The AI Memory Supercycle Arrives
The AI memory supercycle has stopped being a thesis and started being a balance sheet. On October 28, SK Hynix reported third quarter revenue of 104.6 trillion won, the first time in its history that any single quarter has crossed the 100 trillion won mark. Operating profit reached 83.3 trillion won, a staggering 632 percent jump from a year earlier, and the operating margin hit 80 percent. For a company that makes commodity memory, a business long defined by brutal boom and bust cycles, those are numbers that belong to a software monopoly rather than a chipmaker.
This is what the AI memory supercycle looks like at full force. Memory, the cheap and interchangeable component that sits beside every processor, has become the chokepoint of the artificial intelligence economy. High bandwidth memory, the specialized stacks that feed data to Nvidia accelerators, is sold out well into the future, and even conventional server DRAM is now scarce. SK Hynix finished the quarter with 114.5 trillion won in cash and a net cash position of almost 99 trillion won, a cushion that would have been unthinkable during the last downturn.
The scale of the swing is the real story. A year ago SK Hynix earned 11.4 trillion won in operating profit for the same three months. The AI memory supercycle has multiplied that figure more than sevenfold in twelve months, and management signaled it expects the momentum to carry on as HBM4, its next generation product, begins shipping in volume.
Why China Wants a Decade of Supply
Nothing captures the panic buying better than what is unfolding in China. According to Korean media reports confirmed on October 27, Alibaba, Tencent, and ByteDance have each opened talks with Samsung Electronics and SK Hynix for long term memory supply agreements, some stretching all the way to 2035. Ten year contracts for a commodity that normally trades on three month spot prices are almost unheard of.
The logic is defensive. Chinese cloud providers are building out AI data centers at speed, and they have watched prices climb while availability shrinks. Locking in a decade of supply, even at a premium, is cheaper than being shut out of the AI memory supercycle altogether. For the Korean manufacturers, these deals convert a famously cyclical business into something closer to a utility with predictable, contracted revenue.
SK Hynix has already signed long term agreements with fourteen major customers that cover more than half of its planned production capacity for 2027 and 2028. In plain terms, the majority of the chips the company has not yet made are already spoken for. When supply is booked years in advance, the usual pressure valve of the memory cycle, a glut that crashes prices, has nowhere to form.
HBM4 and a Shortage That Runs to 2030
The engine of the AI memory supercycle is high bandwidth memory. SK Hynix said HBM4 shipments began in earnest during the quarter and will overtake the previous HBM3E generation in both volume and revenue by the fourth quarter. Each new Nvidia and AMD accelerator demands more memory bandwidth than the one before it, and HBM is the only way to deliver it, which hands the memory makers extraordinary pricing power.
Management was blunt about how long this lasts. The company now expects the memory shortage to persist through 2030, citing not only today's wave of AI training and inference but a second wave of physical AI, the autonomous vehicles and robots that will each carry their own stacks of silicon. That is a remarkable statement in an industry where executives rarely forecast beyond a few quarters.
The demand signal is being amplified across the whole supply chain. Nvidia disclosed supply and capacity commitments of 279 billion dollars as of late July, and this month it lined up six suppliers to expand substrate, packaging, testing, and optical capacity. "Demand is extraordinary. Capacity is the constraint," chief executive Jensen Huang said. The entire ecosystem is now racing to clear bottlenecks that the AI memory supercycle has exposed.
The Scramble to Build Capacity
Everyone who touches an AI chip is raising prices and expanding plants at the same time. TSMC, which assembles the most advanced accelerators, is weighing a 10 to 20 percent increase in the cost of its CoWoS packaging, the technology that bonds logic and memory together. SK Hynix plans to ramp its M15X fab to maximum output by year end and bring its Yongin site into full operation next year, warning investors that investment will rise significantly.
The spending is spreading geographically too. Solidigm, the SK Hynix owned flash unit, is in talks with the United States Commerce Department over a 45 billion dollar NAND plant on American soil. The AI memory supercycle is pulling tens of billions of dollars of fabrication capacity toward the countries that want to control it, a redrawing of the chip map that will take years to finish.
For the Korean economy the stakes are enormous. SK Hynix alone plans to return roughly 100 trillion won to shareholders this year through dividends and buybacks, money that flows straight into the Seoul market. Memory has gone from a cyclical afterthought to the beating heart of a national economy. Readers tracking the broader infrastructure story can follow it in Blumefield's continuing coverage of the AI buildout.
The Risk Hiding Beneath the Boom
Every supercycle carries the seeds of its own correction, and this one is no exception. The same long term contracts that give manufacturers visibility also commit buyers to volumes they may not need if AI spending cools. Nvidia itself has cautioned that customer side constraints, from land and power to buildings and financing, could slow the pace of data center deployment even when components are available.
The warning signs are already audible. The Bank of Korea cautioned this week that a slowdown in American technology spending next year could cut South Korean chip exports by 50 billion to 100 billion dollars, a measure of how concentrated the risk has become. When a single demand driver accounts for an entire industry's growth, any wobble in that driver echoes through balance sheets, stock markets, and national accounts alike. The Bank of Korea is far from alone in watching the data closely.
For now, the AI memory supercycle is a seller's market of historic proportions, and the firms that make the chips are enjoying margins they have never seen. The question that will define 2027 is whether the demand that justified a decade of contracts proves as durable as the spreadsheets assume. If it does, this record quarter will read like an early milestone. If it does not, the industry will relearn its oldest lesson about what goes up. Investors can follow the next chapter with Blumefield.