HBM Memory — Market

Verified claims and product-axis read for HBM Memory. Every fact below is sourced; every product judgment traces back to underlying signals.

Verified facts

• TSMC Q3 2025 gross margin of 59.5% slightly exceeds Q4 2025 gross margin of 59.0%. ↗ (financial)
• Substrate and Interposer Design Principal roles requiring ABF substrate + signal integrity skills posted Feb 27, 2026. ↗ (other)
• Samsung's Memory Business is structurally exposed such that record HBM quarters still hit a 50% PS bonus cap for engineering staff. ↗ (other)
• Director, HBM Product Marketing role (March 25, 2026) with 'AI memory strategy' specialty implies HBM is being treated as a distinct product line with its own marketing org. ↗ (other)
• HBM Customer Applications Engineer postings (May 4, 2026) suggest HBM vendors are deploying field-applications engineers directly with AI accelerator customers. ↗ (other)
• Memory test for HBM Reliability roles posted May 2, 2026 indicate ongoing reliability validation as field-deployment scales. ↗ (other)
• Strix Halo platforms have ~256 GB/s memory bandwidth, M5 Max has ~614 GB/s, per HN comment — far below HBM's TB/s bandwidth. ↗ _(community_sentiment)_
• HN community discussion characterizes HBM as having 'completely shattered the memory wall for individual cores/compute units.' ↗ _(community_sentiment)_
• Multiple HN comments suggest non-HBM datacenter GPUs (e.g., RTX6000) are 'several times slower' than HBM-equipped datacenter GPUs. ↗ _(community_sentiment)_
• HBM3E 12-high (36GB) at Samsung cleared NVIDIA qualification in Q4 2025, with Q4 2025 first shipments per dimension data. ↗ (other)

Top products (engine read)

CoWoS-L Advanced Packaging (HBM Interposer + Base Die) — 2.5D advanced package

Opportunity: CoWoS-L is the binding constraint on HBM-attached AI accelerator output — verified fact e02380e3 confirms TSMC sole-source status for NVIDIA, and demand_query 8f82b1f4 captures market frustration that supply still can't meet demand. Amkor's Arizona buildout + ASE's CoWoS-like 2.5D flows are direct second-source plays.

CoWoS-L is the choke point for the entire HBM → AI accelerator chain. TSMC owns the leading-edge supply (Rubin commitment); Amkor + ASE are scrambling for the overflow / second-source role, with US (Arizona) capacity gated to 2027. The economic margin sits with whoever owns interposer + base die integration.

HBM3E 12-high 36GB Stack — HBM stack

Opportunity: Hyperscaler AI accelerators (NVIDIA Blackwell/H200/GB300, AMD MI300) need 12-high 36GB HBM3E at volumes that exceed current supply — demand_query 8f82b1f4 notes 'HBM Supply Curve Gets Steeper, but Still Can't Meet Demand', and demand_query 01dc22c3 highlights skyrocketing HBM pricing pushing Micron through $45B.

HBM3E 12-high 36GB is the current ramping product across all three memory vendors; SK Hynix leads NVIDIA share, Micron took #2 NVIDIA slot on GB300, Samsung finally cleared NVIDIA qual in Q4 2025. The constraint is wafer supply + CoWoS packaging slots, not demand.

HBM4 12-high 36GB Stack — HBM stack

Opportunity: NVIDIA Rubin (R100) and AMD MI455X are designed around 12-high HBM4 with logic base die on advanced foundry nodes; HBM4 demand query a6562e22 ('What's Different About HBM4') + verified fact b7dd7a82 (SK Hynix + TSMC 3nm base die partnership) show the race to be Rubin's primary HBM4 supplier is the next-leg margin pool.

HBM4 is the explicit roadmap battle for 2026-2027: SK Hynix announced first-development Sep 2025, Samsung sampled Mar 2026 with own 4nm base die, Micron claims high-volume production. The differentiator is the logic base die (TSMC 3nm for SK Hynix vs Samsung Foundry in-house) — winner takes Rubin/MI455X sockets.

AI Accelerator with HBM4 (NVIDIA Rubin / AMD MI455X class) — AI accelerator

Opportunity: Each NVIDIA/AMD/Microsoft AI accelerator socket commits 8-12 HBM stacks × ~36GB; this is the demand engine for everything upstream (HBM3E/HBM4, CoWoS-L, KGSD). demand_query 0d46c5ed (Lisa Su flying to Samsung for HBM) shows accelerator vendors are personally negotiating supply.

AI accelerator SKUs are the demand sink that defines every HBM/packaging program above. The interesting structural shift is hyperscaler in-house silicon (Maia 200 on HBM3e) joining NVIDIA/AMD as HBM buyers — broadening the demand base beyond the GPU duopoly.

See the Products and Strategy modules for the full product list and forward-looking judgment.