SK Hynix said it is adjusting the timing of a shift to next-generation memory technology to match supply conditions and economic feasibility. It is increasing near-term supply with proven processes while preparing for the next wave of demand with customised HBM and memory-tiering technology. It said AI inference is spreading, pushing memory demand above supply, and a structure of splitting data across HBM, DRAM and SSD is taking hold.

In a third-quarter earnings conference call held on Oct. 28, SK Hynix said some customers are moving to reduce the number of HBM stacks as a result of supply shortages. Kim Ki-tae (김기태), head of HBM sales and marketing at SK Hynix, said, "Recent moves by some customers to adjust HBM specifications are a representative example showing that demand is overwhelming supply." He said customer requested volumes are more than double what the company can supply.

Reducing the number of stacks makes it possible to secure more HBM bundles, or cubes, for the same capacity and produce more AI accelerators. The company said even with fewer stacks, capacity per die rises and can offset much of the capacity reduction. It said its mid- to long-term outlook for HBM demand is unchanged.

The demand pressure is also spreading beyond HBM. It said agentic AI and longer-context processing are increasing, and KV cache, context data needed for AI inference, is growing rapidly. Song Chang-seok (송창석), vice president in charge of NAND marketing at SK Hynix, said, "There are limits in capacity and cost to storing all KV cache in HBM and DRAM." He said a tiered memory structure that stores data across HBM, DRAM and SSD depending on access frequency is becoming more important.

SSDs are already being used in some AI systems to move KV cache. Song said its role is expected to expand into a context memory tier that shares large-scale KV cache across multiple servers and GPUs.

SK Hynix plans to respond first by expanding existing processes to meet such demand. While latecomers are pushing new cell structures and next-generation DRAM development, the company plans to make the most of its current 6F² structure before moving to 4F² or 3D DRAM. 4F² is a next-generation DRAM technology that raises integration density by reducing cell area. The company said it has secured technical leeway to further expand the 6F² structure through extreme ultraviolet, or EUV, process innovation. Park Jun-deok (박준덕), head of DRAM marketing at SK Hynix, said, "At this point, the most efficient strategy is to maximise the use of platforms with proven mass-production stability and cost competitiveness to enhance product competitiveness."

SK Hynix expands supply with proven DRAM structures and broadens AI memory tiers.

The company views technology development and application to mass production as separate timelines. It said a new structure must meet not only performance goals but also yield, reliability and cost requirements, and it must undergo customer evaluation. Park said, "There is a need to distinguish between the realisation of a technology and the point of actual commercialisation." He said the company plans to apply each technology to suitable products when its completeness and economic feasibility are secured. The company's roadmap moves step by step from extreme scaling of existing DRAM to 4F² and 3D DRAM. HBM is following the same sequence. It said it completed supplying HBM4E samples in the second quarter and development and evaluation are under way to match customer schedules.

The next competitive market is customer-by-customer design. Kim said, "Along with stable supply of standard HBM, custom HBM optimised for each customer's system and workload will also become more important." He said the company will continue HBM growth based on proven product competitiveness and stable mass-production capability as a first supplier to key customers.

In NAND, it is developing software technology called SALT KV that determines placement across HBM, DRAM and SSD by weighing the reuse value of KV cache and storage costs. It is also developing CMM Hybrid, a CXL-based hybrid memory technology. Over the mid to long term, it is targeting a layer between HBM and SSD with HBF that combines NAND with high bandwidth.

Kim said the impact of changes in stack counts on HBM profitability remains a variable. He said, "Profitability is influenced together by transaction terms by product, product mix, yield, back-end process productivity and costs, so it cannot be judged by stack counts alone." If stack counts fall, more cubes are needed for the same volume, and base-die and back-end processing requirements increase.

The company said it has already reflected this in its production and supply plans. It said it has also completed consultations with key customers on next year's HBM volume and pricing, and expects the profitability gap between HBM and standard DRAM to narrow. It said next year a higher share of HBM4 and higher prices, along with yield and cost improvements, are expected to lift HBM revenue and profit from this year, and profitability above past cycle peaks is expected to continue.