Semiconductors Videos

AI’s Appetite for HBM

The briefing spotlights how artificial‑intelligence workloads are turning high‑bandwidth memory (HBM) into a critical bottleneck for semiconductor manufacturers. HBM, once a niche component, now underpins the most powerful AI accelerators and is being ordered in volumes that dwarf traditional DRAM markets. Suppliers such as SK Hynix have already sold out their 2025 HBM allocations and repeated the feat for 2026, a rarity in the memory sector. An HBM3 stack comprises 13 vertically stacked dies, delivering roughly 1,500 mm² of silicon in a compact package, and each accelerator consumes eight of these stacks. When multiplied across servers, racks, and entire data‑center clusters, the demand translates into tens of thousands of memory dies per AI deployment, pushing DRAM prices upward. The report cites hyperscalers deploying tens of thousands of GPUs simultaneously, reshaping the memory market landscape. While SK Hynix, Samsung and Micron race to expand capacity, China’s CXMT is developing a domestic HBM alternative. Meanwhile, HBM4 pushes stack heights from 12 to 16 layers, intensifying thermal, packaging and interconnect constraints, prompting vendors to adopt advanced bonding techniques and custom controller dies. These dynamics signal that HBM will become the most demanding product line in semiconductor fabrication, with AI workloads testing physical limits. Over‑expansion could trigger a future oversupply, but the immediate pressure is on manufacturers to innovate packaging and thermal solutions, a shift that will affect AI hardware costs and the broader tech supply chain.