SK hynix’s AI memory roadmap: HBM5, AI DRAM, AI NAND and what it means for GPUs

Fresh off record quarterly profits, SK hynix has laid out a long-range memory roadmap that doubles down on HBM and AI-oriented DRAM and NAND. At its SK AI Summit 2025, the company talked up future HBM5/HBM5E, “AI DRAM,” “AI NAND,” GDDR7-next, DDR6, and 400+ layer 4D NAND stretching into the 2029–2031 window. The message is blunt: if AI is the new gold rush, SK hynix wants to be the dominant shovel merchant.

HBM: the main event

HBM is the star of SK hynix’s roadmap for obvious reasons: it is the memory attached to Nvidia’s and AMD’s highest-margin accelerators. SK hynix already holds a strong share in current HBM generations, and the roadmap promises:

HBM4 ramp and optimisation in the nearer term, feeding GB200/GB300-class and next-gen AMD accelerators.
HBM5 and HBM5E in the 2029–2031 timeframe, including custom HBM5 variants co-designed with large AI customers and GPU vendors.
Higher stacks and bandwidth enabled by better TSV processes and packaging, plus power-efficiency gains to keep thermal envelopes manageable.

Custom HBM is the key phrase. If SK hynix can tune stack capacity, bandwidth, and power characteristics to specific accelerator designs, it can lock in long-term contracts and reduce its exposure to spot pricing swings.

AI DRAM: branding or something more?

“AI DRAM” is a marketing-friendly label, but there are real engineering levers underneath it. For AI training and inference, the priorities are bandwidth per watt, predictable latency, and capacity scaling within sensible module envelopes.

LPDDR6 and GDDR7-next: Higher pin speeds for client and accelerator-adjacent memory.
DDR6 and potential 3D DRAM: Denser server memory with better rank utilisation for AI training clusters.
Signal integrity and power delivery work: Essential as pin rates and channel counts rise.

The “AI” label mainly reflects binning and productisation: parts that hit aggressive power/performance targets for AI rather than general server workloads. Expect SK hynix to offer AI-focused SKUs with guaranteed throughput and thermals for specific accelerator partners.

AI NAND and 4D NAND

On the NAND side, SK hynix is talking about 400+ layer 4D NAND around 2029–2031 and “AI NAND” products in the 2026–2028 window. In practice this means:

Very high-density SSDs for AI training datasets and retrieval-augmented generation stores.
Controller-level tuning for AI workloads—optimising write patterns, QoS, and read latency for large sequential and semi-random access.
PCIe Gen5/Gen6 SSDs targeting high-QPS AI storage tiers rather than general enterprise use.

For system designers, the headline is simple: storage capacity and throughput are being dragged forward to keep up with accelerators that can ingest more data per unit time than any previous generation.

How this affects GPU and accelerator design

A credible HBM5/HBM5E roadmap with custom options matters as much to Nvidia and AMD as any one GPU announcement. The constraints on future accelerators are increasingly packaging and memory, not raw compute. If SK hynix can deliver:

Higher-capacity HBM stacks without blowing power budgets.
Predictable long-term pricing for large AI customers.
Flexible configurations co-designed with GPU roadmaps.

then GPU design teams gain room to push die sizes, tile counts, and interconnect complexity, knowing that the memory subsystem will not be the bottleneck.

Risks and pressure points

Execution risk: HBM yields and 4D NAND layer counts are both hard problems. Missing timelines knocks on into GPU and server product launches.
Customer concentration: A handful of hyperscalers and GPU vendors account for a large fraction of demand; any shift in their strategies hits SK hynix directly.
Competitive response: Samsung and Micron will not leave the “AI DRAM/NAND” label uncontested; expect similar roadmaps and aggressive capex.

Editor’s take

SK hynix’s roadmap doesn’t rewrite physics, but it does spell out how the company plans to monetise the AI wave: lean into HBM and AI-flavoured DRAM/NAND and sell shovels to every prospector willing to sign a long-term contract. The winners on the compute side will come and go; the memory vendors plan to get paid either way.