SK Hynix Expands Dalian NAND Line for AI Storage

SK Hynix is advancing its Floating Gate (FG) NAND technology with a new production line in Dalian, China. This move targets the booming demand for high-capacity storage driven by AI inference.

The South Korean memory giant aims to build a "mid-200-layer" facility at its second Dalian plant. Reports indicate this will feature approximately 250 layers of 3D NAND flash memory.

This strategic expansion highlights SK Hynix's unique position in the global semiconductor market. Unlike competitors, it retains legacy FG technology crucial for specific AI applications.

Key Facts and Strategic Timeline

• Location: The new production line will be established at SK Hynix's second factory in Dalian, China.
• Technology: It utilizes Floating Gate (FG) structure rather than the industry-standard Charge Trap (CT) architecture.
• Layer Count: The facility targets a "mid-200-layer" density, estimated at roughly 250 layers.
• Timeline: A pilot line is scheduled for Q3 2026, with full mass production targeted for H2 2027.
• Market Position: SK Hynix remains the only major manufacturer using FG NAND after acquiring Intel's NAND business.
• Current Capacity: The existing Dalian Plant 1 already produces 192-layer FG NAND at scale.

Why Floating Gate Technology Matters for AI

The semiconductor industry has largely shifted toward Charge Trap (CT) structures for 3D NAND. Most major manufacturers, including Samsung and Micron, rely on CT for its scalability and reliability in standard consumer electronics. However, SK Hynix maintains a distinct advantage through its acquisition of Intel's NAND division.

This acquisition allowed SK Hynix to retain expertise in Floating Gate (FG) technology. While CT dominates general-purpose storage, FG offers superior performance for high-density configurations. Specifically, FG is better suited for Quad-Level Cell (QLC) NAND. QLC stores four bits per cell, maximizing capacity but requiring precise voltage control that FG provides more naturally than CT.

AI workloads, particularly inference tasks, drive the need for this specific technology. Inference involves reading massive datasets repeatedly without frequent writing. This creates a read-intensive workload profile. QLC SSDs provide the high capacity needed for large language models at a lower cost per gigabyte. SK Hynix's commitment to FG ensures it can produce these specialized drives efficiently.

The Solidigm Connection

The continuity of FG technology also benefits Solidigm, SK Hynix's US-based subsidiary. Solidigm focuses on enterprise and data center storage solutions. By maintaining the FG roadmap, SK Hynix ensures Solidigm can develop coherent product lines. This avoids the disruption of switching architectures mid-cycle. It allows for steady innovation in enterprise-grade storage hardware tailored for cloud providers and AI farms.

Industry Context: The AI Storage Bottleneck

The rapid growth of artificial intelligence has created a severe bottleneck in data infrastructure. Training large models requires speed, but inference requires volume. As companies deploy AI agents and customer-facing chatbots, the volume of data accessed daily explodes. Traditional HDDs are too slow, while expensive TLC NAND SSDs strain budgets.

QLC NAND emerges as the ideal compromise. It offers densities that approach hard drives with speeds closer to premium SSDs. However, early QLC implementations suffered from slower write speeds and lower endurance. SK Hynix's FG-based QLC addresses these concerns by leveraging mature physics. The floating gate structure provides more stable electron retention over time compared to charge trap alternatives in high-layer counts.

Western tech giants like Microsoft, Amazon, and Google are aggressively expanding their data centers. They require storage that balances cost, density, and performance. SK Hynix's Dalian expansion directly targets this demand. By localizing production in China, they also navigate complex supply chain dynamics. This strategy reduces reliance on single-region manufacturing risks.

What This Means for Developers and Businesses

For software engineers and IT managers, this development signals a shift in available hardware. Expect to see more affordable, high-capacity NVMe SSDs hitting the market by 2028. These drives will be optimized for read-heavy operations typical of vector databases and model weights storage.

Businesses planning AI deployments should consider storage architecture early. Investing in systems compatible with high-density QLC drives now may yield long-term savings. As SK Hynix ramps up production, prices for enterprise-grade QLC storage could become more competitive. This democratizes access to large-scale AI infrastructure for smaller enterprises.

Developers should optimize their data retrieval patterns for QLC characteristics. Understanding the difference between write-intensive training phases and read-heavy inference phases is critical. Leveraging FG-based QLC for inference caches can significantly reduce operational costs. This technical nuance becomes a key factor in total cost of ownership calculations for AI projects.

Looking Ahead: The Road to 2027

The timeline for this project is ambitious but clear. SK Hynix has already completed R&D for 200-layer-plus FG NAND. The immediate focus is equipment procurement for the pilot line in Q3 2026. This phase will validate the manufacturing process and yield rates.

Full mass production is slated for the second half of 2027. This gives the industry a two-year window to prepare. Supply chain partners and system integrators must align their roadmaps accordingly. The success of this line could redefine the hierarchy of memory technologies in the AI era.

If SK Hynix succeeds, it may force competitors to reconsider their abandonment of FG technology. Alternatively, it could cement FG as a niche but vital technology for specific AI workloads. Either way, the Dalian plant serves as a critical testbed for next-generation storage economics.

Gogo's Take

• 🔥 Why This Matters: This isn't just about making chips; it's about fueling the AI inference boom. By sticking with Floating Gate tech, SK Hynix is betting big on QLC NAND's ability to store massive datasets cheaply. For businesses, this means cheaper, denser storage for running AI models, potentially lowering the barrier to entry for enterprise AI adoption.
• ⚠️ Limitations & Risks: Geopolitical tensions remain a significant risk for any manufacturing footprint in China. Additionally, while FG is great for QLC, it faces scaling challenges at extreme layer counts compared to CT. If yields struggle, SK Hynix could face higher costs than rivals using more standardized processes.
• 💡 Actionable Advice: Monitor SSD pricing trends for enterprise QLC drives starting in 2027. If you are architecting AI infrastructure, evaluate your read/write ratios. Shift read-heavy workloads like model serving to QLC-based storage to capitalize on potential price drops. Keep an eye on Solidigm's product announcements for direct integration opportunities.