holds, 0 Sells. That is a consensus rating of 4.93 out of 5.00 on both names. Effectively nobody on the street was neutral, let alone negative.<\/p>\n\n\n\n

Targets kept climbing into the peak. Shinhan put SK Hynix at \u20a93,800,000, or 5.8x book for a memory manufacturer. Nomura applied 5.0x book to Samsung for a \u20a9670,000 target, raised on June 23, five days after the stock had already peaked. UBS’s operating profit forecasts sat 24% and 52% above consensus for 2026 and 2027. The bull case rested on long-term supply agreements “making profitability sustainable” and a memory shortage “unlikely to resolve before 2027.”<\/p>\n\n\n\n

None of these arguments were foolish. That is precisely the point. When 89 of 90 ratings on a market’s two dominant stocks are Buys, the question is no longer whether the thesis is right. The question is who is left to buy.<\/p>\n\n\n\n

Then the structure took over<\/h2>\n\n\n\n

Four mechanisms turned a rational re-rating into an unstable system.<\/p>\n\n\n\n

Concentration<\/strong>. By mid-2026, Samsung and SK Hynix together represented more than half of the KOSPI’s market capitalisation, at the peak close to 58%. This wasn’t anyone’s decision; it’s arithmetic. As the two stocks rose, their index weights rose, which pulled in more passive money, which pushed prices higher, which raised the weights again. Two companies effectively became the market.<\/p>\n\n\n\n

Leverage<\/strong>. Korea combines high household savings with housing that has become unattainable for much of a generation. For many younger Koreans, the stock market replaced property as the path to wealth, and they used borrowed money to walk it faster. Debt used for equity investment exceeded \u20a960 trillion by late May. Margin loans hit a record \u20a938.63 trillion on June 24, within days of the market top, which is precisely when margin debt always peaks.<\/p>\n\n\n\n

Product design<\/strong>. Leveraged single-stock ETFs launched domestically in May 2026 and were adopted at speed. These products rebalance daily to maintain constant leverage: they buy after prices rise and sell after prices fall, mechanically, regardless of value. In a rising market, an invisible tailwind. In a falling one, a forced seller arriving every afternoon.<\/p>\n\n\n\n

One-way ownership<\/strong>. Foreign investors had withdrawn roughly US$110 billion from Korean equities during 2026, well before the collapse. Leveraged domestic retail absorbed the supply. The ownership base shifted from diversified, patient capital to concentrated, borrowed capital, the opposite of what you want going into a shock.<\/p>\n\n\n\n

The cascade: one week in July<\/h2>\n\n\n\n

On Tuesday, July 28, the index fell as much as 7.6%, dragged by SK Hynix (-11%) and Samsung (-9%), as AI fatigue deepened and crowded positioning began to unwind. On Wednesday morning it plunged as much as 13%, triggering a circuit breaker for the second consecutive day, unprecedented in KOSPI history, as SK Hynix’s earnings disappointed a market primed for perfection. That evening, Finance Minister Koo Yun Cheol convened an emergency meeting of all top financial authorities. The measures announced, market stabilisation steps and curbs on retail access to leveraged ETFs, were dismissed by market watchers as “tinkering at the margins.”<\/p>\n\n\n\n

Thursday showed how broken the price-discovery mechanism had become. Samsung reported that 250-fold jump in semiconductor net income, a massive beat, and the stock managed just +0.7%. The index surged as much as 5.5% on the news, then gave it all back to close down 1.2% at 5,593.56, its lowest close since April 7. Record earnings could no longer move prices, because prices were no longer being set by earnings. They were being set by margin calls, daily ETF rebalancing and forced liquidation: a loop in which falling prices trigger selling that has nothing to do with what anything is worth. The won slipped toward 1,447 per dollar, and the rout dragged the MSCI Emerging Markets Index to a three-and-a-half-month low.<\/p>\n\n\n\n

Most telling of all: with the KOSPI trading at its cheapest valuation ever recorded, global investors including Pictet, Robeco and Eastspring publicly declined to buy the dip, citing volatility too violent to stomach. When forced sellers dominate a tape, even cheap markets find no bid. Analysts now say a daily close above 6,000 is the minimum signal of a credible recovery. It hasn’t come.<\/p>\n\n\n\n

The lesson, and why it travels<\/h2>\n\n\n\n

The comfortable reading of Korea is “retail mania in a small, concentrated market.” The uncomfortable reading is that every ingredient exists elsewhere, at larger scale. Index concentration in the US is at multi-decade highs, with a handful of AI-linked mega caps driving benchmark returns. Passive ownership keeps growing, meaning ever more capital buys weights, not valuations. Leveraged and options-based products increasingly dominate short-term flows. And AI is the single narrative underwriting all of it: real, transformative, and priced as if years of future growth were already banked. Korea wasn’t a different machine. It was the same machine with fewer shock absorbers: two stocks instead of seven, faster leverage adoption, a smaller pool of institutional capital to lean against the wind.<\/p>\n\n\n\n

None of this says AI is a mirage. Hyperscalers are still spending roughly $800 billion this year, and on July 30, mid-crash, Samsung told the market that memory supply constraints will worsen in 2027 and persist through 2028, while SK Hynix plans to double capacity against a shortage its chairman sees lasting to 2030. The supercycle thesis did not die in July; the positioning built on top of it did.<\/p>\n\n\n\n

It says something more precise: great businesses become dangerous investments when market structure removes investors’ freedom to choose whether they sell<\/strong>. A market can be up 30% on the year, its champions up 60-92%, its earnings at all-time records, and still be in freefall, because earnings determine value over years while concentration, leverage and mechanical flows determine prices over weeks.<\/p>\n\n\n\n

The question to ask about your own market is not whether the fundamentals are strong. Korea’s were; 89 Buy ratings out of 90 said so. The question is: if prices fell 20%, how much of the selling that followed would be voluntary?<\/p>\n\n\n\n

In Seoul this July, the answer was: almost none of it.<\/p>\n","protected":false},"excerpt":{"rendered":"

It Wasn’t AI That Broke. It Was the Market’s Machinery for Pricing It. How South Korea erased roughly $2 trillion in forty days, while its two biggest companies reported record earnings. This is being written mid-crash, with the tape still moving. In July 2026 the KOSPI suffered a record 33% monthly decline, worse than October […]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[],"class_list":["post-448","post","type-post","status-publish","format-standard","hentry","category-uncategorized"],"blocksy_meta":[],"_links":{"self":[{"href":"https:\/\/karolpelc.com\/blog\/wp-json\/wp\/v2\/posts\/448","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/karolpelc.com\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/karolpelc.com\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/karolpelc.com\/blog\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/karolpelc.com\/blog\/wp-json\/wp\/v2\/comments?post=448"}],"version-history":[{"count":3,"href":"https:\/\/karolpelc.com\/blog\/wp-json\/wp\/v2\/posts\/448\/revisions"}],"predecessor-version":[{"id":453,"href":"https:\/\/karolpelc.com\/blog\/wp-json\/wp\/v2\/posts\/448\/revisions\/453"}],"wp:attachment":[{"href":"https:\/\/karolpelc.com\/blog\/wp-json\/wp\/v2\/media?parent=448"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/karolpelc.com\/blog\/wp-json\/wp\/v2\/categories?post=448"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/karolpelc.com\/blog\/wp-json\/wp\/v2\/tags?post=448"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}