HBM Memory — What Matters This Week (May 2026)

1. Top 3 Things Happening (User View)

1. **Samsung ships HBM4 12-high 36GB samples on a Samsung Foundry 4nm logic base die at >2 TB/s/stack**, chasing NVIDIA/AMD qual before 2H 2026 volume (Samsung company module).

• Why it matters to: `HBM-MEM-SYS-ENG-01` and `HBM-PROCURE-SOURCING-01` — HBM4 is the first generation where the base die is a leading-edge logic process, not a legacy DRAM node; signal-integrity sign-off and second-source leverage both change.
• Action: Open a parallel HBM4 SI validation track using Samsung's 4nm-base sample kit this quarter; do not wait for SK Hynix-only qual data to lock vendor terms.

1. **TSMC guided CoWoS capacity to roughly double again in 2026 to ~150k wpm and locked NVIDIA Rubin (R100) on N3P with CoWoS-L and 8-site HBM4 stacks** (TSMC company module, Q4 2025 print).

• Why it matters to: `HBM-PACKAGING-ENG-01` and `HBM-SUPPLY-ANALYST-01` — 8-site HBM per package (vs 6 on Blackwell) is the real capability shift; CoWoS slot reservations for Rubin close before Q3 2026.
• Action: This week, reconfirm CoWoS-L slot reservations against Rubin's 8-site footprint; flag any 6→8 site assumptions in 2026 BoMs that haven't been redone.

1. **SK Hynix declared world-first HBM4 development complete (12-high, 36GB, >2TB/s) ready for Rubin mass production, with HBM already >40% of DRAM revenue in Q3 2025** (SK Hynix company module).

• Why it matters to: `HBM-SUPPLY-ANALYST-01` — first-mover qual on Rubin is the single biggest 2026 allocation variable; OpenAI's Stargate LOI alone commits Samsung + SK Hynix to up to 900k DRAM wafers/month (market module, signal `168ed0d0`).
• Action: Re-price SK Hynix on a Rubin-attached HBM4 ASP, not a DRAM-cycle ASP; the >40% HBM/DRAM mix breaks the historical memory-cycle correlation.

2. Strategic Implications (Our View)

1. **HBM has stopped being a memory product and is now a co-designed logic+memory SoC.** Samsung's 4nm logic base die and SK Hynix's HBM4 spec both put a leading-edge logic process under the DRAM stack.

• Evidence: Samsung company module — "per-stack bandwidth exceeding 2 TB/s and a 4nm logic base die".
• Implication for next 30 days: Our HBM coverage must add a foundry-process axis (base-die node) to every vendor comparison; readers covering memory without foundry context will miss the actual differentiator in 2H 2026 qual results.

1. **The CoWoS bottleneck is moving from "capacity" to "site-count per package."** TSMC is doubling wpm but Rubin's 8-site HBM4 footprint absorbs much of the gain; merchant alternatives (ASE FOCoS-Bridge, Amkor Arizona) are quietly being positioned as the hedge.

• Evidence: TSMC company module guides CoWoS "to again roughly double in 2026"; ASE company module — "advanced packaging ... would account for roughly 10% of ATM revenue in 2026, more than double 2024 levels."
• Implication for next 30 days: Run a standalone piece mapping 2026 advanced-packaging supply by site-count, not wafer-count — that's the metric procurement and sell-side analysts are missing.

1. **The hiring signal is a cold-start senior cohort, not a hype wave.** 30 roles in 90d with no prior-period baseline, skewed to Principal/Staff/Director with advanced packaging as the top specialty (hiring module).

• Evidence: Hiring module — "all 30 roles tagged level=unspecified ... heavy on 'Principal' (5+), 'Staff' (4+), and two 'Director' slots".
• Implication for next 30 days: Treat HBM hiring as a leadership-tier rebuild, not a junior ramp; jobseeker-targeted content should be aimed at 10+ YOE packaging/memory architects, not new grads.

3. Newsletter Issue Spotlight

• Lead story: Samsung's HBM4 12-high samples ship on a 4nm logic base die — memory has become a foundry product.
• 3-second hook: HBM4's base die is now a 4nm logic chip. Memory just became a SoC.
• Image direction: V16 isometric 3D exploded view — a single HBM4 stack lifted off a CoWoS-L interposer; 12 DRAM dies floating in a column above a glowing 4nm logic base die clearly labeled as a foundry chip, with a Rubin compute die docked adjacent on the same interposer. Cool blue/silver palette, faint wafer-map watermark underneath.
• Subject line: HBM4: memory is now a 4nm logic chip

4. Reverse-Hype Watch

• Overhyped right now: "Samsung is closing the HBM gap." Samsung HBM3E NVIDIA qual is still *in progress* per the Samsung module (`betting Pyeongtaek P4 and HBM4 close the SK Hynix gap`); a 4nm base die HBM4 sample is not the same as Rubin volume qual. The narrative has moved faster than the qual data.
• Underrated right now: **ASE FOCoS-Bridge as a merchant CoWoS alternative.** ASE company module shows advanced packaging guided to ~10% of 2026 ATM revenue (>2x 2024), backed by ~US$1.6B incremental capex at K28 — but it barely surfaces in mainstream HBM coverage versus TSMC CoWoS-L.

5. Pattern Library Updates

• Success pattern observed: **HBM generation leadership compounds into accelerator lock-in.** SK Hynix's HBM3E 12-high lead on Blackwell flowed directly into being first-to-complete HBM4 for Rubin (SK Hynix module — "first to complete HBM4 development, ramping for Rubin"). Memory leadership is sticky across one accelerator generation.
• Failure pattern observed: **Treating second-source narrative as if qual were complete.** Samsung's HBM3E qual cycle on NVIDIA ran through most of 2024–2025 even with announced wins; readers/analysts who modeled share-shift on announcements rather than on qualified-shipping data were repeatedly early. Avoidance strategy: gate any share-shift call on a named qualified product shipping in volume (e.g., Micron HBM3E 12-high on GB300 per Micron module), not on samples or capex.