Samsung and SK hynix Pivot to AI Storage to Break the 'Memory Wall'

Samsung and SK hynix are expanding their AI memory strategies beyond High Bandwidth Memory (HBM) to target the broader AI data center storage market. This strategic shift aims to resolve critical data bottlenecks that currently hinder the performance of massive AI workloads.

Both companies are diversifying their portfolios to include AI-optimized data center storage, with significant investments in next-generation technologies such as Compute Express Link (CXL) and Processing-in-Memory (PIM). These solutions integrate memory and storage more efficiently, reducing the latency and power consumption associated with moving vast amounts of data between processors and storage units. Specifically, SK hynix is scaling its AI-NAND strategy, including enterprise SSDs (eSSDs), to manage the immense data volumes required for AI inference.

The Shift Beyond HBM

For several years, HBM has been the primary driver for AI accelerators, most notably NVIDIA GPUs, due to its extreme bandwidth. However, as AI models grow in scale and complexity, the industry has encountered the "memory wall"—a performance ceiling where processor speed far outstrips the ability of memory and storage to provide data. While HBM solves the immediate bandwidth need for the GPU, it does not address the broader capacity and latency issues found across the entire data center architecture.

Solving the Memory Wall

This transition toward a comprehensive storage strategy indicates a fundamental shift in how AI infrastructure is built. By implementing CXL, which allows for more flexible memory pooling and expansion, and PIM, which performs computations directly within the memory chip, Samsung and SK hynix are attempting to optimize the entire data path. This approach reduces the distance data must travel, thereby lowering energy costs and increasing processing speeds.

Industry Implications

Maintaining dominance in the AI supply chain now requires more than just providing the fastest memory; it requires solving the systemic bottlenecks of the data center. By controlling both the high-speed HBM layer and the broader AI-optimized storage layer, these companies can offer a full-stack memory solution. This prevents a scenario where AI progress is stalled not by the intelligence of the models, but by the physical limitations of the hardware storing the data.

Future Outlook

Industry observers are now watching for the widespread adoption of CXL-based memory modules in commercial data centers. While the technical foundations of PIM and CXL are established, the next phase will be the integration of these technologies into standard server architectures. The success of this pivot will depend on how quickly cloud service providers adopt these new standards to support the next generation of generative AI.