A Trend Report | July 2026 On July 28, the South Korean stock market stopped looking like a market and started looking like a controlled demolition. The KOSPI lost more than 10 percent in a single session, triggering another market-wide circuit breaker. Samsung Electronics fell more than 13 percent. SK Hynix fell more than 14 percent. Leveraged ETFs built on those two stocks — vehicles that had turned ordinary Korean retail investors into paper millionaires only weeks earlier — went into a freefall that would later be measured in terrifying round numbers. It was the kind of session that normally happens once in a generation. By late July, it was happening roughly once a week. Just six weeks earlier, the KOSPI had set an all-time high of 9,385.59. On July 29, it closed at 5,663.24. That is a drop of about 40 percent in the kind of short window that investors usually associate with emerging-market currency crises or the collapse of a leveraged housing bubble. The Philadelphia Semiconductor Index, which had more than doubled over the previous year, fell more than 20 percent from its late-June record and entered bear-market territory. The Direxion Daily Semiconductor Bull 3X ETF, a favorite tool of momentum traders, lost more than half its value from its June peak. Here is the strangest part: none of this happened because the semiconductor industry stopped making money. In the first 20 days of July, South Korean semiconductor exports reached $22.1 billion, up 180.6 percent from a year earlier. Chips accounted for 40.3 percent of the country's total exports, up from 21.9 percent a year earlier. The first ten days of July alone saw semiconductor exports surge 193 percent year on year to a record $11.2 billion. Global memory inventory levels were so tight that distributors were measuring supply in weeks, not months. Samsung Electronics reported an operating profit of 89.5 trillion won for the second quarter, up 1,814 percent year on year. SK Hynix reported an operating profit of 60.5 trillion won, up 557 percent. So the correction was not about collapsing demand. It was about the price of a trade that Korea — and much of the global investment community — had become dangerously dependent on.

South Korea's economic transformation over the past three years is difficult to overstate. Semiconductors have gone from roughly one-fifth of the country's exports to more than 40 percent. Samsung Electronics and SK Hynix together now account for about 52 percent of the KOSPI's total market capitalization. The country's five major conglomerates contribute more than 40 percent of export value. No other major economy has concentrated its national fortune so tightly around a single technology category. That concentration has been enormously profitable. Korean semiconductor exports grew at a pace that looked like a typo — 162.6 percent year on year in the first half of 2026. The country's current account surplus is expected to approach $300 billion, nearly double the previous year. Government tax revenue from the chip boom has been so strong that policymakers created a "Future Response Fund" to recycle excess corporate tax payments into AI infrastructure spending. But there is a difference between being a winner and being a hostage. The same data that made Korea look invincible also made it fragile. If AI memory demand slows even modestly, the country's export engine, its corporate profit pool, and its stock market are all exposed at the same time. The economic term is "concentration risk." The more honest description is that Korea used its once-in-a-generation technological advantage to make a single bet — and then made that bet even bigger by letting retail investors leverage it. The employment numbers tell the uncomfortable side of the story. Between 2023 and 2025, South Korean semiconductor exports grew about 75 percent, but the industry added only about 1,000 net jobs. The automation intensity of chip manufacturing means the export boom has not produced broad-based prosperity. Instead, it has produced a peculiar kind of boomtown economy: record tax revenue, record corporate profits, record capital spending — and a labor market that has not improved nearly as much as the headline numbers suggest. In the second quarter of 2026, more than 480,000 university-educated Koreans were unemployed, a five-year high. Youth unemployment rose to 7.2 percent. The AI boom did not create a middle class in Korea. It created a concentrated class of chip winners and a much larger group of people who could only participate through the stock market.

The Leverage Trap

What transformed a routine correction into a rout was not weak fundamentals. It was leverage. In late May, the Korea Exchange approved 16 new 2x leveraged ETFs tracking only two stocks: Samsung Electronics and SK Hynix. These weren't exotic products for institutional players. They were marketed, discussed, and traded by ordinary retail investors who had watched the KOSPI double and wanted a way to double faster. Within roughly two months, assets under management in leveraged products went from about 5 trillion won to 76 trillion won. Korean retail investors bought 99.2 trillion won of Korean stocks in the first half of 2026, a national record. The problem with daily-reset leveraged ETFs is that they are designed to multiply one-day returns. They tend to bleed in volatile markets even when the underlying stock goes sideways. When the underlying stock goes down sharply, the compounding effect is brutal. By mid-July, the losses were no longer theoretical. Ritesh Jain of Pinetree Macro estimated that around 1.2 million leveraged retail accounts hit margin calls on July 13 alone, with something like 320,000 to 460,000 accounts fully liquidated. Citigroup analysts calculated that roughly one in 30 Korean adults had been caught in the forced-selling wave. Aggregate retail losses on leveraged products reached 56.3 trillion won. The human toll showed up in Reddit threads and local media reports. One Korean office worker told reporters he had invested 80 million won of his wedding savings into a leveraged ETF and now faced the prospect of postponing his marriage. Another retail investor who had made five years' worth of salary during the rally said she had given it all back in a matter of days. In a widely shared post, a Reddit user insisted, "Even if SK Hynix stock falls to zero, I won't sell." That may sound like conviction. It also sounds like denial. The regulatory response arrived late and in pieces. The Financial Services Commission raised margin requirements from 10 million won to 30 million won, limited individual leverage exposure, and expanded the minimum trading unit from one share to twenty. The government said it had reserved about 10 trillion won for a market-stabilization fund. None of these measures stopped the bleeding. The KOSPI kept falling because most of the forced selling was coming from leveraged positions that had already been triggered.

The Bulls Are Not Wrong

After a 40 percent crash, it is easy to dismiss everyone who was bullish as having been delusional. That would be a mistake. The bull case for AI memory is built on a set of facts that have not changed. Samsung and SK Hynix are not merely optimistic about AI demand; they have signed contracts that lock in demand for the next half-decade. Samsung has announced long-term supply agreements with the world's five largest data center operators, with five more close to being signed. These agreements are structured to cover roughly 60 to 70 percent of Samsung's long-term production capacity, run for at least five years, and include prepayments and minimum pricing provisions. That is not a hope. That is a firewall. SK Hynix has been even more aggressive in using long-term agreements. The company said it has signed LTAs with ten industry customers, with differentiated pricing mechanisms designed to smooth the industry's historical boom-and-bust cycles. Management expects DRAM demand to grow in the mid-20 percent range and NAND demand in the high teens. On the company's earnings call, executives stated that AI infrastructure investment "will remain strong beyond 2027." The company's LTAs are structured to match customer demand timing rather than lock in fixed quantities, which SK Hynix says should avoid the inventory gluts that wrecked memory markets in previous cycles. Samsung's memory business executive vice president, Jaejune Kim, was even more direct. He told analysts that the supply shortage in 2027 is expected to worsen compared with this year, and that the shortage is likely to continue into 2028. Given that Samsung controls close to 38 percent of the global DRAM market and roughly 21 percent of the HBM market, that statement carries real weight. The supply-side math supports the bullish view. HBM consumes roughly three times as much wafer capacity as standard DRAM. Even with aggressive capacity expansion, the industry's overall wafer shortage is estimated at 50 to 60 percent. TrendForce expects HBM to consume about 22 percent of DRAM wafer input by the end of 2026 and 30 percent by 2027 — yet HBM's share of total DRAM bit supply will still be only around 9 percent in 2026 and 13 percent in 2027. In other words, the industry is pouring an enormous and growing share of its capacity into a product that does not produce very many bits. That is the definition of structural tightness. Wolfe Research senior analyst Chris Caso put it bluntly: "There is simply not enough physical space to manufacture more semiconductors." He expects the AI chip supply constraint to persist well beyond 2028.

The Bears Are Not Wrong Either

The bear case is simpler, and for that reason more uncomfortable. The companies paying for all this infrastructure — the hyperscalers — are not charities. They are spending staggering sums on AI capex because they believe they have to, not because they enjoy it. UBS estimates that hyperscaler capital expenditure will rise 76 percent in 2026 to $673 billion. That is an extraordinary number. But in 2027, the firm expects growth to slow to 25 percent. In 2028, just 6 percent. JPMorgan's projection is even sharper: roughly 100 percent growth in 2026, dropping to 22 percent in 2027 and 7 percent in 2028. The question is not whether hyperscaler spending will keep growing. The question is whether it can keep growing fast enough to justify the earnings expectations embedded in semiconductor stock prices. As Empirical Research noted, there is now a growing mismatch between moderating capex growth and lofty revenue expectations for chip suppliers. Either the hyperscalers will have to upgrade their capex trajectories again, or the revenue growth penciled in for their suppliers will have to come from somewhere else. BofA's July global fund manager survey showed that 82 percent of fund managers viewed semiconductors as the most crowded trade on earth. Zero percent reported being short the sector. For market historians, that combination — maximum crowding, zero positioning against it — is itself a warning sign. Some professional investors have started quietly acting on that warning. Alexis Bossard, global equity portfolio manager at Edmond de Rothschild Asset Management, told Reuters he had cut his semiconductor exposure. "Once they stop increasing their capex, it will definitely be a relief for hyperscalers and a negative signal for the semi industry," he said. Meanwhile, hedge funds have reduced their net exposure to semiconductor and momentum stocks by roughly 5 percent of total market cap, one of the largest reductions on record. The "big short" crowd is not exactly cheering, but Michael Burry announced a bearish position on AI and semiconductor stocks, including Nvidia. And then there is the uncomfortable detail hiding inside the Korean export data. The export boom is real, but it is overwhelmingly price-driven. DRAM contract prices rose 90 to 95 percent quarter over quarter in the first quarter of 2026. Samsung's overall average selling price jumped 146 percent compared with 2025 averages. But SK Hynix's bit shipment growth in Q1 was nearly flat. That means the current export miracle is not a story of Korea shipping more memory. It is a story of Korea charging much, much more for the same amount of memory. That is an extraordinary position to be in while supply is tight. It is also a fragile one. If AI memory prices ever start to normalize, the export numbers will not merely slow. They will collapse.

A Cycle That Has Broken Its Own Rules

The semiconductor industry has historically operated on a simple rhythm: boom, overcapacity, bust, consolidation, repeat. The current cycle has refused to follow the script. At a June seminar hosted by the Korea Economic Research Institute, senior research fellow Do Young-woong showed how the standard inventory cycle broke down. In previous cycles, the market moved from peak boom into an early downturn where both inventory and shipments declined. This time, inventory has decreased while shipments continue to increase. The semiconductor market entered a boom phase in June 2023, hit a peak in March 2024, and appeared to be turning down — but then, in Do's words, "the clock turned backward." By April 2026, the industry was somehow still in the peak boom phase. "The pattern moved clockwise smoothly until March 2024, then became tangled like a knotted thread," Do said. "It's difficult to tell from the inventory cycle whether it will move left again into a downturn or maintain the boom." The obvious culprit is the unprecedented scale of AI infrastructure spending. In the first quarter of 2026, capex at the four largest hyperscalers rose as much as 108 percent year on year, with total spending expected to reach around $725 billion for the full year. To put that in context, Amazon alone raised its 2026 cash capex guidance by $20 billion in July, bringing the total to $220 billion. OpenAI has lifted its compute spending plan through 2030 from $600 billion to $750 billion. The long-term agreements add another layer of complexity. Korea Investment & Securities research head Yoo Jong-woo notes that LTAs now include price-binding provisions, which is a significant change from past cycles. "The earnings visibility of semiconductor companies has greatly improved," he said. But LTAs cut both ways. They may provide a floor under prices, but they can also cap the upside for chipmakers while handing more leverage to hyperscalers. SemiAnalysis, the semiconductor industry research firm, argued that SK Hynix actually missed investor expectations in the second quarter precisely because the shift toward LTAs slowed the pace of price increases. In other words, the very mechanism that protects memory makers from a downturn also keeps them from fully exploiting a shortage. That is a reasonable trade-off in a mature industry. It is a frustrating one in a mania.

Korea's Answer: Bet Bigger

One might expect a 40 percent stock market crash to cool South Korea's national ambitions in AI and semiconductors. Instead, the government has doubled down. In late June, President Lee Jae-myung's administration unveiled a package of more than $576 billion in chip investment. Samsung and SK Hynix plan to invest around 800 trillion won — roughly $518 billion — to build two new fabrication sites each in the country's southwest. Samsung Group separately pledged 1,000 trillion won for Korean investment over the next decade across chips, AI data centers, batteries, and displays. Across all initiatives, the government's total AI and semiconductor investment commitment is estimated at 1,461 trillion won. The breakdown is striking: 800 trillion won for semiconductor production bases in the southwest, 81 trillion won for HBM packaging clusters in the Chungcheong region, 550 trillion won for AI data centers, and 30 trillion won for next-generation semiconductor R&D. The funding plan includes a Future Response Fund built on excess corporate tax revenue, plus a new sovereign wealth fund that will receive an initial injection of at least 20 trillion won from state-owned stakes and inheritance tax proceeds. Industry Minister Kim Jung-kwan described the moment in almost existential terms: "Semiconductors and AI robots are a national survival race. All fiscal and institutional means must be mobilized." The critics are not silent. Building four massive fabs in a country that already struggles with electricity supply and construction labor will not be easy. The power demand from the four planned facilities alone could equal 70 to 80 percent of the current annual electricity consumption of the country's southwestern regions. Professor Lee Jong-ho of Seoul National University warned that demand might not stay strong for the full 20- to 30-year life of these investments. "It would be ideal if demand remained strong," he said, "but no one can know that with certainty. If demand were to decline, the consequences would be severe." That is the Korean conundrum in one sentence. The country is so dependent on the AI trade that it cannot afford to slow down. But the faster it runs, the more expensive a stumble becomes.

What Comes Next Is Not a Forecast

No one can confidently predict where the KOSPI goes from here. The disagreement among smart people is too large and too honest. Goldman Sachs has maintained a 12-month KOSPI target of 12,000 points, arguing that the index's decline is far out of line with the earnings outlook. Citi has described the pullback as a buying opportunity. Meritz Securities argues that the market is misjudging the fundamentals, noting that DRAM demand satisfaction rates are only 75 to 80 percent in the second half of 2026, with supply gaps widening. Morgan Stanley, by contrast, has turned bearish on storage chips, pointing to NAND module makers' inventory rising to about 13 weeks and cooling demand downstream. The bank has also warned that Chinese memory producers are expanding faster than the market expects. CXMT, China's leading DRAM maker, saw its global market share jump to 8 percent in the first quarter of 2026, and its capacity is expected to grow from roughly 280,000 to 290,000 wafers per month at the end of 2025 to around 500,000 by 2028. YMTC has pushed its share of the NAND market to about 13 percent. In a downturn, that added Chinese capacity could be ruthless. The honest answer is that the semiconductor market's future will be determined by a narrow set of variables. The first is what hyperscalers actually say about 2027 capex. The second is how quickly HBM4 ramps and how aggressively prices rise. The third is whether the LTA structure holds up in a downturn or starts to unravel. The fourth is whether governments and companies build so much capacity that the 2028 shortage becomes a 2029 glut.

The Haunting

The 2026 semiconductor pullback is a stress test of the AI trade and of South Korea's decision to hitch its entire economic wagon to that trade. At its core, the problem is not that AI demand is falling. All evidence says AI demand is still growing at a pace that nearly any industry would envy. The problem is that the market had priced the AI trade not for strong growth, but for uninterrupted perfection. And in a famous cyclical industry, perfection is a very fragile assumption. That is why the Korean market feels haunted even as Samsung and SK Hynix report record profits. It is haunted by the possibility that the world's most important technology transition is happening at a speed that cannot be sustained. It is haunted by the knowledge that Korea's export machine, fiscal health, corporate earnings, and stock market have all become functions of one number: the AI memory price. "Valuations in semiconductor stocks had priced near-perfect demand, for what has been a cyclical area in the past, so was always going to leave stocks vulnerable at some point in what has been a rapid rise," said Toni Meadows of BRI Wealth Management. What makes this pullback harder to process is that it came with no bad news attached. There was no demand shock. There was no inventory bubble. There was no surprise downgrade from a major hyperscaler. There was only a market suddenly noticing that everyone was on the same side of the boat — and then scrambling to get off first. By late July, the KOSPI's 12-month forward price-to-earnings ratio had fallen to 4.5x, more than three standard deviations below its historical average. Samsung Electronics was trading at about 4.8x forward earnings, its lowest level since 2000. By historical standards, the market is cheap. But cheap can keep getting cheaper when the economy's largest growth engine is at the center of a global leverage unwind. The question now is not whether the AI memory cycle has ended. Nothing in the supply data says it has. The question is whether the price of that cycle had become a bubble on top of a boom — and whether the human and economic damage from that bubble's unraveling will shape the next stage of the AI trade. South Korea's AI dependency is no longer an abstract policy concern. It is a lived reality, measured in circuit breakers, margin calls, and the smashed accounts of ordinary investors who believed that the AI boom would never look down. The memory industry has a history of teaching that lesson the hard way. This time, the lesson is not about demand. It is about price, leverage, and what happens when an entire country becomes the trade.