SK Hynix, South Korea's leading memory maker, is tightening its strategic ties with graphics giant Nvidia by developing a high-performance solid-state drive (SSD) optimized specifically for AI inference workloads. The collaboration marks a significant expansion of a partnership that previously centered on supplying high-bandwidth memory (HBM) for Nvidia's AI GPUs. This time, the two companies are pushing hard into NAND flash storage innovation, with the goal of building an AI SSD that is 10 times faster than today's drives.
The AI SSD Performance Goal and the 100 Million IOPS Milestone
At a recent tech conference, SK Hynix Vice President Kim Cheon-seong made an announcement that stirred the industry. According to him, the Korean memory maker is working actively with Nvidia to deliver a tenfold leap in SSD performance. Citing Korea's Chosun newspaper, Kim said his company is developing a new type of SSD that is 10 times faster, in tandem with Nvidia.
The two companies call the new concept by different internal codenames: "Storage Next" at Nvidia and "AI-NP," short for AI NAND Performance, at SK Hynix. Development is currently at the proof-of-concept stage, with a prototype expected to be finished before the end of 2026. It is an ambitious timeline, but it also shows how seriously both sides are taking the effort to redefine what storage can do.
The technical target SK Hynix has set for this next-generation AI SSD is 100 million input/output operations per second (IOPS). To put that in perspective, the figure is far higher than the throughput of typical enterprise SSDs today. Reaching that level of performance would be a major leap in hardware architecture, effectively narrowing the gap between memory (RAM) and storage in AI infrastructure.
Tackling Bottlenecks and Pressure on the Supply Chain
The push behind this collaboration comes from the data-access bottlenecks that current AI workloads run into. Today's large-scale inference models depend on constantly retrieving an enormous number of model parameters. Traditional memory technologies such as HBM (High Bandwidth Memory) or DRAM cannot support that task efficiently at scale and cost. The vision for "Storage Next" is to enable a "pseudo-memory layer" built on NAND flash chips and advanced controller technologies designed specifically for AI computing rather than for ordinary data storage alone.
The joint project also hints at potential ripple effects across the components market. The NAND supply chain is already under heavy pressure from growing demand for cloud and AI services, and a dedicated AI SSD could add to that strain. Industry observers have raised the possibility of a supply crisis similar to the one that hit DRAM if high-performance NAND solutions like this become common in mainstream AI applications.
Both SK Hynix and Nvidia appear to be sharply focused on overcoming challenges in throughput and energy efficiency, which are becoming the defining limits of current-generation AI infrastructure. By integrating more advanced controller and NAND architectures, the two companies are positioning flash storage to play a more active computational role in machine learning workloads, something traditional storage cannot achieve economically.
