SK Hynix's Post-HBM Bet Could Redraw the AI Memory Race

SK Hynix Isn’t Just Defending HBM. It’s Building a Memory Ecosystem.

At SEDEX 2026 in Seoul, SK Hynix’s head of development, Ahn Hyun, delivered a message that was far more ambitious than a typical product roadmap. He didn’t just talk about scaling HBM — he outlined a complete upgrade path for AI memory, starting with what comes after HBM and ending with a new class of storage designed specifically for inference.

The pitch was stark: AI demand is “just beginning,” and the data-center memory market will grow roughly 20-fold by 2030, at a compound annual rate of 109 percent. The implication is that whoever controls the memory stack controls the pace of AI progress.

“Memory is no longer just a component,” Ahn said. “It is the factor that will determine whether AI can grow further.”

The 3D-Stacked DRAM Bet

SK Hynix’s most consequential announcement was its push into 3D-stacked DRAM — a technology that places DRAM directly on top of a GPU or AI accelerator using silicon through-vias, rather than stacking memory chips beside the processor the way HBM does.

The difference matters. HBM already delivers massive bandwidth, but it sits adjacent to the GPU, connected by a relatively long signal path. 3D-stacked DRAM eliminates much of that distance, potentially multiplying bandwidth by tens of times compared to current HBM implementations while cutting power consumption. In effect, SK Hynix is positioning this technology as the missing layer between on-chip SRAM — the fastest but smallest and most expensive memory inside a GPU — and off-chip HBM.

The engineering challenges are serious. Heat dissipation when you sandwich DRAM on top of a logic die is a well-known problem, and advanced packaging capabilities are not something you acquire overnight. But SK Hynix has spent the last three years perfecting HBM packaging, and the company is clearly treating 3D-stacked DRAM as its next natural step.

If it works at scale, it changes the arithmetic of AI training. GPUs paired with 3D-stacked DRAM could process larger models with fewer memory bottlenecks, and the power efficiency gains would matter enormously at data-center scale.

Custom HBM Is the Quiet Shift

While 3D-stacked DRAM grabbed the headlines, SK Hynix also signaled a structural change in how it sells HBM itself. The company expects custom HBM — where the base die at the bottom of the stack is designed to a specific customer’s requirements — to grow steadily alongside standard HBM products.

This is a subtle but important pivot. NVIDIA, AMD, and the growing number of in-house AI accelerator designers from Google, Amazon, and Microsoft all have different memory profiles. Some prioritize raw bandwidth; others prioritize power efficiency or capacity. Custom HBM lets SK Hynix lock in these customers with designs that are harder to replicate and more deeply integrated into their chip architectures.

Standard HBM remains important — it’s the product that scales volume. But custom HBM is where the margins and the relationships deepen. SK Hynix is essentially saying: we will build the memory that fits your chip, not just the chip that fits our memory.

HBF Targets the Inference Explosion

Perhaps the least discussed but strategically sharpest move in Ahn’s presentation was the emphasis on HBF — High Bandwidth Flash. This is a NAND-flash-based memory technology that uses TSV and logic-process techniques to deliver both high capacity and high bandwidth, landing somewhere between HBM and SSD in the performance hierarchy.

HBM excels at speed but struggles with capacity and cost. SSDs offer capacity but lack the bandwidth that AI inference workloads increasingly demand. HBF aims to occupy the middle ground — a memory tier optimized for the kind of large-scale, read-heavy workloads that define inference.

As AI models move from training into production, inference is becoming the dominant memory consumer. Every chat response, every image generation, every search query pulls massive datasets through memory. An inference-optimized memory tier that is cheaper than HBM but faster than SSD could capture a market segment that will dwarf training workloads within a few years.

SK Hynix framed HBF explicitly around this shift: “The capacity and bandwidth of HBF are suited to the repeated large-data reads that define AI inference.”

What This Means for Samsung

Samsung has historically led in memory technology but has lagged behind SK Hynix in HBM qualification and shipping volume. NVIDIA’s HBM3e supply, for example, went almost entirely to SK Hynix, a deal that fundamentally reshaped the competitive landscape.

SK Hynix’s expansion into 3D-stacked DRAM, custom HBM, and HBF is a broader play than simply adding more HBM SKUs. It is an attempt to define the entire memory architecture for the next generation of AI chips — not just the fastest tier but every tier below it.

For Samsung, the threat is existential in a new way. If SK Hynix becomes the default memory partner for multiple GPU architectures — each with its own custom HBM and its own 3D-stacked DRAM configuration — Samsung risks being pushed to the commodity end of the market, competing on price for standard HBM while SK Hynix captures the high-value custom work.

Samsung has responded with its own HBM3e efforts and a public commitment to close the gap, but catching up on custom memory partnerships requires more than process improvements. It requires deep integration with GPU designers, and those relationships take years to build.

The Bigger Picture

Ahn Hyun’s speech was notable not for any single product announcement but for the scope of the vision. SK Hynix is no longer just a memory supplier — it is positioning itself as the architect of a multi-tier memory system that spans from on-GPU stacks to flash-class inference memory.

The numbers support the ambition. A 109 percent compound growth rate in AI infrastructure spending through 2030 is not a projection based on incremental demand. It is a bet on a fundamental reordering of computing — one where memory, not just processing, determines how fast AI can advance.

SK Hynix is betting that the next phase of AI memory competition will be won not by whoever makes the fastest HBM, but by whoever controls the most complete memory stack. If that bet pays off, the company doesn’t just maintain its lead — it redefines what a memory company does in the AI era.

Samsung will be watching closely. So will every GPU designer who needs to decide whose memory architecture to build around next.