When Chey Tae-won, chairman of South Korean conglomerate SK Group, speaks about semiconductors, he does so with the authority of someone who controls one of the world’s largest memory chip producers, SK hynix. But his latest warning – that memory has become a national security priority – is not a mere public relations exercise. It signals a tectonic shift that directly affects the artificial intelligence ecosystem, and on-premise infrastructure in particular.

The reason is brutally simple: without fast memory, workloads revolving around Large Language Models collapse. In recent years we have witnessed a frantic race for compute power, but the real bottleneck for inference and training is increasingly shifting to memory – bandwidth, capacity, physical proximity to the silicon. High Bandwidth Memory (HBM) is now an indispensable ingredient for high-end GPUs, while VRAM determines the size of models that can be hosted without resorting to aggressive quantization or costly distributed architectures. Those designing on-premise deployments know this well: ordering a server without first mapping the memory supply chain is like building a house without foundations.

This is where geopolitical geometry comes into play. Today, virtually all DRAM and advanced memories come from three players: Samsung, SK hynix, and Micron. In a world riven by trade tensions and aggressive industrial policies, this concentration becomes a systemic risk. A single factory earthquake, an escalation of tariffs, or fresh export restrictions can sever the pipeline that sustains data centers and private installations. This is not a theoretical scenario: companies already experienced it during the GPU shortage, and the same script now threatens to play out for memory.

The SK chairman’s warning is therefore not an isolated alarm but a reflection of a structural transformation. Governments are beginning to treat the ability to manufacture memory chips as an attribute of sovereignty, on par with energy reserves or cyber defense. The US CHIPS Act and the European Chips Act are attempts to bring production back onto domestic soil, but the timelines to build fabs are measured in years, while bottlenecks are already here. For organizations handling sensitive data that choose on-premise to keep data residency under control, this introduces a new layer of fragility: it is no longer enough to protect model access; you must guarantee the physical continuity of components.

Consequently, IT decision-makers will have to learn to evaluate the total cost of ownership (TCO) of an AI infrastructure not just in euros per FLOP, but also in terms of supply chain resilience. Suppliers that can diversify sourcing or lock in production in regions deemed secure will gain a lasting competitive advantage. At the same time, the case strengthens for modular on-premise systems that can adapt to different memory generations without full node replacement, and for architectures that account for the decreasing predictability of deliveries.

Ultimately, SK’s call certifies a truth many industry practitioners already sensed: memory is no longer a commodity. It is a geopolitical asset that determines which models can be trained, where, and with what degree of independence. In the age of Large Language Models, control over the silicon raw material that stores model weights has become a strategic game. And the next moves – from fabs under construction to stockpiling policies – will be watched with the same attention reserved for military maneuvers.