SK Hynix predicts AI infrastructure performance scale to grow 20-fold by 2030

SK Hynix's head of development, Ahn Hyun, predicted at the Korea Semiconductor Exhibition (SEDEX 2026) in Seoul that the performance scale of AI infrastructure is expected to grow approximately 20-fold by 2030. This forecast stems from the rapid expansion of the AI inference market across multiple dimensions.

These dimensions include increasing task complexity, a surge in the volume of information processed per instance (context length), and a significant increase in inference requests. These intertwined factors are driving a substantial growth in demand for AI computing infrastructure.

As AI inference needs escalate, the role of memory in storing and transferring vast amounts of data becomes increasingly critical. Specifically, the "Key-Value Cache" (KV Cache) is expanding, necessitating the distribution of data across various storage media such as High Bandwidth Memory (HBM), DRAM, and SSDs. The resulting data migration is becoming a new bottleneck, impacting performance and increasing system energy consumption.

In response, SK Hynix is developing next-generation storage technologies. The company is advancing 3D-stacked DRAM, which integrates DRAM directly on top of logic chips like GPUs and NPUs to enhance transfer performance and reduce power consumption. Additionally, High Bandwidth Flash (HBF) aims to apply HBM-like technologies to NAND flash, creating a new storage tier between HBM and DRAM. SK Hynix targets HBF to achieve around 50% of HBM's bandwidth and a tenfold increase in capacity.