At a recent conference, SK hynix vice-president Kim Cheon-seong revealed that his company had entered into a partnership with Nvidia to create a new generation of SSDs. The aim is to increase the speed of these future SSDs by a factor of 10. Development is still at the proof-of-concept stage, with a prototype planned for late 2026.
To take the figures a step further, Kim points out that the aim is for this new generation to turbo-charge at 100 million input/output operations per second (IOPS), a much higher throughput than conventional enterprise SSDs. Reaching such a scale would represent a major architectural breakthrough, bridging the gap between memory and storage in AI infrastructures. Because, as you may have guessed, it’s all about AI. The main reason lies in the way current models work. It is imperative to have uninterrupted access to a wide variety of model parameters, a requirement that traditional HBM or DRAM memories cannot satisfy, or at least not under satisfactory economic and efficiency conditions. This new generation of SSDs will offer a kind of pseudo-memory, perfectly optimized for AI workloads.
Super SSD = Super shortage?
It’s all very exciting, of course, but in the end, there’s another problem on the horizon that won’t make things any easier for us users. Growing demand for AI-related hardware is already putting considerable pressure on NAND memory supply chains. This situation is well known, but it has been partially overshadowed by concerns about “classic” RAM. The widespread use of SSDs specifically designed for artificial intelligence could exacerbate these pressures. Industry specialists warn of the risk of a supply crisis comparable to the one we are currently experiencing with DRAM, should these high-efficiency NAND technologies be adopted for AI-related uses. Unless, of course, the AI bubble bursts in the meantime…
