In the last week of July 2026, two things happened to South Korea at once. Its two largest companies signed the biggest supply commitments in the country’s industrial history. And its domestic retail investors, the cohort that had funded the rally in those same two companies, were being forcibly margin-called out of the trade.
Both facts are true. Neither is a correction of the other. Together they describe a country whose position in the global economy has been permanently rewritten in about eighteen months, and whose own capital markets have not survived the rewriting intact.
On July 24, at an AI summit in San Francisco attended by President Lee Jae Myung, Nvidia and SK Group signed letters of intent covering more than $500 billion of AI infrastructure. The core of it is a long-term partnership under which SK Hynix secures next-generation memory supply for Nvidia and the two co-develop high-bandwidth memory for AI training, agentic workloads, and physical AI. SK Telecom will build a two-gigawatt AI factory running Nvidia’s Vera Rubin systems on SK Hynix HBM4, with the first facility due in 2027. Nvidia, Naver, and Brookfield will expand Naver’s Korean data center operations, with Nvidia putting $1 billion directly into Naver. SK Hynix separately intends to double wafer capacity by 2030.
The next day, Samsung disclosed a memorandum of understanding with Broadcom worth more than $200 billion through 2030. Samsung will supply HBM4 and HBM4E for Broadcom’s next-generation AI accelerators, manufacture Broadcom products on 2-nanometer and below processes at Pyeongtaek, and provide 2.3D and 2.5D advanced packaging on the same node. Samsung is selling a turnkey stack — memory, logic, packaging — rather than a component.
What actually changed is the contract structure, not the demand
The instinct is to read these as demand announcements. They are not. They are scarcity announcements.
Nvidia has more pricing power than any company in the semiconductor supply chain. It does not pre-commit half a trillion dollars over multiple years to lock in a component unless that component is the binding constraint on its own growth. The language from Nvidia’s side was explicit about securing stable HBM supply. Broadcom, similarly, is dual-sourcing away from a single leading-edge foundry not because capacity is cheap but because it is not available.
Korea has therefore been converted from a cyclical exporter into contracted infrastructure. For thirty years the memory business was defined by the fact that customers bought on spot, suppliers overbuilt, and prices collapsed on a schedule. What was signed in San Francisco is the opposite arrangement: multi-year, co-developed, capacity-reserved, with the customer carrying the commitment risk. That is a structural change in the character of the business, and it is the strongest argument that this cycle does not end the way the previous ones did.
It is also the argument most likely to be oversold. Both agreements are letters of intent and memoranda of understanding, not binding volume-and-price contracts. Both were announced at a state-attended summit, packaged into a Korea–US cooperation framework approaching a trillion dollars in headline value, at a moment when trade terms between the two countries remain live. The dollar figures are five-year framework ceilings. They are not backlog, they are not revenue, and they are the first thing that gets repriced if the political context shifts.
The export boom is a price, not a volume
Korean customs data for the first twenty days of July showed chip exports up 180% year on year to $22.1 billion, with computer-related shipments up roughly 232%. Those are extraordinary numbers, and they are being widely read as evidence of accelerating AI unit demand.
They mostly are not. Contract DRAM prices rose 90 to 95% quarter on quarter in early 2026, with NAND up more than 50%, as wafer capacity shifted toward HBM and away from conventional bits. Strip the average selling price effect out of a 180% export figure and the underlying unit growth is a far less dramatic story. This matters in two directions. It confirms that Korean memory makers currently hold genuine pricing power, which is the more durable of the two things a supplier can have. And it guarantees that the year-over-year comparison in the second half of 2027 becomes nearly impossible to clear on anything other than continued price escalation, which is precisely what the announced capacity expansions are designed to prevent.
Meanwhile the capital is moving outward. Korean foreign direct investment into the United States rose more than 100% year on year in the first quarter, to $10.2 billion. Korea’s AI boom is increasingly being built somewhere other than Korea, for customers headquartered somewhere other than Korea, under agreements signed in San Francisco. The wafers are Korean. The demand curve, the design authority, and a growing share of the physical plant are not.
The domestic investor got the exposure and none of the return
Which brings the story to the part that will define the next twelve months more than any memorandum.
Since single-stock leveraged ETFs launched in Korea in May, domestic retail bought a net 14 trillion won of them, against roughly 2 trillion won from foreign investors. The KODEX SK Hynix Single Stock Leverage ETF has fallen about 70% from its June high and roughly 50% since its May 27 debut. The underlying stock fell 19.5% in the month to July 16; Samsung fell 24.3%. The gap between a 19.5% decline in a stock and a 70% decline in a fund tracking twice its daily move is not a mystery — it is what daily rebalancing does to a levered position in a volatile tape, and it is what these products are built to do.
The scale of the buildup was systemic. Assets in the 25 largest leveraged Korea-focused ETFs rose from roughly 15% of the category in January to about 30% by June. Retail accounted for around 70% of trading volume in a $4.3 trillion market. Margin debt hit records. The Bank of Korea warned that leveraged retail positioning had reached historic highs and was concentrated in semiconductors. Oxford Economics downgraded Korean equities to neutral at the end of June, citing leveraged positioning and the likelihood that brokerages would pull back credit. Korea’s exchange has tripped seven market-wide circuit breakers in 2026, more than half of every such event since the mechanism was created. Forced liquidations ran to 2.3 trillion won over roughly two and a half months. The head of the Financial Supervisory Service publicly expressed regret that the products had been approved at all. Regulators then raised the cash requirement to trade them by a factor of ten.
The KOSPI fell 25% from its June peak by July 16. The companies signing the largest contracts in their history were, at that moment, down roughly a fifth to a quarter in a month.
What comes next
Three things determine whether the transformation holds.
The first is conversion. Letters of intent become meaningful when they become take-or-pay volume at fixed price. Watch for the point at which either agreement is restated in a filing with committed quantities. Until then the correct discount on the headline numbers is large.
The second is the supply response. SK Hynix doubling wafer capacity by 2030 and Samsung selling turnkey foundry plus HBM plus packaging are both answers to a shortage, and both land in 2028 and beyond. The question is not whether AI demand persists. It is whether demand in 2029 is still growing faster than the wafers being committed today. Nothing announced in San Francisco addresses that, and the market will begin pricing it well before the fabs open.
The third is the one nobody is modeling. The marginal buyer of Korean chip equity for the past six months was a domestic retail investor using two-times daily leverage. That buyer has been wiped out, forcibly liquidated, and then regulated out of the product by a tenfold increase in the cash requirement. The bid does not come back quickly. Korea has just signed itself into the center of the global AI supply chain at the precise moment its domestic equity market lost the flow that supported its valuation. The gap between the operating business and the share price can persist for a long time when the natural owner of the shares has been removed.
The durable position in all of this is not the Korean memory makers. It is the layer they have to buy from. Doubling wafer capacity, moving to HBM4 and beyond, and adding 2.3D and 2.5D packaging on a 2-nanometer base all increase the tool intensity per wafer regardless of which of Samsung or SK Hynix wins the socket, and regardless of what the memory price does in 2029. Hybrid bonding and atomic layer deposition capture the transformation without carrying the cycle.
South Korea has been changed by the AI boom. Its shareholders have not yet been paid for it.