🔥36氪•Freshcollected in 2m
SK Hynix predicts HBM prices to double by 2027
💡Critical supply chain insight: HBM shortages will continue to drive up AI infrastructure costs through 2030.
⚡ 30-Second TL;DR
What Changed
HBM prices expected to double by 2027 due to extreme supply constraints.
Why It Matters
The sustained HBM shortage will likely keep AI training and inference costs high for developers. Companies may need to secure long-term supply agreements to avoid hardware bottlenecks.
What To Do Next
Evaluate your long-term compute requirements and consider diversifying hardware vendors to mitigate potential HBM-related supply chain risks.
Who should care:Founders & Product Leaders
Key Points
- •HBM prices expected to double by 2027 due to extreme supply constraints.
- •SK Hynix faces long-term demand that exceeds current expansion plans.
- •Chinese memory firms like CXMT and YMTC are emerging as key market variables.
🧠 Deep Insight
AI-generated analysis for this event.
🔑 Enhanced Key Takeaways
- •SK Hynix has shifted its capital expenditure strategy to prioritize HBM3E and HBM4 production, dedicating over 60% of its DRAM wafer capacity to high-bandwidth memory products by mid-2026.
- •The transition to HBM4 requires a fundamental shift in manufacturing processes, specifically the adoption of 12-layer and 16-layer stacking using advanced MR-MUF (Mass Reflow Molded Underfill) technology.
- •Major hyperscalers, including NVIDIA and AMD, have signed multi-year take-or-pay agreements with SK Hynix, effectively locking up the majority of the company's HBM supply through 2028.
- •SK Hynix is accelerating the development of 'custom HBM' solutions, where logic dies are tailored to specific AI accelerator architectures, moving away from standardized JEDEC-only specifications.
- •The emergence of Chinese competitors like CXMT is primarily impacting the legacy DRAM and LPDDR markets, forcing SK Hynix to exit lower-margin commodity memory segments to focus exclusively on high-margin HBM.
📊 Competitor Analysis▸ Show
| Feature | SK Hynix (HBM3E/4) | Samsung (HBM3E/4) | Micron (HBM3E) |
|---|---|---|---|
| Stacking Tech | MR-MUF | TC-NCF | TC-NCF |
| Market Position | Leader (NVIDIA primary) | Challenger (Qualifying) | Fast Follower |
| Capacity Focus | Aggressive HBM-first | Balanced DRAM/HBM | Niche/High-Efficiency |
🛠️ Technical Deep Dive
- HBM4 Architecture: Utilizes a 2048-bit wide interface, doubling the bandwidth of HBM3E, and integrates a base die manufactured on a 12nm-class logic process.
- MR-MUF Technology: SK Hynix's proprietary thermal management technique that improves heat dissipation by 20% compared to traditional Non-Conductive Film (NCF) methods, critical for 16-high stacks.
- Power Efficiency: New HBM4 designs target a 30% reduction in power consumption per gigabyte per second (GB/s) compared to HBM3E, addressing thermal throttling issues in massive AI clusters.
🔮 Future ImplicationsAI analysis grounded in cited sources
HBM will account for over 40% of total DRAM revenue by 2027.
The rapid shift of wafer capacity toward HBM, combined with doubling price projections, will fundamentally alter the revenue mix of major memory manufacturers.
Custom HBM will become the industry standard for flagship AI chips.
Standardized HBM is increasingly insufficient for the specific thermal and bandwidth requirements of next-generation AI accelerators, forcing a move toward bespoke logic-die integration.
⏳ Timeline
2023-06
SK Hynix begins mass production of HBM3, securing early dominance in the AI memory market.
2024-03
SK Hynix officially commences mass production of 8-layer HBM3E for AI accelerators.
2025-05
SK Hynix announces the successful development of 16-layer HBM3E, setting a new industry standard for capacity.
2026-02
SK Hynix unveils its HBM4 roadmap, emphasizing custom logic die integration for major AI partners.
📰
Weekly AI Recap
Read this week's curated digest of top AI events →
👉Related Updates
AI-curated news aggregator. All content rights belong to original publishers.
Original source: 36氪 ↗