HBM Supply Struggle: SK Hynix, Samsung, Micron Expand

NVIDIA, a leader in AI chipsets, saw its shares rise 2.7% on Wednesday (11th), closing at $134.91 per share. This marks the fourth consecutive trading day of gains, solidifying its market cap at $3.23 trillion. The upward momentum is driven by pivotal technological advancements, notably CoWoS advanced packaging led by TSMC (2330-TW) (TSM-US) and the widespread adoption of high-bandwidth memory (HBM).

NVIDIA's cutting-edge H200 chip, the first to utilize the HBM3E memory standard, underscores its leadership in AI acceleration technologies. With the burgeoning AI sector, memory giants like Samsung, SK Hynix, and Micron (MU-US) are increasingly prioritizing HBM production.

The popularity of HBM has created significant market dynamics, stabilizing the memory chip industry, potentially causing shortages and price increases in general DRAM, and intensifying technological competition.

Categorized into DDR, LPDDR, GDDR, and HBM, these memory types serve diverse sectors: DDR for consumer electronics, servers, and PCs; LPDDR for mobile devices, smartphones, and automotive applications; and GDDR primarily for GPU-based image processing.

HBM, introduced in 2014 with HBM1 offering 128GB/s bandwidth and 4GB memory, outperformed contemporaneous GDDR5. HBM2, launched in 2018, featured 256GB/s bandwidth and 8GB memory, with up to 8 layers of DRAM die stacking. HBM3, introduced recently, supports up to 6.4Gbps transfer speeds and a peak rate of 819GB/s with 16GB memory. SK Hynix's HBM3E enhancement, starting mass production in 2024, offers up to 8Gbps transfer speeds and 24GB capacity.

Demand for HBM has outpaced supply capacity this year for SK Hynix, Samsung, and Micron, with next year's capacity largely reserved. In response, these companies are aggressively expanding production capabilities. SK Hynix plans to increase 1b DRAM monthly production from 10,000 wafers in Q1 to 90,000 by year-end, reaching 140,000-150,000 wafers in H1 next year—a substantial increase from Q1.

Samsung anticipates increasing HBM production capacity 2.9 times by the end of March compared to last year, while Micron is constructing advanced HBM testing and production lines in the US and plans to commence production in Malaysia, aiming to meet rising AI-driven demand.

In this competitive landscape, SK Hynix leads with its superior HBM3 performance, securing major NVIDIA orders, while Samsung focuses on cloud-based customer orders and Micron prioritizes HBM3E products.

Investments are significant, with SK Hynix planning a $74.8 billion investment by 2028, allocating 80% to HBM research and production and advancing the production timeline for next-generation HBM4 chips to next year.

According to industry analysis, the dynamic gap in HBM demand over the next two years is estimated at 5.5% and 3.5% of production capacity. Dolphin Investment Research suggests that HBM, from undersupply at the end of last year, is expected to shift to oversupply by the end of this year.