SK Group and NVIDIA outline $500 billion-plus AI infrastructure and memory initiative

SK Group and NVIDIA have outlined a proposed partnership valued at more than 500 billion dollars that spans AI-factory construction and next-generation memory. The companies signed letters of intent covering SK Telecom's plan for a two-gigawatt NVIDIA Vera Rubin DSX AI factory in Korea and a long-term agreement between SK hynix and NVIDIA to secure and develop memory, including high-bandwidth memory used by modern AI accelerators. The first factory is planned to come online in 2027.

The announcement links two constraints that have become central to AI infrastructure: compute capacity and the memory needed to keep accelerators supplied with data. Training and serving large models require thousands of processors connected by fast networks, but performance can be limited when memory bandwidth is insufficient. HBM packages sit close to accelerators and provide much higher bandwidth than conventional memory, making them a critical component of data-centre systems even though they are more difficult and expensive to manufacture.

SK Telecom's planned facility would use the NVIDIA DSX full-stack architecture, combining accelerated computing, systems, software and partner technologies. The companies say it is intended to support sovereign, physical, agentic and enterprise AI services and to address demand across the Asia-Pacific region. By offering infrastructure through a cloud, SK Telecom could spread the cost of a large build-out across more customers instead of requiring each business to purchase and operate its own accelerator cluster.

The memory partnership is designed to provide a more stable supply of future HBM generations while giving SK hynix a role in co-developing products for changing AI workloads. The companies say they will optimise memory solutions for large-language-model training, agentic systems and physical AI. In practice, that kind of cooperation can include early technical planning, packaging requirements and software tuning, although the announcement does not provide a detailed production schedule or a final commercial contract.

The size of the headline figure should be treated carefully. The 500-billion-plus initiative describes the broad scope of a strategic plan and letters of intent, not a single cheque paid on the announcement date. The factory still requires power, construction, equipment deliveries, network capacity and customers. Memory commitments also have to be matched with manufacturing yields and demand, while the performance of a new AI system depends on how well its software uses the hardware.

For Korea, the partnership is another attempt to move from being a leading buyer of AI hardware to hosting infrastructure and developing the components that make it possible. SK Group brings a large semiconductor and telecommunications footprint, while NVIDIA supplies the computing platform and software stack. Their combined position could help local companies access capacity and give Korean memory technology a more direct connection to the design of future AI systems.

The next evidence will be financing, construction and the delivery of the first Vera Rubin systems, followed by details about HBM generations and customer workloads. If those steps proceed, the partnership could become one of the most visible examples of AI factories being planned around both compute and memory supply. If schedules or demand change, the letters of intent may be revised. For now, the announcement shows how the AI infrastructure race is spreading across data centres, chips and the supply chains that connect them.