HBM Memory Product Landscape — May 2026

Product Categories

• **HBM3E 12-High Stacks in Volume**: 36GB twelve-stack HBM3E DRAM in mass production today against named hyperscaler accelerator qualifications (NVIDIA Blackwell / Blackwell Ultra).
• **HBM4 12-High Stacks at Sampling**: 36GB twelve-stack HBM4 at >2.0 TB/s per stack — shipped as customer samples since June 2025, but not yet customer-qualified at volume.
• **In-House 2.5D Advanced Packaging (TSMC CoWoS-L / SoIC)**: Vertically integrated silicon interposer + advanced packaging on TSMC, supporting >800 mm² compute die and 8-site HBM4 stacks.
• **Merchant OSAT 2.5D (ASE VIPack / Amkor 2.5D)**: Third-party advanced packaging alternatives to CoWoS, supporting 8–12 HBM stacks per package, positioned as second-source overflow rather than primary volume.

Comparison Table

| Product | Company | Lead Tech Spec | Calibration Score (0–100) | Use Case | |---|---|---|---|---| | HBM3E 12-High 36GB (Blackwell Ultra qual'd) | SK Hynix | ~1.25 TB/s/stack, MR-MUF stack | 88 | NVIDIA GB200 / GB300 mass production, Mar 2025 → present | | HBM4 12-High 36GB (dev completed Sept 2025) | SK Hynix | >2.0 TB/s/stack, TSMC N3 logic base die | 88 | NVIDIA Rubin volume ramp, H2 2026 | | HBM3E 12-High (NVIDIA qual Sept 2025) | Samsung Electronics | ~1.2 TB/s/stack, NCF stack | 52 | NVIDIA Blackwell second-source supply, Q4 2025 ship-start | | HBM4 12-High 36GB (4nm Samsung Foundry base die) | Samsung Electronics | >2.0 TB/s/stack, 4nm in-house logic base | 55 (pre-qual) | Samples Mar 2026 → Rubin / AMD HBM4 qual gate H2 2026 | | HBM3E 12-High 36GB (NVIDIA GB300 shipping) | Micron Technology | ~1.25 TB/s/stack | 75 | NVIDIA Blackwell Ultra GB300 volume shipments, Sept 2025 → present | | HBM4 12-High 36GB (June 2025 customer sample) | Micron Technology | >2.0 TB/s/stack (+60% vs HBM3E) | 70 | Second-hyperscaler HBM4 qualification is the 2026 gate | | CoWoS-L (NVIDIA Rubin / Micron HBM4 base die) | TSMC | 8-site HBM4 on >800 mm² die, ~75k wpm exit-2025 | 85 | NVIDIA Rubin R100, AMD MI400-class, AWS Trainium 2026 builds | | VIPack / FOCoS-Bridge (NVIDIA Blackwell back-end) | ASE Technology | 8–12 HBM3E/HBM4 stacks on >2× reticle interposer | 75 | Merchant CoWoS-overflow allocation, 2025–2026 | | 2.5D Advanced Packaging (TSMC Arizona partner) | Amkor Technology | "Multiple HPC/AI" 2.5D flows; Peoria first-prod 2027 | 70 | US-domestic CoWoS-class back-end (real volume 2027) |

Differentiation Map

• **Best for NVIDIA Blackwell / Blackwell Ultra inference shipping today**: SK Hynix HBM3E 12-high — first to 12-high mass production with MR-MUF stack-yield advantage (39652c4a, 4b1f789e).
• **Best for NVIDIA Rubin GA in H2 2026**: TSMC CoWoS-L paired with SK Hynix HBM4 — only stack with three consecutive years of on-cadence CoWoS doubling and completed HBM4 development (b0978465, bdb41bd3, e6ed7351, 39652c4a, d9a7b1ad).
• **Best for cost-relieving CoWoS allocation pressure on Tier-2 customers**: ASE VIPack / FOCoS-Bridge — only merchant >2× reticle interposer publicly disclosed at 8–12 HBM stacks (cfd0d86f, 07916e28).
• **Best for vertical-integration moat on the HBM4 base die**: Samsung HBM4 12-high with 4nm Samsung Foundry logic base — only HBM supplier not dependent on TSMC for the logic base die (601b7bcf).
• **Avoid if banking on US-sourced volume HBM packaging in 2026**: Amkor Peoria — first-production target is 2027; "US-sourced" AI packaging today still ships from Asia sites (3bf92143).
• **Avoid if treating Samsung HBM4 4nm-base differentiation as already-shipped**: same "samples now, volume 2H 2026" wording that slipped Samsung HBM3E by ~3 quarters and cost ~55% YoY division OP (745dabbc, 775edbb3).

Tool: Interactive Comparator

[Link to /tools/hbm-memory-compare] — Phase G will wire filters on calibration score, per-stack bandwidth, base-die foundry, max HBM stacks supported on package, and NVIDIA-qualification status, sourced from the comparison table rows above.

Reverse-Hype Warnings

The single most overpromised position in this landscape is Samsung's HBM4 12-high "samples March 2026, customer qual 2H 2026" narrative (745dabbc): the wording is verbatim the script Samsung used for HBM3E 12-high in 2024, a programme that ultimately slipped roughly three quarters of NVIDIA qualification and cost the memory division ~55% YoY OP in Q2 2025 (775edbb3). The 4nm Samsung Foundry base die (601b7bcf) is genuinely the only vertical-integration lever among HBM suppliers, but until it carries a named NVIDIA or AMD HBM4 qualification it is a moat on paper, not at the customer interface — score upside to ~75 is binary on the next 6–9 months. TSMC CoWoS-L sits at the opposite extreme: the product is supply-constrained rather than overhyped, with allocation rationing through 2024–2025 forcing NVIDIA, AMD and AWS into quarterly capacity negotiation (37dafbd7) — bullish on demand, but the same dynamic creates real customer-churn risk the day ASE FOCoS-Bridge or Samsung I-Cube hit yield parity. The underrated product in the set is Micron HBM3E 12-high on GB300 (6979b188): it is widely dismissed as the third HBM source despite carrying the same NVIDIA Blackwell Ultra qualification that anchors SK Hynix at 88 — the 13-point gap to SK is yield history and case-study volume, not 2025-shipped capability (4aac4e72). Also underrated: TSMC's SoIC / hybrid-bonding readiness for the HBM4-plus 3D step (cd8e59e9, c06c041a) — the thermal-mechanical-reliability and hybrid-bonding postings of March 2026 are a leading indicator that known-good-stack and warpage on 3D HBM stacks remain unsolved at AI-accelerator volume by *any* incumbent. Finally, Amkor's 11.8% Q2 gross margin (c73672c9) is the cleanest reverse-hype datapoint here: HBM-class 2.5D capability does not translate to OSAT pricing power when TSMC is the upstream partner, regardless of how flattering the Arizona / CoWoS-class language sounds.