AI Is Reshaping the Memory Market—and the Squeeze May Last for Years

AI infrastructure is changing the memory market from a familiar boom-and-bust cycle into a supply-allocation problem. Hyperscalers building AI clusters need high-bandwidth memory (HBM) alongside accelerators, high-density DRAM in servers, and substantial NAND storage. Manufacturers are prioritizing those higher-value products while demand from PCs, phones and other electronics continues, leaving less slack across conventional memory markets. 28

Why AI demand affects both DRAM and NAND

HBM is the most visible constraint. It is specialized DRAM used close to advanced AI processors, and its production depends on scarce advanced capacity and packaging. IDC identifies AI infrastructure demand—especially HBM and high-density DRAM in GPU servers—as the primary driver of the present imbalance, with server demand growing faster than supply can respond. 8

The effect is broader than HBM. AI systems also require large, fast storage pools, supporting enterprise NAND demand. As memory makers devote production and investment to AI-oriented products, conventional DRAM and NAND buyers must compete for capacity or accept higher prices. IDC says this reallocation toward higher-margin AI memory has contributed to shortages affecting consumer electronics. 48

Why the shortage is expected to persist through 2027

Adding memory output is not instantaneous. New capacity must be built, equipped, qualified and ramped to acceptable yields; advanced HBM also requires specialized packaging. Synopsys CEO Sassine Ghazi said the crunch would continue through 2026 and 2027, citing the time needed to bring new capacity online. 12

Micron has likewise said AI is reshaping demand faster than the industry can add supply and expects tight conditions well into 2027 and beyond. 7 IDC’s supply outlook illustrates the constraint: it projected 2026 DRAM supply growth of 16% year over year and NAND supply growth of 17%, both below historical norms. 14

That does not mean every memory product will face identical shortages. Contract terms, product generation, supplier allocation and end-market demand can produce very different conditions in HBM, server DRAM, client DRAM and NAND. But the common theme is that supply additions are arriving more slowly than AI-led demand has grown.

Why prices are flowing into PCs and smartphones

Memory is a material part of a device’s bill of materials. IDC says smartphones and PCs are contending with higher component costs that are changing device economics. 8 Manufacturers may respond by absorbing the costs in margins, raising prices, adjusting memory or storage configurations, or changing product plans.

IDC’s 2026 scenarios estimated smartphone average selling prices could rise 3% to 5% in a moderate case and 6% to 8% in a pessimistic case. These are scenarios, not a guarantee that every handset will increase by that amount. 14

For downstream component suppliers, the main exposure is indirect: a higher-cost PC, phone or peripheral market can alter customers’ shipment volumes, product mix and inventory needs. The material provided does not establish a specific financial effect, forecast or guidance for Sonix Technology, so a company-specific impact cannot be quantified from this evidence.

What Phison’s NAND warning means

Phison CEO Khein-Seng Pua has warned that NAND flash could see its worst shortage in 2027. 17 Separately, reporting on his comments says he was told it could take about four years for new NAND capacity to catch up with demand—an outlook that points to 2030 or later if it proves accurate. 19

This is best understood as a risk outlook, not a settled market forecast. The more widely supported near-term conclusion is that tight memory supply and elevated pricing can persist through 2027. A 2030 outcome depends on how quickly capacity ramps, whether yields improve, how much production remains committed to AI memory, and whether AI infrastructure demand continues at its current pace. 71219

When new capacity could help

SK hynix’s more than $4 billion Indiana project shows both the scale of the response and the long lead time involved. The company plans to complete the cleanroom by October 2028 and begin volume production of HBM4E in the third quarter of 2029. CEO Kwak Noh-jung said he expects the memory shortage to persist through the end of 2030.

That facility should add meaningful U.S. HBM packaging capability from late 2029 onward. But it is not an immediate cure for all memory categories: a packaging plant aimed at next-generation HBM does not by itself create enough additional wafer supply or resolve constraints across conventional DRAM and NAND.

The practical takeaway

The memory squeeze is no longer only a semiconductor-market story. It is becoming a product-planning issue for device makers and a procurement issue for the companies that supply them. AI data centers are consuming memory capacity that would otherwise serve a wider set of electronics markets, while replacement capacity takes years to arrive.

Expect the clearest pressure to remain in AI-focused memory and servers, with spillover into consumer DRAM and NAND prices through at least 2027. Relief may begin unevenly as new capacity ramps, but claims of a full market reset by a specific year should be treated cautiously—especially when the additional capacity is targeted at HBM rather than the entire memory supply chain.