SK Group chairman outlines pivot to memory as a service, signaling seismic shift for AI infrastructure

SK Group Chairman Chey Tae-won wants to stop just selling memory chips. He wants to sell memory like a cloud service.

The concept, called Memory-as-a-Service (MaaS), would fundamentally change how data centers consume one of their most critical resources. Instead of buying stacks of physical chips and bolting them into servers, customers would tap into pooled, composable memory infrastructure on demand.

From chips to subscriptions

The MaaS strategy leverages CXL (Compute Express Link) interconnect technology, which allows memory to be disaggregated from individual servers and shared across an entire data center. Rather than each server having its own dedicated memory that sits idle when not in use, CXL lets you create a giant shared pool that any server can draw from as needed.

Chey first signaled this strategic direction as early as 2022. The timing wasn’t accidental. AI workloads were beginning their exponential climb, and the memory demands of training large language models were exposing the limitations of traditional server architectures.

SK Hynix, the semiconductor arm of SK Group, is already the dominant force in high-bandwidth memory (HBM), which is the specialized memory that sits atop AI accelerator chips like Nvidia’s GPUs. Moving from selling those chips to offering memory infrastructure as an ongoing service would create recurring revenue streams.

A shortage that won’t quit

Global memory chip shortages, particularly for HBM, are projected to persist until at least 2030. Wafer production currently lags behind demand by more than 20%, and the gap isn’t closing fast enough.

SK Hynix is responding by committing to double its wafer production capacity over the next five years.

The company is also deepening its relationship with Nvidia through a collaboration to build an AI-focused factory in Korea. This facility would center on GPU-as-a-service models, essentially creating vertically integrated AI infrastructure that bundles compute and memory together as a managed offering.

What this means for the AI hardware supply chain

The shift toward MaaS and composable memory represents a broader industry trend where hardware companies are borrowing playbooks from software companies. Selling products is fine. Selling services built on those products is better.

The competitive landscape is also worth monitoring. Samsung and Micron, SK Hynix’s primary rivals in the memory market, will face pressure to develop their own MaaS offerings or risk being relegated to commodity chip suppliers while SK Hynix captures the higher-value service layer. Samsung has been investing heavily in its own HBM capabilities, but SK Hynix currently holds a significant lead in production volume and customer relationships, particularly with Nvidia.