Highlights of an expert call from a Senior Directo

Highlights of an expert call from a Senior Director at ARM who was a Former Director at Western Digital regarding memory:

Unique Nature of the Current Memory Upcycle

> AI Data Center Expansion: Unlike past commodity-driven cycles (such as 2017–2018 or 2021–2022), this upcycle is uniquely propelled by sustained, multi-year AI data center buildouts from hyperscalers like Amazon, Microsoft, and Meta.

> Longer Cycle Duration: While past cycles typically flipped every 18 months, strong demand and constrained supply mean suppliers are already largely sold out through 2027, with market softening potentially expected around 2028 as new fabs come online.

Memory Bandwidth & Capacity Constraints (Training vs. Inference)

> Training Phase: Demand for HBM, LPDDR, and storage remains robust to support trillion-parameter models, though the baseline of model creation is becoming more established.

> Inference & The Agentic World: Inference is the primary revenue driver (exemplified by Micron's performance) and the "token economy". As multi-user agent systems scale, demand for faster token generation explodes, creating a "double whammy" of high demand for both GPU-attached HBM and CPU-orchestrated LPDDR bandwidth.

Supply Growth, Market Share, and Competitors

> Current Market Shares:

HBM Share: SK Hynix leads at approximately 58% to 60%, followed by Samsung and Micron at around 20% each.

DRAM Share: Samsung leads overall with roughly 40%, followed by SK Hynix (30%), Micron (20%), and others/China vendors (10%).

> Strategic Allocation: Vendors are balancing high-margin LPDDR capacity (driven by tight markets in smartphones and PCs) alongside HBM.

> The CXMT Threat: China's CXMT is expected to capture growing single-digit market share, primarily focusing on older nodes and LPDDR products for domestic data centers and smartphone brands (Xiaomi, vivo, OPPO).

Long-Term Supply Agreements (LTAs) & Pricing Outlook

> Shift in LTAs: Data center customers are pushing for 3-year or even 5-year LTAs to secure guaranteed supply, giving memory vendors stronger pricing power compared to historical 1-year agreements.

> Pricing Trajectory:

HBM/DRAM/NAND: Broad consensus expects prices to trend upward through 2026 and 2027 due to supply constraints.

HBM Pricing Trajectory: Blended average selling prices (ASPs) per stack are estimated around $330 to $360 for HBM3E in 2026, projected to rise to about $500 for HBM4 in 2027 as it gets introduced, and expected to soften slightly to around $450 for HBM4E in 2028 as supply catches up.

Company Optimism & Outlook

> SK Hynix: Highly praised for its market leadership in HBM, though integrating logic dies for future iterations poses a competitive frontier.

> Samsung: Viewed favorably due to an inherent margin and pricing advantage via its in-house logic foundry (Samsung Foundry), positioning it to potentially claw back some HBM share in the 2027–2028 timeframe.

> Micron: Expected to maintain its steady ~20% market share while benefiting strongly from current supply-constrained margins.

$Micron Tech(MU.US) $EWY $DRAM