Memory-procurement lead at hyperscaler / accelerator vendor.

Audience Profile

• **Age / Experience:** 12-25 years
• **Current role:** Strategic Sourcing / Memory Procurement / Component Engineering
• **Top pain points:**
• Three-vendor negotiation leverage
• Long lead-times force 18-month forward commits
• Quality variance between vendor lots
• **Top decision blockers:**
• Vendor-neutral qualification cost
• Internal demand forecast accuracy

What This Segment Needs

• **Information:** Quarterly HBM allocation share, CoWoS/advanced-packaging wpm commitments, lot-level yield variance, JEDEC qual report dates per accelerator partner.
• **Tools:** Vendor-neutral KGSD test harness, multi-supplier lot-tracking dashboard, forward-commit scenario model tied to your accelerator roadmap.
• **Services:** Independent failure-analysis lab access, second-source qualification engineering, geopolitical exposure briefings.

Top 5 Companies for You (Fit Score)

| Rank | Company | Score | Why | |------|---------|-------|-----| | 1 | TSMC | 83/100 | FY2025 revenue +34% YoY; CoWoS doubling to ~75,000 wpm by end-2025, 2x again in 2026; Micron HBM4 base die (Apr 2025) + NVIDIA Rubin (Sept 2025) wins. Your HBM forward-commit math runs through their CoWoS allocation queue. | | 2 | SK Hynix | 79/100 | Q3 2025 OP 11.4T KRW; HBM >40% of DRAM revenue; world-first HBM4 dev completed 2025-09-26; HBM3E 12-high shipped to NVIDIA March 2025. Strongest HBM track record; China export controls cap supply flexibility. | | 3 | Micron | 78/100 | FQ4 2025 data-center revenue +100% YoY; HBM run-rate >$8B; HBM3E 12-high 36GB shipping in volume to NVIDIA GB300; $2B Manassas expansion (Jun 12, 2025). Diversifies Korean-only sourcing; 2 CFO changes in <12 months. | | 4 | ASE Technology | 75/100 | Q3 2025 revenue NT$165.8B (+11% YoY); US$1.6B K28 capex for 2.5D/3D; 2026 advanced packaging ~10% of ATM (>2x 2024); NVIDIA Blackwell back-end partner for HBM3E. Second-source path around the TSMC CoWoS bottleneck. | | 5 | Samsung | 73/100 | Q3 2025 OP 12.2T KRW reverses Q2 -55% YoY; HBM3E 12-high passed NVIDIA qual 2025-09-04, Q4 2025 ship start; OpenAI Stargate LOI 2025-10-01; 20T KRW P4 Phase 2 capex. Recovering after multi-quarter HBM3E slip. |

Deal-Breakers (Your Hard Preferences)

No hard preferences declared for this segment.

How to Evaluate Any Vendor in this Niche (Checklist)

• [ ] **Check growth signals:** Ask for committed advanced-packaging wpm and your specific allocation % through Q4 2026 — not headline capex.
• [ ] **Check comp data:** Cross-reference vendor's reported HBM ASP against TrendForce/DRAMeXchange monthly; gap >8% on the quote means renegotiate.
• [ ] **Check learning signals:** Request the latest JEDEC HBM qual report date for the exact stack height + your accelerator (GB300 / MI400 / in-house ASIC).
• [ ] **Check stability signals:** Pull top-3 customer % of HBM revenue from the latest 10-Q / 공시 — if you'd be customer #4+, expect allocation cuts in tight quarters.
• [ ] **Check culture signals:** Ask the account team to name the on-call FA engineer for your lot — vague answer = no real escalation path.
• [ ] **Check geopolitical exposure:** % of HBM/packaging capacity on a single island (TW) or under US-China export-control overlap; require dual-region within 18 months.
• [ ] **Check lot variance:** Request lot-level KGSD pass-rate σ over last 4 quarters — your "quality variance" pain lives here, not in headline yield.

Reverse-Hype Watch

• **"Sample → volume" HBM4 language (March / 2H 2026):** Samsung used the identical pre-qual phrasing on HBM3E 12-high, which slipped 3+ quarters before passing NVIDIA qual on 2025-09-04. Treat any vendor's "samples in" date as +9 months for volume.
• **OpenAI Stargate "up to 900,000 DRAM wafers/month":** LOI co-signed with SK Hynix — capacity claim not yet backed by firm POs. Discount it when modeling your own allocation squeeze.
• **CoWoS allocation politics:** TSMC's 75,000 wpm headline says nothing about how it's split — Microsoft/Meta/Google volume re-forecasts can shrink your slice 30-50% inside a quarter.

What's under-reported for memory procurement: lot-level yield σ per fab line (vs. headline yield), TSV-bonder uptime and maintenance-window cadence that drives 2-3 week mid-quarter supply gaps, and Tier-2 hyperscaler de-prioritization patterns when Tier-1 customers re-forecast. None of this surfaces in earnings calls — you only see it in your own qual data and supplier weekly ops calls.