NVIDIA and SK Group unveil $500B+ AI initiative with HBM4 and 2GW Korean data center

NVIDIA and SK Group unveiled one of the largest compute-investment commitments of the year: a $500 billion-plus AI initiative anchored by a long-term SK Hynix partnership to jointly develop high-bandwidth memory (HBM) for AI training and agentic workloads. The flagship project is a 2-gigawatt AI data center from SK Telecom, powered by NVIDIA's Vera Rubin chips and SK Hynix's HBM4 memory, targeted to come online in 2027. NVIDIA and Naver separately plan to expand Naver's Korean AI data center.

The deal mechanics tie together the full stack — NVIDIA compute, SK Hynix memory, and Korean power/data-center capacity — and reflect the week's clear theme of massive compute-and-memory buildouts (echoing the NVIDIA/SK, Microsoft/Mistral European, and Alphabet's $195–205B capex figures). HBM4 co-development is particularly strategic: memory bandwidth, not just GPU count, is increasingly the bottleneck for training and long-context agentic inference.

Developer forums gave the announcement limited vocal reaction during the window, though engineers noted memory-supply implications — the more HBM committed to hyperscale AI, the tighter supply for everyone else. The scale also reinforces NVIDIA's near-record stock and its central position in the AI-infrastructure trade. What to watch: whether the 2027 timeline holds, how HBM4 supply is allocated, and whether Korea emerges as a major sovereign-AI compute hub alongside the European buildouts.