SK Hynix, the SK Group subsidiary that supplies high-bandwidth memory (HBM) to Nvidia and other AI hardware makers, sits at the center of a supply chain that has become as strategically sensitive as the GPUs themselves. Chey's remarks frame AI memory not as a component shortage in the traditional sense but as a geopolitical chokepoint: governments negotiating for guaranteed allocation, treating access the way they might treat energy or rare-earth supply.
The 60-100% demand increase Chey cited for 2027 reflects the compounding effect of frontier model training runs growing larger and inference workloads scaling with adoption — both of which are memory-bound in ways that pure compute capacity additions don't fully address. His characterization that no company currently has meaningful new capacity coming online points to the multi-year lead times required to bring new memory fabrication capacity to market, meaning any shortfall identified today will not be resolved by 2027 regardless of investment decisions made now.
The comments arrive as memory pricing and allocation have already become a recurring flashpoint in AI hardware supply chains through 2026, with hyperscalers reportedly securing long-term supply commitments well in advance of need.