SK Hynix 54 Trillion Won CapEx Locked In; Nvidia Rubin Ultra Downgrading HBM Stack Amid Structural Supply Crunch
SK Hynix has formally committed 54 trillion won (~$39B USD) to new fabs at Yongin Y2 and Cheongju M17, with first meaningful HBM production not expected until 2027–2029 after 12–18 month yield ramp periods — offering no near-term supply relief. Concurrently, Nvidia is actively testing at least three downgraded HBM configurations for its next-generation Rubin Ultra GPU — including a reduction from the planned 16-stack HBM4E to 12-stack and potentially 8-stack HBM4/HBM4E variants — a direct product-specification response to confirmed HBM supply constraints, with AMD's MI400 reportedly following a parallel dual-stack strategy.
The combination of a multi-year CapEx commitment with no meaningful output before 2027–2029 and Nvidia actively spec-stepping its flagship GPU to accommodate HBM scarcity confirms that HBM is the binding constraint on AI accelerator supply chains through at least 2027 — any AI server BOM model should price in sustained HBM premium and flag Rubin Ultra yield and configuration risk as a lead-time variable. The Lam Research advanced packaging growth revision from >40% to >70% independently corroborates accelerating CoWoS/2.5D packaging demand pull, signaling that packaging capacity investment is compounding alongside — not ahead of — HBM supply, keeping the overall bottleneck structural rather than transient.
Sources
- SK Hynix Newsroom — SK hynix Invests 54 Trillion Won in Yongin Y2 and Cheongju M17 to Secure Mid-to-Long-Term Production Base for AI Memory Demand
- SK Hynix Newsroom — SK hynix Invests 54 Trillion Won in Yongin Y2 and Cheongju M17 to Secure Mid-to-Long-Term Production Base for AI Memory Demand
- SK Hynix Newsroom — SK hynix Invests 54 Trillion Won in Yongin Y2 and Cheongju M17 to Secure Mid-to-Long-Term Production Base for AI Memory Demand
- TechTimes — SK Hynix Board Locks Korean AI Memory Through 2033; US Has Zero HBM Wafer Fabs
- Donga — 엔비디아, HBM 공급부족 대응… 차세대 GPU에 저사양칩 검토
- V (Daum/아시아경제) — '살 안 빼고 못 찍어내면 말짱 도루묵'…'최고 성능 자랑' 버리고 'HBM 강제 다이어트' 시키는 AI 공룡들[칩톡]
- AOL / Lam Research — Lam Research (LRCX) Q4 2026 Earnings Call Transcript
- Debuglies — Monopolizing Compute: Geopolitical Impact of Semiconductor Inflation
Related data & tools
Storylines
Follow this signal over time — every brief threaded by entity.
Silicon Analysts Research
All benchmarks and data are derived from publicly available sources (earnings calls, press releases, analyst reports, regulatory filings). Figures are estimates for educational purposes only and should not be used as the sole basis for business or investment decisions. Public sources only — no insider data. Generated from the cited reports; verify against the originals before acting. Full Terms & Data Provenance