South Korea’s SK Hynix unveiled a broad lineup of next-generation memory products and technology strategies tailored for AI infrastructure, expanding its collaboration with key customer Dell Technologies.

SK Hynix announced on the 26th that it participated in the “Dell Technologies Forum 2026 (DTF 2026)” held at COEX in Seoul’s Gangnam district on the 25th, showcasing sixth-generation high-bandwidth memory (HBM) HBM4, AI server DRAM, enterprise solid-state drives (eSSD), and client SSDs (cSSD). The exhibition space was arranged in a server rack format so attendees could view memory products applied in AI servers at a glance.

In the HBM lineup, HBM4 wafers were displayed alongside HBM4 and the seventh-generation HBM4E. DRAM products on display included SOCAMM2, 3DS RDIMM, RDIMM, CSODIMM, LPCAMM2, LPDDR5X, and GDDR7, covering server, mobile, and graphics applications.

eSSD products targeting AI data centers were also unveiled. Notable offerings included the liquid-cooling-enabled “PEB210 E1.S,” the high-capacity, low-power “PS1110 E3.S,” and the “PS1101 E3.S”—SK Hynix’s first QLC (quad-level cell)-based eSSD.

SK Hynix delivered its first keynote address at the forum as a partner company representative. Joo Young-pyo, Vice President of System Architecture at SK Hynix, spoke on the theme “The Invisible Engine: Technology Completing the AI Era,” explaining the role and development direction of next-generation memory technology.

He emphasized that in the agentic AI era, incremental HBM improvements alone will struggle to keep pace with exploding data demands, arguing that “3D stacking” technology—placing memory directly on top of GPUs—is necessary. This is because in environments where multiple AI agents divide roles, repeatedly perform inference, and exchange information, the volume of data that memory must process grows far more rapidly than before.

Citing the example of a GPT-4-class mixture-of-experts (MoE) model, Joo explained that storing weights alone requires 3.5 terabytes (TB) of memory, and generating 20 tokens per second requires reading 4.4TB of data per second from memory. For coding agents, processing demands could reach hundreds to thousands of tokens per second.

“In AI systems, how much bandwidth memory can supply determines overall system performance,” he said. “If you ask whether continuing to refine HBM as we do today will keep supplying sufficient performance going forward, unfortunately the answer is no.”

The solution SK Hynix proposed is a 3D integrated structure that stacks memory directly on top of GPUs or logic dies. Vertically connecting logic and memory through hybrid bonding can dramatically increase the number of channels linking the two, boosting bandwidth even within limited area. However, the challenge remains that heat generated by logic directly affects the memory stacked above it.

“Placing memory on top of a GPU—one of the hottest chips in the world—is not easy,” Joo said. “Not only memory, logic, and GPU developers, but also server and infrastructure companies must work together on solutions starting from the cooling and system design stages.”

Beyond 3D stacking, SK Hynix also presented processing-in-memory (PIM) and Compute Express Link (CXL)-based memory pooling as core technologies for next-generation AI infrastructure. PIM combines storage and compute functions to reduce power consumption during data movement, while memory pooling reduces idle time when high-performance GPUs wait due to data transfer latency, improving AI infrastructure operational efficiency.

At the forum, Dell also identified cost efficiency amid agentic AI proliferation as a key challenge. Yoo Sang-mo, President of Dell Technologies Korea, said in his keynote: “61% of enterprise leaders feel they need to integrate AI into the workplace, but 64% still lack a clear execution roadmap.” Matt Dunphy, Senior Vice President at Dell, emphasized that a hybrid infrastructure strategy—combining on-premises and cloud based on cost efficiency—is essential.

An SK Hynix representative said, “Through this event, we further solidified our partnership with Dell, a key partner, and demonstrated our technological capabilities as a full-stack AI memory creator. We will continue to expand strategic partnerships with major global big tech companies to strengthen our leadership position in the AI memory market.”

NAND product planning executives also introduced eSSD products for AI data centers and next-generation cSSD technologies for on-device AI, along with relevant use cases. SK Hynix plans to expand collaboration with Dell from R&D to business operations based on their long-term strategic partnership.