DRAM Rally Persists as AI Memory Costs Soar

The boom in AI accelerators has turned high-bandwidth memory into the memory industry's priority product. Samsung Electronics, SK hynix and Micron are steering constrained wafer capacity toward HBM, which consumes more manufacturing capacity per bit than conventional DRAM. That shift leaves fewer chips for PCs, smartphones and general-purpose servers, making the price surge more than a cyclical rebound and raising the risk that expensive memory will weigh on device demand and data-center budgets.

On Aug. 26, South Korea's Trade Statistics Promotion Institute said the DRAM export unit price for Aug. 1-20 reached $92,183 per kilogram, up 401% from a year earlier and 12.5 times the January 2023 trough of about $7,400. One kilogram was worth the equivalent of 620 grams of gold. TrendForce expects HBM to absorb 30% of the top three suppliers' DRAM wafer input by end-2027, while Goldman Sachs forecasts the DRAM supply deficit will widen to 5.9% in 2027 from 5.0% in 2026.

The expansion of AI inference services and continued data center construction by major cloud providers are boosting demand for server DRAM. Meanwhile, memory manufacturers are shifting capacity toward high-bandwidth memory, further tightening supplies of conventional DRAM. Daiwa Securities expects the supply-demand imbalance and price rally to persist through 2027, shaping the earnings cycle for memory producers.

Daiwa Securities said in its latest report that Nanya Technology benefited from rising DRAM prices, with second-quarter 2026 revenue beating market expectations and the company returning to profit from a loss. This indicates that the industry recovery is already showing up in its financial results. Daiwa therefore sharply raised its price target for Nanya Technology to NT$650, citing expectations that AI inference demand will continue to support pricing and earnings.

The expansion of AI server capacity has driven a surge in demand for high-bandwidth memory. Suppliers have responded by prioritizing wafer and advanced packaging capacity for higher-margin products, squeezing supplies of conventional DRAM. Market research firm SigmaIntell says the imbalance has created an AI-driven memory supercycle affecting costs for servers, PCs and smartphones.

SigmaIntell’s latest assessment expects DRAM price momentum to reach a turning point in the second quarter of 2026, with the quarterly increase narrowing to about 30%. Growth is projected to cool further in the second half as capacity allocation gradually stabilizes and customer stockpiling subsides. AI server demand will continue to support the market in the third quarter of 2026, but mounting pressure in consumer markets and a high comparison base will further curb price gains.