Who makes HBM for AI data centers
HBM (HBM3E, HBM4)
constrained
High-bandwidth memory stacked directly on the same package as an AI accelerator, giving the GPU or ASIC very fast, close-range access to the memory it needs to stay fed.
HBM sits on the accelerator package itself, next to the compute die, rather than on a separate module across the board — the physical packaging that puts it there belongs to Layer 1. It is the first thing a new accelerator generation is designed around, since the number of stacks and their capacity set a hard ceiling on how large a model an accelerator can hold in memory before it has to reach further out to DRAM. AI training and inference workloads scale primarily by adding parameters and context length, both of which are bound by per-accelerator memory capacity, so HBM content has risen with every accelerator generation. The taxonomy flags HBM as the most supply-constrained component in the stack, and also notes that HBM4 base die is fabricated at a foundry partner for some suppliers or produced in-house by others — a qualification detail worth tracking, since a supplier's base-die sourcing shapes its cost position and lead times. For an investor, the things to watch are supplier qualification status on each new accelerator design win, and how stacking-tool and base-die capacity trail the memory makers' own output.
AI delta: Larger models and longer context windows push accelerator designs toward more HBM stacks and higher per-stack capacity with each generation, keeping HBM the tightest link in the memory chain.
The most supply-constrained component in the stack.
Companies on this part (6)
| Company | Role | Ownership | Ticker | Price | 1d | Confidence | Note |
|---|---|---|---|---|---|---|---|
| BESI | manufacturer | public | BESI.AS | 145 EUR | -7.9% | inferred | |
| Disco | manufacturer | public | 6146.T | 49290 JPY | -3.2% | inferred | |
| Hanmi Semiconductor | manufacturer | public | 042700.KS | 191000 KRW | -0.8% | disclosed | stacking tools · source |
| Micron | manufacturer | public | MU | 878 USD | +3.8% | disclosed | · source |
| Samsung Electronics | manufacturer | public | 005930.KS | 215500 KRW | +0.2% | disclosed | · source |
| SK Hynix | manufacturer | public | 000660.KS | 1277000 KRW | +4.4% | disclosed | 62% of HBM revenue in Q2 2025, per Chosun Biz via Astute Group, 2025-09-24 -- a single-quarter figure, not a running average — leader · source |
Who competes here
The same table read the other way: every company above competes for HBM sockets. Confidence tags matter — an inferred placement is a lead, not a confirmed rivalry.
Seen in real machines
HBM is named in 9 of the rack teardowns and supercomputer breakdowns on this site. Each line below is a piece of that machine. The companies shown are the ones named on that line, not necessarily the maker of this part.
- Nvidia GB300 NVL72 · rack teardown
- HBM3E (288 GB per GPU) — SK Hynix, Micron, Samsung Electronics
- AMD Helios (MI455X) · rack teardown
- HBM4 (432 GB per GPU, 31 TB per rack) — Samsung Electronics
- Google TPU7x (Ironwood) · rack teardown
- HBM3E (192 GB per chip, 8 stacks, 7.37 TB/s bandwidth) — Samsung Electronics, SK Hynix
- Frontier · supercomputer
- JUPITER · supercomputer
- Isambard-AI · supercomputer
- GPU memory (96 GB HBM3 per Hopper GPU, 384 GB per node) — SK Hynix
- Eagle (Microsoft Azure) · supercomputer
- GPU memory (80 GB HBM3 per H100, 640 GB per VM) — SK Hynix
- Eos · supercomputer
- GPU memory (80 GB HBM3 per H100, five stacks; 640 GB per DGX H100) — SK Hynix
- Fugaku · supercomputer
- A64FX die fabrication and 2.5D packaging with four on-package HBM2 stacks (32 GiB, 1,024 GB/s per node) — TSMC
Nearby parts
- Next in layer: DRAM
- Layer: Layer 2 — Memory & storage
From the blog
- October's shortage headlines were about memory and chip tools, not GPUs 2026-10-19
- Half the supplier bench, a sixth of the coverage 2026-09-07
- The five AI data center parts this catalog flags as tight or constrained 2026-08-30
- Introducing the catalog: every part of an AI data center, and who makes it 2026-08-28
Watch
Independent videos — no affiliation; nothing loads from YouTube until you click.
Latest feed
- [AI Infrastructure Insight] How AI infrastructure will be designed for the future — newsroom:sk-hynix, 2026-10-22Effective real-world AI application requires improvements across faster compute, wider memory bandwidth, higher-performance networking, better storage, and more stable power infrastructure.
- Why Chip Design AI Benchmarks Must Go Beyond Functional Correctness — EE Times, 2026-10-21The increasing need for specialized architectures across various domains like physical AI and edge devices is driving a demand for custom silicon optimized for specific performance, memory bandwidth, and energy targets.
- Samsung shifts M2800 modem to 2nm, frees 4nm capacity for HBM base dies — DigiTimes, 2026-10-21Samsung is moving its next-generation modem to a 2nm process, thereby freeing up 4nm capacity to support the production of HBM base dies.
- Breaking AI's memory wall: former TSMC VP charts a path with HBM, optical links and 3D stacking — DigiTimes, 2026-10-21As AI performance gains are challenged by memory bandwidth limitations, the industry is advancing HBM, advanced packaging, and 3D integration techniques to connect memory and compute more densely.
- [AI Ecosystem] The semiconductor paradigm shifts toward memory — newsroom:sk-hynix, 2026-10-21The industry trend is shifting towards a greater reliance on memory components, suggesting changes in accelerator and memory demand.
- [AI Ecosystem] The semiconductor paradigm shifts toward memory — newsroom:sk-hynix, 2026-10-21The semiconductor industry is shifting its focus towards memory solutions, indicating a strong trend in memory technology.
- AI memory demand lifts Samsung, SK Hynix 2027 outlook — DigiTimes, 2026-10-19Increased demand from AI workloads is boosting the 2027 outlook for memory manufacturers Samsung and SK Hynix.
- Nvidia's custom HBM pulls Samsung onto TSMC's 3nm process, report says — DigiTimes, 2026-10-19Nvidia's use of custom High Bandwidth Memory (HBM) is influencing Samsung to utilize TSMC's 3nm process.
- Interview: SEMI CEO sees US$3 trillion chip industry by 2035, as trusted partners reshape supply chain — DigiTimes, 2026-10-16SEMI's CEO predicts the semiconductor industry will achieve $2 trillion in revenue by 2035, driven by AI and quantum computing growth.
- South Korea's September ICT Exports Surge Past US$75B on AI Demand — DigiTimes, 2026-10-16South Korea's ICT exports reached over US$75 billion in September 2026, heavily supported by semiconductor exports that surpassed $60 billion due to global AI infrastructure spending.
Questions about HBM
- What is HBM?
- High-bandwidth memory stacked directly on the same package as an AI accelerator, giving the GPU or ASIC very fast, close-range access to the memory it needs to stay fed.
- Which companies make HBM for AI data centers?
- This catalog lists 6 companies on HBM, 6 of them with listed equity: BESI, Disco, Hanmi Semiconductor, Micron, Samsung Electronics, SK Hynix.
- How does AI change demand for HBM?
- Larger models and longer context windows push accelerator designs toward more HBM stacks and higher per-stack capacity with each generation, keeping HBM the tightest link in the memory chain.
- HBM — is it a supply bottleneck?
- This catalog flags HBM as "constrained". The most supply-constrained component in the stack. The bottlenecks page lists every part carrying a sourced supply constraint.
- How reliable are these HBM supplier placements?
- The 6 placements on this part break down as 4 disclosed, 2 inferred. "disclosed" means the company or a filing says so; "reported" comes from trade press or a reference-design list; "inferred" is a taxonomy lead with no citation yet.
Sources: placements without a source URL come from the taxonomy seed and are tagged accordingly. How this data is built.· Markdown · JSON