AI-Fueled HBM Boom Sends DRAM Prices Soaring

Memory costs had broadly declined for nearly two decades as chipmakers improved manufacturing processes and expanded capacity. That trend is now being disrupted by the rapid buildout of generative AI infrastructure. Graphics processing units require large volumes of high-bandwidth memory, or HBM, a more complex and capacity-intensive form of DRAM, prompting suppliers to redirect production and tightening availability across the broader market.

Global DRAM prices have risen sharply within a matter of months as demand for HBM used in AI servers and GPUs outpaces supply. The increase has pushed memory costs back to levels last seen around 2007, reversing an almost 20-year hardware deflation trend. The impact is spreading beyond AI accelerators and servers, adding cost pressure for makers of smartphones, automotive electronics and other memory-dependent products.

The rapid expansion of generative AI servers is driving demand for high-bandwidth memory, or HBM. Samsung and other major memory-chip makers have reportedly shifted more than 90% of their capacity toward HBM, squeezing supplies of conventional DRAM. Because DRAM is widely used in smartphones and PCs, higher prices could further increase the cost and retail price of finished devices.

Samsung is reportedly planning to raise DRAM prices by as much as 20% in the third quarter. If implemented, it would be the company's third price increase within a year. UBS has also raised its pricing forecasts, predicting that DRAM prices will rise 32% quarter on quarter in the third quarter and NAND Flash prices will gain 30%. Smartphone and PC brands face cost pressure, but the impact on consumers' purchasing decisions is expected to be limited.

The rapid expansion of generative AI and high-performance computing is driving demand for high-bandwidth memory, or HBM, in AI servers. HBM shares some production capacity and supply-chain resources with conventional DRAM, creating a crowding-out effect that has also spread to NAND flash memory. The memory market is consequently shifting from a short-term cycle to a structural supply-demand imbalance.

Lenovo recently warned at an international supercomputing conference that high memory prices could persist beyond 2030, with costs gradually passed on to consumer electronics such as PCs and smartphones. Lenovo did not disclose specific price increases, amounts or effective dates. It said AI servers would continue to absorb HBM capacity, ultimately leaving consumers to pay higher prices.

Rapid growth in AI infrastructure is driving demand for high-bandwidth memory and advanced DRAM. Samsung Electronics, SK hynix and Micron have consequently shifted production resources toward advanced processes, squeezing supplies from mature nodes. The consumer DRAM shortage has now spread to DDR2, affecting costs and shipment plans for low-end consumer electronics and industrial control equipment.

To maintain shipments amid tight supplies, brand manufacturers and original design manufacturers have begun downgrading specifications on some products and switching to older-generation memory. That move has further increased DDR2 demand. The market expects third-quarter DDR2 contract prices to rise another 35% to 40% from the second quarter. With little new capacity for mature processes likely in the near term, the increase could continue into subsequent quarters.

The AI boom is fueling demand for high-bandwidth memory, but constrained capacity and tight supply are also pushing up conventional DRAM prices. Bernstein said revenue per wafer for conventional DRAM has surpassed that of HBM, suggesting memory-chip makers' profit momentum is broadening beyond a single high-end product. Taiwan's memory supply chain is also expected to benefit.

Bernstein's latest forecast sees the DRAM rally and the industry's windfall-profit cycle continuing through 2027. Earnings per share at the three major memory-chip makers could also peak in 2027 as HBM prices rise before reversing in 2028. Higher prices will boost supplier revenue but also increase hardware costs for building AI servers and training large models.

Generative AI has expanded rapidly since ChatGPT launched in December 2022, fueling demand for high-bandwidth memory and expectations of supply shortages. That has positioned companies including Samsung Electronics and SK hynix as major beneficiaries of AI infrastructure spending. But memory chips trade much like commodities, leaving prices and profits vulnerable to sharp reversals when supply and demand fall out of balance and raising questions about whether the stock rally can last.

As of May 25, 2026, Samsung Electronics and SK hynix had risen 114% and 186%, respectively, since the start of the year, while Micron had gained 141%. BlueBox Asset Management warned of a boom-bust cycle, while Standard Chartered advised taking some profits. New plants are expected to begin mass production from late 2027 to 2028. Nomura, meanwhile, set 12-month price targets of 4 million won for SK hynix and 590,000 won for Samsung.

The boom in generative AI and data-center construction is driving demand for high-bandwidth memory, or HBM. Samsung Electronics, SK hynix and Micron are shifting resources toward higher-margin products, squeezing supplies of conventional DRAM and NAND. HBM is a critical component in AI accelerators, and shortages could affect server delivery times, electronics costs and the semiconductor industry cycle.

Micron recently reported better-than-expected results, while the market estimates SK hynix’s revenue could reach $54.8 billion. Samsung and SK hynix have said their main memory capacity is fully booked through 2027. The industry estimates that even with manufacturers accelerating HBM capacity expansion, the supply shortfall will still be about 40% at the end of 2027. New capacity may not begin gradually easing the pressure until after 2027 at the earliest.