Samsung Electronics recently announced a pivotal shift in its strategy, dedicating a large part of its Pyeongtaek P4 factory to producing next-generation HBM (High Bandwidth Memory).
This decision has significant implications for the entire memory market. First, it formalizes the industry's 'HBM-first' approach. By prioritizing HBM, which is essential for AI accelerators, Samsung is structurally reducing the number of silicon wafers available for conventional DRAM products like DDR5 (used in PCs and servers) and LPDDR5 (used in mobile devices). This creates a deliberate supply constraint, setting the stage for sustained price increases. Major cloud service providers have already been securing their supply through long-term agreements (LTAs), anticipating this very scenario.
On top of this supply shift, a powerful new demand driver has emerged. NVIDIA recently highlighted its new CPU, 'Vera', which utilizes a new type of memory module called SoCAMM2. This module uses LPDDR5 memory, which was traditionally reserved for mobile phones. Now, it's moving into data centers. A single Vera CPU can support up to a massive 1.5 terabytes of this memory, creating a huge new source of demand for LPDDR5 that directly competes with the mobile and PC markets.
These two forces—a strategic supply cut in conventional DRAM and a surge in new server-side demand—are creating a perfect storm. The competition for memory isn't just between HBM and regular DRAM anymore; it's now also between servers and mobile devices for the same LPDDR5 supply. This dynamic strengthens the likelihood of a widespread and prolonged memory shortage across the board.
This development isn't a surprise but rather the culmination of signals we've seen over the past several months. From Micron announcing its 2026 HBM supply was "sold out" late last year to market analysts repeatedly raising their price forecasts, the writing has been on the wall. Samsung's P4 announcement is the final, definitive piece of the puzzle, confirming that the AI-driven memory supercycle is not only real but is also fundamentally reshaping the supply landscape for all types of memory.
Glossary
- HBM (High Bandwidth Memory): A high-performance type of DRAM used in high-end GPUs and AI accelerators. It stacks memory chips vertically to achieve much faster data transfer speeds than conventional memory.
- DRAM (Dynamic Random-Access Memory): The most common type of computer memory used in everything from PCs and servers to smartphones. It's the "working memory" a system uses to run applications.
- SoCAMM2 (Small Outline Compression Attached Memory Module 2): A new, more compact, and power-efficient memory module format. It's gaining traction in servers for its ability to use low-power LPDDR memory, which was previously limited to mobile devices.