The HBM4 Base Die
▶ See it in 3D The Memory Stack region
New in HBM4: the bottom die of every stack is now a logic chip made in a foundry. SK Hynix uses TSMC; Samsung makes its own. It decided who won Rubin's allocation — and gave TSMC a toll on memory.
Each HBM stack sits on a "base" chip that links it to the GPU. With HBM4 that base chip is made on a logic process instead of a memory one: SK Hynix has TSMC make it, Samsung makes its own. It sounds like a detail. It is the moment memory companies became dependent on the logic-chip world, and TSMC got a slice of memory profits.
What it is
An HBM stack is 8–16 DRAM dies sitting on a base die that talks to the GPU. Through HBM3E that base die was made in the memory fab. HBM4 moves it to a logic process — a foundry-made die with buffers, test logic and soon compute — because the interface got too fast and too wide for DRAM-process transistors. From here, an HBM stack is part memory, part chip.
The business
Two camps, one winner-takes-more. SK Hynix + TSMC ("One-Team"): the best logic die on earth under the best HBM on earth — the combination NVIDIA qualified first and gave ~60–70% of Rubin HBM4 to. Samsung goes it alone with an in-house 4nm base die and the only turnkey memory-plus-logic-plus-packaging offer — a real differentiator for custom HBM. Micron lands in between. Structurally, TSMC gains a toll on every SK Hynix stack, and the base die becomes the place hyperscalers will customise their memory.
Bottleneck severity
HBM4 doubles the bus to 2,048 bits and puts real logic in the bottom die of the stack — and a memory fab cannot print leading-edge logic. So the base die is now made on a foundry node: SK Hynix on TSMC 12nm (evaluating 3nm for HBM4E), Samsung on its own 4nm. That drags the world's scarcest packaging line into the world's scarcest memory part — and hands TSMC a slice of HBM economics it never had.
Who makes it
SK Hynix000660.KS
#1 HBM maker (~62% share). Qualified HBM4 at NVIDIA first and holds an estimated 60–70% of Rubin's HBM4 allocation, with a TSMC-made logic base die. Sold out through 2027; its CEO calls 2027 potentially the tightest supply year the industry has seen. The bellwether of the whole cycle.
TSMCTSM
Makes the GPU die AND the CoWoS packaging that bonds HBM to it — and now the logic base die under SK Hynix's HBM4 too. CoWoS is heading for ~130k wafers/month by end-2026 and is still sold out into 2027 (NVIDIA alone holds ~60%); it is farming 240–270k wafers/year of front-end steps out to Amkor and SPIL. A Taiwan disruption stops the whole stack.
Samsung005930.KS
#2 in HBM. Began HBM4 mass production in Feb 2026 and is qualified at NVIDIA — with an in-house 4nm base die, the one turnkey memory+logic+packaging offer. P5 Pyeongtaek pulled forward to late 2027; ~$73B cumulative capex through 2026.
MicronMU
#3 HBM, the only US-HQ DRAM maker. Qualified on Rubin HBM4; sold out through 2027 and meeting only ~50–65% of what customers ask for. New Idaho capacity doesn't land until 2027–28 — so the squeeze is structural, not seasonal.
BE SemiconductorBESI
Hybrid-bonding tool leader — the technique HBM4 / future stacking needs. Tiny float, huge optionality.
NVIDIANVDA
Buys the most HBM, NVLink and optics on earth and pre-buys supply to lock rivals out: ~60% of CoWoS, $2B each into Coherent and Lumentum (Mar 2026), $2B into Marvell. B300 (288 GB HBM3E) is the shipping reference GPU; Rubin (288 GB HBM4, 22 TB/s) samples Q4 2026, volume Q1 2027.
Metrics that move this layer
- Rubin HBM4 allocationSK Hynix ~60–70% · Samsung ~25–30% · Micron the rest — the scoreboard of the two approaches
- HBM4E base-die nodeTSMC 3nm confirmation = the moat deepens
- Custom HBM winshyperscaler-specific base dies = the next design-win war
- 16-Hi qualificationQ4 2026 samples to NVIDIA