As the focus of AI use shifts from training large language models (LLMs) to inference applications, memory prices continue to soar. Taiwan-based market research firm TrendForce forecast that, with this reshaping of demand, second-quarter DRAM contract prices this year will jump 58% to 63% from the previous quarter.
On the 2nd, according to TrendForce, as cloud service corporations (CSPs) shift their data center build-out from AI-dedicated servers to general-purpose servers, the spectrum of memory demand is widening significantly. Demand that had previously concentrated on HBM3E (5th generation), LPDDR5X, and high-capacity RDIMMs is now spreading evenly across RDIMM product lines of various capacities.
On the supply side, there is no slack. With DRAM suppliers' inventories already very low in the second quarter, production lines are being prioritized for high-capacity RDIMMs for AI servers. As a result, TrendForce projected it will be difficult to meet on time the demand from general customers such as PC OEMs and smartphone manufacturers.
Another driver of the price surge is a shift in CSPs' stance. Large CSPs that prioritize supply stability are taking a relatively flexible position on price increases, leading smaller customers to tilt toward accepting higher prices to secure supply volumes, according to the analysis.
As the top three—Samsung Electronics, SK hynix, and Micron—focus their capabilities on securing advanced process technologies, gaps are opening in the mature-process market. Taiwan's Nanya Technology, Winbond, and PSMC are filling this niche with mature-process products and expanding their foothold.
The latest price outlook shows that, beyond a simple supply-demand imbalance, the AI infrastructure shift is reshaping the entire memory ecosystem. TrendForce said, "The spread of inference workloads is lifting demand for general-purpose servers, and this, in turn, is clearly forming a chain effect that translates into upward price pressure across all DRAM product categories."