Memory strategy expands from faster chips to system-level data management
SK hynix is extending its artificial intelligence memory strategy beyond high-bandwidth memory as the rapid growth of AI inference and agent-based services creates new demands for storing and moving data.
The company has recently put greater emphasis on technologies including High Bandwidth Flash (HBF), Processing-In-Memory (PIM) and SALT-KV. Rather than simply increasing HBM performance, the approach combines different types of memory and storage to improve the efficiency of AI systems.
SK hynix showcased the technologies at AI Infra Summit 2026 in Santa Clara, California. The event drew about 6,000 participants, while the company doubled the size of its exhibition space from the previous year.
AI inference changes the memory equation
Much of the AI semiconductor boom has so far revolved around GPUs and HBM. Training large language models requires enormous computing power, while HBM feeds large volumes of data to processors at high speed.
The rise of commercial AI services is adding another challenge.
As more users interact with AI models and conversations and tasks become longer, the amount of data that must be retained and retrieved increases sharply. Keeping all of that data in high-performance HBM can become costly and impractical from a capacity standpoint.
HBF is designed to fill part of that gap.
The technology applies through-silicon via, or TSV, technology used in HBM to NAND flash, seeking to combine greater storage capacity with higher bandwidth. It is intended to create a new memory tier between HBM and conventional solid-state storage.
That could become increasingly relevant for inference workloads where large amounts of data need to remain readily accessible but do not necessarily require HBM-level performance.
PIM tackles the cost of moving data
SK hynix is also advancing PIM, which integrates computing functions directly into memory.
In conventional computing architectures, data repeatedly moves between memory and processors. The larger AI workloads become, the more that movement can contribute to latency and power consumption.
The company has developed AiM, a memory product with computing capabilities, as well as the AiMX accelerator card and servers equipped with AiMX. At the summit, SK hynix demonstrated large language model applications running on the technology.
PIM itself is not a new concept. Its relevance, however, is growing as AI inference places greater pressure on memory bandwidth and energy efficiency.
For SK hynix, it also represents a potential business opportunity beyond the HBM market that has driven much of the company's recent growth.
Deciding where AI data should reside
SALT-KV approaches the problem from the software side.
Large language models and AI agents store previously processed information in what is known as a KV cache. As conversations become longer and AI agents carry out increasingly complex sequences of tasks, the amount of cached data can grow substantially.
SALT-KV distributes that data across HBM, DRAM and SSDs based on factors such as its likelihood of being reused and the cost of storing it.
Frequently accessed data can remain in faster memory, while less frequently used information can be shifted to larger and less expensive storage.
SK hynix demonstrated the technology at the summit with an LLM generating and running a game.
The project is notable because it takes the company's development work beyond memory hardware itself. SK hynix is increasingly looking at how different memory products are actually deployed within an AI system.
Preparing for the market beyond HBM
The direction is also reflected in SK hynix's broader AI strategy.
CEO Kwak Noh-jung said at the company's 2026 Global Forum in the United States that SK hynix intends to open new possibilities for AI with next-generation memory, building on the position it established in HBM.
Discussions at the forum covered memory design, process and device technology, advanced packaging and AI systems, with greater emphasis on connecting those technologies at the system level.
The company's investment footprint is moving in a similar direction. SK hynix is developing its semiconductor cluster in Yongin, South Korea, while building an advanced packaging production and R&D base for AI memory in Indiana in the United States.
None of this suggests that HBM is becoming less important. Demand for high-performance HBM is expected to remain central as AI accelerators continue to increase in computing power.
But the expansion of AI from model training into inference and autonomous agents is creating memory requirements that cannot be addressed by HBM alone.
HBF, PIM and SALT-KV show where SK hynix is looking next: maintaining its position in HBM while pursuing new opportunities emerging around the increasingly complex memory architecture of AI systems.
