Who Makes HBM

Who makes HBM is the roles map for high-bandwidth memory: which firms build the DRAM stacks, which partners attach those stacks beside an accelerator on advanced packaging, and which tools and materials gate how fast supply can rise.

EXPLAINEDPlayersLast reviewed: 04 Oct 2026Next review: after next Micron / Samsung / SK hynix HBM capacity or allocation print (ask, SCA lock, or volume note)

For investors on the semiconductors layer, the useful question is which role gets paid when AI CapEx binds scarce stacks and packaging — not which ticker “wins.” Early October 2026 clocks show allocation locks and opening asks — ready book, not settled cash until stacks ship and invoices clear. See what HBM is; for the attach gate, see what CoWoS is.

Education only. Not investment advice. Do your own research.

Who makes HBM?

Who makes HBM splits into three roles. DRAM stack makers fabricate and stack the memory cube. Foundry and advanced-packaging partners attach that cube beside the accelerator. Tool and materials suppliers gate how fast cubes can ship. JEDEC’s HBM4 standard (JESD270-4, 16 April 2025) sets the interface envelope — up to about 2 TB/s on a 2048-bit bus — but does not name who clears which role.

Who builds the DRAM stacks?

Stack makers are the memory suppliers that ship HBM cubes — principally SK hynix, Samsung, and Micron. SK hynix began 12-layer HBM3E volume production on 26 September 2024; Samsung announced HBM4 commercial shipments on 12 February 2026 (both company-reported). Those dates show the role, not a ranking. Contrast board-level memory in HBM vs GDDR or HBM vs DDR5 — same DRAM family tree, different attach and who clears first when AI orders hit.

The cash clock shifted in early October 2026. Micron’s fiscal Q4 release (30 Sep 2026) and Korea Herald (1 Oct 2026) carried the outlook: more than 75% of fiscal 2027 output already committed, talks shifting toward 2028, and most calendar 2027 HBM bits sold at higher YoY prices — company outlook, not collected cash. Committed volume still has to ramp, ship, and clear invoice. See the Micron >75% 2027 output Wire.

Separately, Maeil Business (2 Oct 2026) reported Samsung opened 2027 HBM4 talks at mid-to-high about $4/Gb — more than three times a ~$1.50/Gb HBM3E reference. That is an opening ask, not a settled ASP. Cash clears when qualified volume locks and the customer pays. See the Samsung mid-$4/Gb HBM4 Wire.

Who attaches HBM to the accelerator?

HBM only becomes useful when it sits beside the GPU or ASIC on an advanced package. TSMC’s CoWoS family places SoCs and HBM stacks on a shared package; CoWoS-L has been in volume since 2024 (TSMC CoWoS page). That is attach — not DRAM fabrication. GPU designers who specify stacks still wait on cubes and package slots — see who makes AI chips.

SK hynix’s 28 September 2026 newsroom note says it won Partner of the Year for validating HBM5 with TSMC CoWoS — early validation, not volume. When package slots are short, cash sticks at attach; stack makers can wait even when cubes exist. See CoWoS explained.

Who supplies the tools that gate stacks?

Building taller HBM cubes needs stacking equipment. Thermal-compression (TC) bonders stack DRAM dies — a step memory makers buy when they expand HBM capacity. Yonhap (8 Jun 2026) reported SK hynix ordered Hanmi TC bonders in a ₩44.2 billion filing for HBM4 — reported, not guided. See memory process tools for HBM.

Who gets paid — and where is the stuck step?

Who gets paid tracks which role clears first when AI CapEx orders arrive. Stack makers clear when scarce cubes ship under settled contracts; packaging clears when CoWoS-class slots bind; toolmakers clear when stacking capacity expands. The stuck step is usually stacks plus attach — and allocation locks or asks can rise while installs and invoices still wait.

Korea markets door where memory allocation bites: Korea HBM — who gets paid when AI memory stays scarce. Yield / wafer-shift door: Samsung HBM4 yield. Analysis on where the bind moves next: If packaging catches up, where does the stuck step move?.

Common misconceptions

• People assume “who makes HBM” means one company. Actually three roles share the chain: stack makers, package attach, and tools/materials. Missing any one role leaves the accelerator waiting.
• People assume TSMC makes the HBM cube. Actually CoWoS attaches memory makers’ stacks beside logic. The foundry/packaging partner does not fabricate the DRAM stack.
• People treat an allocation lock or opening ask as collected cash. Actually committed volume and mid-$4/Gb asks are ready book. Cash clears on ship and invoice — and on settled contract ASP, not the first number in the room.
• People treat a Partner of the Year award as volume. Actually SK hynix’s HBM5–CoWoS recognition is early validation with the packaging partner — not a shipment print.

FAQ

Related terms

• HBM — what high-bandwidth memory is.
• CoWoS — the packaging attach that sits next to HBM.
• Memory process tools — equipment that gates how fast HBM capacity can rise.
• HBM vs GDDR — stack-on-package versus board-level graphics DRAM.
• Who makes AI chips — designers who specify stacks next to the attach gate.

Markets and stack doors

Layer: semiconductors. Markets: Korea HBM memory · Samsung HBM4 yield.

Further reading

Anger prints: Micron >75% 2027 output Wire · Samsung mid-$4/Gb HBM4 Wire · where the stuck step moves if packaging catches up.

Get the Wire — short notes when memory or packaging clocks move.

Sources