In the cleanrooms where SK Hynix etches memory chips, the machines are running flat out, and the chief executive says they will be for years. Kwak Noh-jung, who runs the world’s second-largest memory maker, said the global shortage of memory chips is expected to last until the end of 2030, with little risk of the oversupply that has punctuated past cycles.
In remarks reported on Sunday, Kwak said memory chips have stopped being a simple standardized commodity in the AI era. Demand is now differentiated and customized, tied to specific AI systems, which changes how the industry balances supply and demand. He said he sees no signs of oversupply or of a looming memory downturn, because demand from AI customers remains strong.
Kwak acknowledged the risk explicitly. If the industry has crossed the AI peak, or if another downturn arrives, challenges will appear, he said. But he argued that the next downturn would be different from the slumps of the past few decades: not a sharp collapse, but a slow easing of demand, possibly a plateau.
His confidence carries weight because he runs the company that was burned hardest by the last crash. In 2022 and 2023, SK Hynix’s profits collapsed as memory prices fell more than 70 percent from their peaks, and the company posted deep losses before AI demand rescued it. The memory industry’s institutional memory is full of booms that ended in ruin, which is why a CEO predicting six more years of shortage is a notable statement.
The supply picture supports him so far. AI data centers consume high-bandwidth memory, HBM, the vertically stacked chips that sit next to accelerators, and standard DRAM in record volumes. Every wafer dedicated to HBM is a wafer not making regular chips, and HBM capacity does not convert back to commodity DRAM on demand. The shortage is compounded by a slow-motion capacity build: fabs take two to three years to construct and qualify, and equipment lead times remain long.
Others in the industry are reaching similar conclusions. Samsung and Micron have signaled tight supply into 2027, and the three major makers have sold their 2027 production capacity in advance, according to industry reports. The question analysts are asking is not whether supply is tight, but what happens if AI investment stalls: a demand cliff would arrive faster than new capacity, and the industry would face the same glut it faced in 2022, only with more capacity in the ground.
Kwak’s answer is that the structure has changed. In the past, memory was a commodity bought on price, and the cycle was driven by overbuilding. Now, HBM stacks are co-designed with individual customers, contracts run years long, and the largest buyers have locked in supply in advance. The industry, in his telling, has moved from selling chips to selling commitments.
The comments frame the bet underpinning a wave of new investment. SK Hynix is building new fabs in Korea, including a sprawling complex south of Seoul, and is weighing a joint venture in Japan, its first serious consideration of a fab outside Korea. The company is also preparing to hand part of its HBM production to outside foundries, a supply-chain decision that would have been unthinkable a decade ago.
The stakes extend beyond the company. Memory has become a pricing signal for the entire AI boom: spot prices for HBM have reached multiples of contract prices, and the cost of standard DRAM is climbing again. When the memory cycle turns, it moves the economics of every server, every phone and every AI data center built on top of it.
The other side of the ledger is the industry’s capacity discipline. Memory makers were so badly burned by the 2022 crash that they have refused to build speculative capacity since, and the current shortage is partly a product of that caution: even with prices rising, none of the three major makers has announced a greenfield fab on the scale the market would have built in a comparable boom a decade ago. The result is a supply curve that rises slowly, on purpose.
That discipline is also a bet. If AI demand keeps growing, the industry will earn exceptional returns for years, and the shortage will extend precisely because suppliers refuse to overshoot. If demand stalls, the industry will be smaller than it might have been, but profitable rather than devastated. The memory makers have, in effect, chosen to be late rather than early, and the CEO’s 2030 horizon is the public statement of that choice.
Kwak’s comments also carry a message for customers. The largest buyers have already begun signing long-term agreements and reserving capacity years in advance, and the message to everyone else is that the era of buying memory off the shelf, at spot prices, when needed, is over for the foreseeable future. Procurement teams that have treated memory as a commodity input are being forced to treat it like power: contracted, planned and hedged.
For now, Kwak said the company sees no signs of the cycle turning, and his words will be tested by the capacity decisions being made today. Every fab announced this year comes online around 2029 or 2030, precisely when he says the shortage ends. The industry has heard confident CEOs before, and the memory market has a long history of proving them wrong. The difference this time, he argues, is that demand is under contract years in advance, and the contracts are signed by the largest companies in the world.
