The global shift from AI model training to large-scale inference has set off a sweeping upgrade cycle across data center memory — pushing demand well beyond graphics processors into a full infrastructure stack encompassing high-bandwidth memory, server DDR5, low-power modules, and enterprise-grade solid-state drives.
Per-Server Memory Doubles as AI Inference Demands Rise
As AI models move from training into deployment, the computational architecture around them must evolve as well. HBM handles high-speed model data transfer close to the processor core; DDR5 and low-power formats such as SOCAMM serve CPU-side operations including data pre-processing and service scheduling; and enterprise SSDs absorb the storage layer — housing model weights, vector databases, retrieval datasets, and persistent key-value caches.
Micron has quantified the scale of this transformation, noting that per-server memory capacity has approximately doubled over the past three years. With inference workloads growing in complexity, overall system performance is now increasingly constrained by memory bandwidth and capacity rather than compute power alone — a structural shift that is reshaping how chipmakers compete and how much memory each server requires.
Samsung and Micron Push HBM4 Into Mass Production
Samsung Electronics announced in February 2026 that its fourth-generation high-bandwidth memory — HBM4 — had entered mass production and begun commercial shipment. Built on sixth-generation, 10-nanometer-class 1c DRAM with a 4-nanometer logic base die, the chips represent a significant leap in both process and architecture. By May 2026, Samsung had delivered 12-layer HBM4E samples to key customers.
Micron, meanwhile, is advancing a 36GB, 12-layer HBM4 module engineered specifically for Nvidia's Vera Rubin platform, targeting single-stack bandwidth of more than 2.8 terabytes per second. The company is also developing SOCAMM2 modules for low-power server memory alongside a PCIe Gen6 SSD lineup — including the 9650, which achieves sequential read speeds approaching double those of its PCIe Gen5 predecessors, and the 6600 ION, which stretches single-device capacity to 245TB.
The three-way competition among Samsung, SK Hynix, and Micron has grown far more multidimensional than earlier HBM generations. Beyond per-die costs, the battleground now spans advanced process nodes, die-stacking architecture, logic base die design, packaging, thermal management, and customer qualification cycles. Advanced DRAM wafers and extreme ultraviolet lithography equipment are shared resources across all three product lines — meaning HBM, high-density DDR5, and enterprise SSDs are effectively competing for the same constrained manufacturing capacity.
Micron reported revenue of $41.46 billion for Q3 2026, up from $23.86 billion in the prior quarter, and projected a non-GAAP gross margin of approximately 86% for the following quarter — reflecting the pricing power generated by AI-driven demand, supply constraints, and a product mix weighted heavily toward high-value memory.
Taiwan's Nanya Technology Records Outsized Gains on DRAM Pricing
The supply-demand imbalance has delivered significant benefits to Taiwan's own memory producers. Nanya Technology — Taiwan's largest dedicated DRAM manufacturer — reported second-quarter 2026 revenue of NT$82.55 billion, a 68.2% increase from the first quarter. The jump was driven primarily by pricing: the company's average DRAM selling price rose more than 60% quarter-over-quarter, even as bit shipments held roughly flat.
Those pricing gains fed directly into margins. Nanya's gross margin expanded from 67.9% in Q1 to 79.5% in Q2, illustrating a defining feature of the memory industry — high fixed costs mean that when average selling prices rise without a corresponding increase in costs, profit improvements can be substantial.
Winbond Electronics, another Taiwan-listed memory maker, is similarly positioned to capture niche memory opportunities as supply constraints ripple through the market, according to Pioneer Investors Weekly.
Both companies' gains point to a structural pattern: when the largest global producers direct their most advanced wafer capacity toward premium products like HBM and high-density server DDR5, supply of standard-grade DRAM can tighten independently of overall demand — generating pricing leverage for specialized producers further down the value chain.
*Adapted from Wealth Invest Weekly, Issue 2414. By Lu Tai-de. （Related： Beijing Watch | Kimi K3 Built on Stolen Anthropic Tech, White House Charges ｜ Latest ）
