Nvidia, the company driving the global artificial intelligence industry, said the memory chip supply bottleneck will persist into early 2028, as its sweep of high-bandwidth memory (HBM) supply spills over into orders for standard server DRAM and NAND flash, deepening the shortage across the memory market.
On the 27th, Nvidia told an earnings conference call that the memory supply shortage would continue into early 2028. Colette Kress, Nvidia's chief financial officer, said the current memory shortage would act as a bottleneck to the company's growth at least through fiscal 2028, which runs from February 2027 to January 2028.
That effectively locks in a structural shortage for roughly the next 16 months. Industry watchers expect the supply crunch to run even beyond Nvidia's projection, lasting through the end of 2028, given the pace of AI infrastructure buildout.
Memory prices are already climbing beyond market expectations, the company said. Kress said Nvidia was seeing extreme pricing in memory and that prices would rise further next year.
One of the biggest drivers of the sustained price increases is aggressive advance buying by Nvidia, the industry leader. In a regulatory filing the same day, Nvidia said its purchase commitments to suppliers, including memory, more than doubled to $279 billion (about 385 trillion won) from $119 billion (about 164 trillion won) in the previous quarter.
With Nvidia locking up HBM volumes with astronomical sums, other big tech firms and cloud service providers now have to offer higher prices and unusually generous terms to secure chips. As the three major memory makers — Samsung Electronics (005930.KS), SK hynix (000660.KS) and Micron — concentrate production lines on higher-margin HBM, shortages are cascading across the memory market, from standard server DRAM to enterprise solid-state drives (eSSD).
Rising profitability for the latest AI accelerators suggests HBM prices will keep climbing for some time. Nvidia Chief Executive Jensen Huang said that measured by revenue generation per gigawatt of data center capacity, the company's latest AI accelerator, Vera Rubin, would reach $40 billion, up from $18 billion for Hopper and $25 billion for Blackwell. With each AI server generating enormous added value, big tech buyers and other customers have little choice but to accept further price increases for HBM, a core component.
In an unprecedented seller's market, earnings forecasts for South Korea's two memory giants are being revised upward almost daily. According to financial data provider FnGuide, the consensus estimate for Samsung Electronics' operating profit this year is 394.4153 trillion won, up 804.6% from a year earlier, on revenue of 741.1692 trillion won, up 122.2%. In 2028, when the memory bottleneck is expected to peak, the company is projected to post unprecedented results of 1,023.3649 trillion won in revenue and 563.3759 trillion won in operating profit.
SK hynix is likewise projected to post 2026 operating profit of 266.264 trillion won, up 464.0% from a year earlier, on revenue of 345.1216 trillion won, up 255.3%. Operating profit is then seen at 397.6897 trillion won in 2027 and 414.7592 trillion won in 2028. Both companies would set record results every year from this year through 2028.
