Global Memory Shortage Intensifies
Surging AI demand is fueling a global shortage of critical memory components like HBM and 3D NAND, described as a looming crisis. Social media users report HBM DRAM pricing has increased over 30% year-to-date, constraining AI deployments. In response, foundries like TSMC are expanding globally while Samsung is leveraging its integrated memory and foundry services to win new orders from big tech.
- SK hynix, which held a 55% share of the HBM market in 2025, is expected to supply 70% of the HBM4 for Nvidia's next-generation "Rubin" GPU platform. The company's entire HBM production for 2026 is reportedly sold out, reflecting the extreme demand. - While AI training is computationally expensive, ongoing inference costs for widely used models can account for 80-90% of the total lifetime cost of an AI system. This economic reality is driving hyperscalers to develop custom silicon, like Google's TPUs and Amazon's Trainium chips, which can offer 30-40% better power efficiency for these specific, repetitive workloads. - The shift to next-generation Gate-All-Around (GAA) transistor architecture is a critical factor for future chip performance and efficiency. Samsung was the first to mass-produce chips with GAA at its 3nm node, while competitors like TSMC and Intel are expected to integrate the technology in their 2nm and 20A processes, respectively. - The memory shortage is having a cascading effect, with the prioritization of HBM production leading to price hikes for other memory types. DRAM prices surged by as much as 258% in a six-month period, and SanDisk is reportedly doubling the price of its enterprise-grade 3D NAND in the first quarter of 2026. - Advanced packaging is a significant bottleneck, with demand for TSMC's CoWoS (Chip on Wafer on Substrate) technology persistently outstripping supply. This constraint directly impacts the availability of finished AI accelerators, as HBM stacks require complex packaging techniques with long lead times. - Micron has aggressively re-entered the HBM market by skipping the HBM3 generation to focus on HBM3E and is now shipping HBM4 samples. The company's market share is projected to grow to around 20% by the third quarter of fiscal year 2026, challenging the dominance of SK hynix and Samsung. - Demand for custom-built, ASIC-based AI chips from hyperscalers is forecasted to drive an 82% increase in HBM consumption for that segment, accounting for one-third of the total HBM market. This "build vs. buy" decision allows companies with sufficient scale to optimize for specific workloads, reducing long-term operational costs and dependency on third-party chip suppliers. - The transition to HBM4, expected to enter mass production in 2026, promises significant performance gains, with bandwidth exceeding 1.5 TB/s per stack. However, manufacturing 16-layer HBM4 stacks presents formidable technical challenges, requiring wafers to be thinned to approximately 30 micrometers.
