SK hynix Unveils Full-Stack AI Memory, 3D Memory Strategy

At the SK hynix Future Forum 2026 on September 8, the company's executives laid out a vision that goes well beyond shipping faster memory chips. CEO Kwak Noh-Jung had first announced the strategic reframing at the SK AI Summit in November 2025, telling employees and partners that the company would evolve from a full-stack AI memory provider into a full-stack AI memory creator. The Future Forum made clear what that word change means in practice: SK hynix no longer intends to sell discrete components to GPU makers — it intends to co-design the memory system as an integrated solution, from the logic base die inside an HBM stack to the SSDs that hold cold data in the training cluster.

Three pillars

The strategy rests on three pillars, each with its own product families.

Custom HBM is the first and most consequential. The idea is to move selected computing functions off the GPU or ASIC and onto the base die that sits at the bottom of every HBM stack. That base die is fabricated on a logic process — SK hynix has partnered with TSMC for 12 nm and 3 nm nodes — rather than on a conventional DRAM process. By pulling logic closer to the memory, the company aims to cut data-movement power, improve inference efficiency, and lower total cost of ownership for data-center operators.

The second pillar, AI DRAM (branded AI-D), splits into three tracks. AI-D Optimization covers low-power, high-performance modules such as MRDIMM, SOCAMM2, and LPDDR5R, targeting data-center TCO reduction. AI-D Breakthrough tackles the memory wall with ultra-high-capacity solutions including CXL memory modules and processing-in-memory. AI-D Expansion extends HBM and other DRAM technologies into robotics, automotive, and industrial automation — domains where memory requirements increasingly resemble those of AI training clusters.

The third pillar, AI NAND (AI-N), similarly subdivides. AI-N Performance focuses on ultra-high-IOPS SSDs for large-scale AI inference, with samples planned for late 2026. AI-N Bandwidth applies HBM-style stacked architecture to NAND flash — a category now known as HBF, or High Bandwidth Flash. AI-N Density targets PB-scale capacity using QLC technology, blending SSD speed with HDD economics.

3D Memory: erasing the boundary

The concept that ties these pillars together is what SK hynix calls 3D Memory. At the Future Forum, Vice President Euicheol Lim described it as the combination of 3D-stacked DRAM, HBM, and HBF into workload-optimized configurations. Korea University professor Changhwan Shin, speaking at the same event, offered the more provocative framing: 3D is not simply a technology for stacking chips vertically, but a technology that removes the distinction between memory and computing chips.

That blurring of boundaries is visible most concretely in the HBF initiative. Western Digital's SanDisk brand first proposed the concept in February 2025; SK hynix and SanDisk signed a memorandum of understanding in August 2025 to jointly develop it. By February 2026 the two companies launched a standardization effort, and in August 2026 the Open Compute Project published the first HBF specification. Read bandwidth is expected to reach 1.6–3.2 TB/s per stack, with single-stack capacity up to 512 GB — an order of magnitude beyond what HBM alone can offer, at a lower cost per bit.

Whether HBF can deliver on that promise in commercial products remains an open question. No HBF part has yet shipped to customers. But the speed of standardization — from concept to OCP spec in roughly 18 months — signals serious industry intent.

Market position and competitive pressure

SK hynix enters this strategic pivot from a position of strength. According to Counterpoint Research, the company held approximately 50 percent of the global HBM market by shipments in the second quarter of 2026, ahead of Samsung at 33 percent and Micron at 18 percent. The company's entire 2026 HBM output is sold out, with order books stretching into the first quarter of 2027 and key customer commitments locked through the end of 2028. SK hynix also supplies an estimated 70 percent of Nvidia's HBM4 volume.

Production milestones reinforce that lead. HBM4 in 12-layer configuration entered mass production in July 2026, with deliveries to Nvidia beginning the same month. The company's 321-layer 3D NAND is in volume production, with 375-layer planned for late 2026.

Yet the competitive landscape is shifting. Samsung's HBM share jumped from 21 percent in the first quarter of 2026 to 33 percent in the second — a 12-percentage-point gain in three months that narrowed the gap with SK hynix from 37 to 17 points. UBS has forecast that by 2027 the two companies could each hold roughly 40 percent of the market. Samsung's advantages as an integrated device manufacturer — logic, memory, foundry, and packaging under one roof — give it structural leverage that a pure memory player cannot easily replicate.

This is the competitive context that gives the Full-Stack strategy its urgency. If HBM becomes commoditized, margins will compress regardless of share. By moving up the value stack into custom logic, system-level optimization, and new categories like HBF, SK hynix is betting that the next phase of AI memory competition will be won not by who ships the most bits, but by who solves the customer's system problem most completely.

What remains unproven

Several elements of the strategy remain aspirational. Custom HBM requires deep co-engineering with each customer — a model that scales poorly if the number of differentiated AI ASIC designs proliferates. 3D-stacked DRAM as a product faces heat-dissipation and process-complexity challenges that SK hynix's own design-research vice president, Kyunghoon Kim, acknowledged at the forum. HBF has no shipping product. And the company's pre-announced U.S. advanced-packaging facility in Indiana will not begin production until the second half of 2028, leaving a multi-year window in which geopolitical supply-chain expectations outpace manufacturing reality.

The bet is real, and the timeline is long. SK hynix is effectively wagering that the AI industry's shift from a scale race to an efficiency race — as SK Group chairman Chey Tae-won put it at the AI Summit — will reward the company that treats memory not as a component but as an architecture.