SEOUL — SK Hynix, the world’s largest maker of high-bandwidth memory, is adjusting the pace of its sixth-generation HBM4 mass production and shifting resources to commodity DRAM, the company said, as shortages push prices for standard memory to levels that make the shift profitable. The move is a rare example of a memory maker choosing breadth over depth in the middle of an AI boom.
The company’s reasoning is arithmetic. HBM now accounts for more than 40% of SK Hynix’s revenue and holds a dominant position in the AI accelerator market, according to the company, but the demand for commodity DRAM — the memory that goes into personal computers, servers and smartphones — has become so severe that the economics favor reallocation. Rather than expand capacity blindly, the company said it is redirecting existing resources to capture more revenue from the general DRAM market.
The shortages are the product of years of capacity discipline. Memory makers spent most of the last downcycle cutting production and delaying new factories, and when AI demand arrived, the industry had no spare capacity to absorb it. Data center demand for HBM pulled wafer capacity away from commodity DRAM, and the resulting shortage has pushed prices up sharply, with spot prices for some DRAM products reaching levels not seen in years.
The shift has implications for customers across the technology industry. Computer makers have warned of higher memory costs, Apple has said price increases are coming, and server builders face allocation decisions. SK Hynix’s decision to reallocate capacity suggests the company believes commodity DRAM prices will stay high enough to justify the move, a bet that analysts read as a signal of confidence in the breadth of the AI-driven cycle.
The reallocation is not a retreat from HBM. SK Hynix remains the leading supplier of the memory stacked inside Nvidia’s accelerators, and its HBM4 products are in qualification with major customers, according to people familiar with the process. The company’s statement described the adjustment as a matter of pacing, not abandonment, and analysts said the shift is possible precisely because HBM’s margins and market position are secure enough to spare capacity.
The balance between HBM and DRAM is the central management question for memory makers. HBM commands premium prices but requires complex packaging and yields, while commodity DRAM sells in far larger volumes with simpler manufacturing. A company that overcommits to HBM risks missing the broader market; one that undercommits risks losing its position in the most profitable segment. SK Hynix’s move suggests its management believes the general DRAM market has reached a point where the marginal won of revenue is better spent there.
The broader memory industry is making similar calculations. Samsung, SK Hynix’s larger rival, has accelerated its own HBM qualification efforts while maintaining commodity production, and Micron has signaled that its capacity is sold out through the year. The collective effect of these decisions is that DRAM prices are likely to stay elevated, which feeds the pass-through to consumers that has become a theme of the current cycle.
The timing of the announcement mattered. It came on the same day the company’s shares fell 12% in the global technology selloff, and a day after SK Hynix surpassed Samsung to become South Korea’s most valuable company. The message to investors was that the company is managing its product mix actively rather than riding a single wave, and that the memory cycle’s strength extends beyond the AI chips that have dominated the headlines.
The risks are the same ones that have defined every memory cycle. If the AI boom slows, HBM demand could soften just as the industry’s new capacity comes online, and the commodity DRAM market, which is already absorbing a flood of reallocated capacity, could swing from shortage to glut. Memory makers have been burned by exactly this sequence before, and the industry’s discipline in the current cycle is a response to the scars of the last one.
The market has already responded to the shortages. Memory contract prices for DRAM have risen for several consecutive quarters, according to industry data, and suppliers have said allocation, not price, is now the binding constraint for many customers. Server makers have been forced to queue for memory, and PC makers have begun shipping machines with less memory than they would prefer in order to manage costs. The situation has made memory the most discussed component in the technology industry, displacing the GPU shortage that dominated headlines in earlier phases of the AI boom, and the companies that control supply have found themselves in a position of unusual power.
The strategic implications extend beyond the current quarter. Memory is a commodity business with brutal cycles, and the industry’s discipline in the current upcycle, with suppliers resisting the temptation to flood the market with new capacity, has been the exception to its own history. The question is how long the discipline holds: if prices stay elevated, the economics of building new fabs improve, and the capacity that is now being planned could arrive just as demand softens, recreating the boom-bust pattern that has defined the industry for decades. SK Hynix’s careful pacing, shifting resources rather than building blindly, is an acknowledgment of that risk, and the company’s management has said it intends to manage the cycle rather than be managed by it.
For now, the reallocation is a bet on the breadth of demand. SK Hynix is telling the market that the AI boom has made memory scarce across the board, not just in the specialized products that serve data centers, and that the company intends to collect the gains wherever they appear. The shift from HBM4 to commodity DRAM is a management decision with a simple logic: when everything is in short supply, sell what the market will pay for.
