Memory subsystems

Why the concentration exists

High Bandwidth Memory stacks eight to twelve individual DRAM dies using through-silicon vias, then bonds them to an interposer alongside the GPU die. This vertical integration enables bandwidth far exceeding conventional DRAM. An NVIDIA H100 SXM5 paired with six HBM3 stacks achieves approximately 3.35 TB/s of memory bandwidth, roughly 65 times the bandwidth of a dual-channel DDR5 desktop system.[20][9]

Manufacturing complexity creates natural supply constraints. A single HBM3E DRAM chiplet can be roughly twice the size of an equivalent-capacity DDR5 chip, which reduces yields per wafer. SK Hynix's HBM3E has increased the number of through-silicon vias by almost six times compared to earlier generations, achieving up to 75 percent lower IR drop for VPP.[6][8]

Demand growth systematically outpaces supply expansion. HBM demand is growing at roughly 80 to 100 percent per year driven by AI accelerator deployments, while supply is growing at 50 to 60 percent per year. The total committed investment across the three HBM suppliers exceeds $50 billion, yet analysts report potential multi-year shortages in HBM supply.[20][19]

What the evidence shows

Who supplies it

Three companies control approximately 95 percent of global DRAM production. Samsung holds about a third of global DRAM capacity and led the market with a 38 percent share in Q1 2026. SK hynix follows with 29 percent share, and together with Samsung controls roughly two-thirds of the global DRAM market.[18][17][1]

SK hynix dominates the HBM segment with 62 percent market share as of Q2 2025, followed by Micron at 21 percent and Samsung at 17 percent. SK hynix holds a near-exclusive position for NVIDIA H100, H200, and B200 GPU packages with over 50 percent of HBM3E output. Micron has overtaken Samsung for the number two HBM spot with 23 percent share.[17][18][9][6]

China's CXMT has emerged as a strategic challenger in DRAM segments adjacent to core HBM stacks. CXMT's FY25 gross margin reached 37.8 percent, and its operating margin reached 70 percent in Q1 2026. South Korean prosecutors have alleged that former Samsung employees shared sensitive trade information that contributed to CXMT's engineering progress.[2][11]

What it depends on, and what depends on it

Memory hierarchy now determines AI infrastructure competitiveness as much as compute performance. SK hynix positions HBM4, HBM4E, SOCAMM, LPDDR, and eSSDs as a connected AI memory portfolio, recognizing that infrastructure competition has moved beyond individual product performance toward the design of the entire memory hierarchy.[3]

Large language models create memory demands that exceed conventional server capacity. For Llama 3 70B loaded in FP16 without grouped-query attention, the KV Cache consumes 160GiB at 64K context length and 320GiB at 128K context length, exceeding the model weights' 140GiB. In the first half of 2026, surging KV Cache demand coupled with limited memory supply resulted in severe memory bottlenecks.[12]

Compute Express Link technology addresses capacity constraints by enabling memory expansion beyond traditional DIMM slots. Micron's CZ120 memory expansion modules offer 128GB and 256GB densities, enabling up to 2TB of added capacity at the server level with bandwidth increase of 38 GB/s. Companies positioned along CXL-based capacity expansion include Enfabrica, Penguin Solutions, Marvell, and Meta.[14][12]

Where it sits in the stack

What would break it

Supply constraints are intensifying rather than easing. SK hynix has stated that 2027 will be the 'worst year' for memory shortage and forecasts the crunch to last until 2030. DRAM prices rose by 172 percent in 2025 alone, prompting memory manufacturers to halt new orders for DDR modules.[13][15]

Capacity expansion faces long lead times despite aggressive investment. Samsung is looking to expand production capacity by around 50 percent in 2026, while SK hynix announced plans to increase infrastructure investment by more than four times the previously announced figure. SK hynix will double its memory wafer capacity within five years, as announced by chairman Chey Tae-won at Computex on June 2, 2026.[5][10]

China's domestic memory manufacturing investment cycle is accelerating orders for locally developed semiconductor equipment, with production capacity expanding across DRAM and NAND projects during 2026. This investment is strengthening China's domestic semiconductor equipment ecosystem while altering purchasing patterns for fabrication tools, potentially reducing reliance on overseas suppliers for selected process stages.[4]

What to watch

Micron's $1.8 billion acquisition of a fabrication facility from Powerchip Semiconductor Manufacturing Corp is expected to close in the second quarter of 2026. Mass production of advanced DRAM chips at the acquired site is expected by 2027, with the first phase estimated to contribute over 10 percent of Micron's total global capacity relative to Q4 2026 levels.[7]

Samsung may launch the first commercial DRAM products based on its sub-10nm process technology at the end of 2026 to the beginning of 2027. The technology uses a high-heat-resistant amorphous InGaO-based vertical channel transistor that can withstand temperatures of 550 degrees Celsius with a channel length of 100nm.[16]

Related nodes

Sources

1. counterpointresearch.com · 2026-06-09T00:00:00
2. sonoranelectronics.com · 2026-01-12T22:30:32
3. news.skhynix.com · 2026-07-03T00:00:14
4. astutegroup.com · 2026-07-06T07:00:00
5. datacenterdynamics.com · January 05, 2026
6. semiwiki.com
7. scmp.com · 21 Jan 2026
8. newsletter.semianalysis.com · 2027
9. semiconductorx.com
10. tomshardware.com · 2026-06-02
11. newsletter.semianalysis.com · 1Q26
12. insights.trendforce.com · Jul 24, 2026
13. tomshardware.com · Q1 2026
14. micron.com
15. en.wikipedia.org · October 1, 2025
16. semicone.com · 2025-12-18
17. fortune.com · Q2 2025
18. introl.com · Q2 2025
19. intuitionlabs.ai
20. gpunex.com