HBM4 Market Booms! Samsung, Micron, and SK Hynix Battle for AI Supremacy, Beneficiary List of 'Two Memory Firms' Revealed

AI servers are driving up demand for HBM and high-capacity DRAM. Although international memory manufacturers are accelerating capacity expansion, production still lags behind demand. Nanya Technology benefits from a significant rise in DRAM pricing, while Winbond secures business through niche-market opportunities.

By Lü Taider

Large language models are shifting from centralized training to large-scale inference. Data center expansions now involve not just GPUs or other AI accelerators, but an entire memory and storage architecture encompassing HBM, server DDR5, low-power memory, and enterprise SSDs. HBM, positioned close to the computing core, handles high-speed transmission of model data. Low-power memory such as DDR5 and SOCA MM supports CPU-side system operations and data preprocessing, while enterprise SSDs meet the high-capacity demands of model weights, datasets, vector databases, and persistent KV caches.

Micron recently pointed out that the memory capacity per server has doubled over the past three years. As AI inference complexity increases, system performance is increasingly constrained by memory bandwidth and capacity, no longer solely determined by computing chip performance.

HBM4 Mass Production Accelerates, Widening the Supply-Demand Gap for AI

Samsung Electronics announced in February 2026 the mass production and commercial shipment of HBM4, utilizing sixth-generation 10nm-class 1c DRAM and 4nm logic base die. In May, it provided 12-layer HBM4E samples to key customers. Micron, meanwhile, is advancing its 36GB, 12-layer HBM4 designed for the NVIDIA Vera Rubin platform, achieving a per-stack bandwidth exceeding 2.8TB/s, while simultaneously developing SOCA MM2 and PCIe Gen6 data center SSDs.

The competition among Samsung, SK Hynix, and Micron in the HBM4 era has evolved from a focus on per-DRAM cost to a comprehensive race involving advanced processes, stacking technology, logic base die, packaging, power consumption, thermal management, and customer certification capabilities.

More importantly, HBM, high-capacity server DDR5, and low-power server memory all compete for advanced DRAM wafers, EUV equipment, R&D personnel, testing, and back-end packaging resources. When memory giants prioritize allocating advanced DRAM capacity to HBM and high-capacity server products, not only does high-end memory supply remain tight, but it may also limit the supply growth of certain standard or mature-spec DRAM products.

Although enterprise SSDs primarily use NAND Flash and do not directly consume the same DRAM wafers, they still compete for capital expenditure, advanced controllers, R&D resources, and data center customer access. Micron reported third-quarter revenue of $41.46 billion in 2026, up from $23.86 billion in the previous quarter, and forecasts a non-GAAP gross margin of approximately 86% for the next quarter—reflecting how AI demand, supply constraints, and high-value product portfolios are significantly boosting memory profitability.

High-Capacity DDR5 and Enterprise SSDs Drive Full-Scale AI Data Center Upgrades

AI servers still rely on CPUs for operating systems, data preprocessing, network and storage management, service orchestration, and partial inference workflows, necessitating increased system DRAM capacity. Micron has provided server ecosystem partners with 256GB DDR5 RDIMM samples, reaching speeds up to 9200MT/s, using 3DS stacking and TSV technology to integrate multiple DRAM dies. A single 256GB module reduces operational power consumption by over 40% compared to using two 128GB modules. This indicates that even without a doubling of server shipments, the memory capacity, bandwidth, and unit value per device continue to rise—meaning memory bit demand is no longer solely determined by end-device volume.

On the other hand, large models require reading massive datasets during training and must store model weights, vector databases, retrieval data, and persistent KV caches during inference. Data centers cannot permanently store all data in the more expensive HBM or DRAM, making enterprise SSDs essential for high-capacity storage. Micron’s 9650 is a commercial PCIe Gen6 data center SSD, with sequential read performance nearly doubling that of PCIe Gen5 products. Another model, the 6600 ION, increases SSD capacity to 245TB. These products enhance performance and capacity respectively, showing that AI data pipeline upgrades are extending from GPUs and memory to enterprise storage devices.

DRAM Prices Surge, Driving Sharp Profit Growth for Nanya Technology

The memory industry is characterized by high fixed costs and clear price cycles. When capacity utilization is already high and bit shipments change little, rising average selling prices increase gross profit and earnings. Nanya Technology’s Q2 2026 revenue reached NT$82.549 billion, a 68.2% increase from Q1. DRAM average selling prices rose over 60% quarter-on-quarter, while bit shipments remained roughly flat. Driven by price increases and improved product mix, gross margin improved from 67.9% in Q1 to 79.5% in Q2. In other words, Nanya’s sharp profit growth in Q2 was primarily driven by supply constraints, price hikes, and a higher proportion of high-value products, resulting in significantly higher revenue and profits.

Article authorized by "Sinotrade Investment Weekly, Issue 2414"