HBM3E / HBM4
▶ See it in 3D The Memory Stack region
If you remember one stop, make it this one. Just three firms make HBM and it is sold out through 2027 — gated from below by TSMC's CoWoS packaging, ASML's EUV monopoly and Ajinomoto's 95% grip on ABF. The bottleneck is packaging, not the chip. Tap "Who makes it" in the panel.
Picture bolting a stack of ultra-fast memory chips directly onto the GPU so data barely has to travel. That's HBM. It is the scarcest, most fought-over part in all of AI: only three companies can make it, it costs 5–10× normal memory, and it is sold out through 2027. If you remember one thing from this whole map, make it this.
What it is
High Bandwidth Memory — DRAM dies stacked 8–12 high with through-silicon vias, bonded to the GPU package via a silicon interposer. A single stack runs a 1,024-bit bus. The single most fought-over component in AI.
The business
The center of gravity. A 3-way oligopoly (SK Hynix, Samsung, Micron) makes the stacks — all three are now qualified on NVIDIA's Rubin HBM4, with SK Hynix holding ~60–70% of that allocation. TSMC CoWoS packaging is the real ceiling on supply (and now makes SK Hynix's base die too); Besi/ASML/Japan materials gate it from below. HBM is ~5–10× the cost/GB of DDR, is sold out through 2027, and is eating the DRAM wafers commodity memory used to get — see The DRAM Cycle.
Bottleneck severity
Only three companies on earth can make HBM (SK Hynix, Samsung, Micron) — and even their output is capped from outside: every stack must be bonded to the GPU on TSMC's CoWoS packaging line (sold out into 2027) and its leading DRAM dies are printed on ASML's EUV machines (a literal monopoly). All three are sold out through 2027 and filling only 50–70% of what customers ask for, so allocation, not price, decides who ships GPUs.
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.
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.
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.
BE SemiconductorBESI
Hybrid-bonding tool leader — the technique HBM4 / future stacking needs. Tiny float, huge optionality.
ASMLASML
Monopoly on EUV lithography — every leading-edge logic + DRAM die needs it. Ultimate chokepoint.
Lam ResearchLRCX
Etch & deposition tools for HBM's through-silicon vias. HBM tool revenue +50% YoY — equipment ordered today = chips in 12–24 months.
Disco6146.T
Near-monopoly on wafer dicing & grinding — every advanced chip is thinned and cut on Disco tools. A quiet utilization barometer.
ASE TechnologyASX
World's largest OSAT. Takes TSMC's CoWoS overflow (SPIL ~60–80k wafers in 2026) and expects advanced-packaging sales to double — a relief valve for the bottleneck, not a substitute at the leading edge.
Ajinomoto2802.T
Yes — the food company. Invented ABF, the insulating film inside every high-end chip substrate. ~95% of the market, running flat out (~2M m²/month), a new pricing system from Q3 2026 up to +30%, and supply to China cut ~30% to serve everyone else. The quietest monopoly in AI.
Ibiden4062.T
~35% of high-end FC-BGA substrates — the board every GPU/CPU package rides on. ¥500B FY2026–28 build-out, but mass production only from FY2027, so the substrate shortage stands through 2026. AI-substrate load index: 1.8× 2024 in 2026, 2.5× by 2028.
Shin-Etsu4063.T
World #1 silicon wafers + photoresist. Japan's quiet grip on materials.
Metrics that move this layer
- HBM bit-growthindustry HBM bit supply YoY — the core volume signal
- HBM share of DRAM wafers~18% end-2025 → ~22% end-2026 → ~30% end-2027: the crowd-out tape
- CoWoS capacity~130k wafers/month by end-2026; still sold out
- HBM ASP / marginmemory-maker gross margin = cycle health
- Rubin HBM4 allocationwho gets qualified at NVIDIA = who wins; 16-Hi samples due Q4 2026