NVIDIA signs $500B memory and AI infrastructure deal with SK Hynix

NVIDIA has secured a multi-year high-bandwidth memory (HBM) supply and co-development deal with SK Hynix as part of a broader $500 billion-plus AI infrastructure initiative with SK Group. The agreement, announced at an AI summit in San Francisco, aims to lock in the critical memory component for NVIDIA's next-generation GPUs and Vera Rubin systems while building large-scale data centers expected to come online in 2027. For AI builders, this deal signals that memory supply constraints are becoming a first-order infrastructure concern, and that global AI deployment is moving beyond a handful of hyperscalers to involve large conglomerates and foreign governments.

What happened

NVIDIA and SK Group unveiled a comprehensive partnership that spans memory supply, co-development, and data-center construction. The centerpiece is a long-term agreement with SK Hynix to secure and co-develop next-generation AI memory, including HBM for NVIDIA's GPUs. SK Hynix is the leader in HBM production, according to analysts. The deal could be worth approximately $500 billion over multiple years.

Separately, NVIDIA committed $1 billion to Naver, a Korean cloud company, to build data centers around its GPUs. SK Telecom, an SK Group affiliate, will build a cloud business using NVIDIA's Vera Rubin systems. NVIDIA said it is targeting enough capacity to require 2 gigawatts of power, indicating a massive buildout with hundreds of thousands of GPUs.

Why AI builders should care

Memory has become the toughest bottleneck in AI infrastructure. SK Hynix CEO Kwak Noh-jung has said the global memory industry is heading for its worst-ever supply shortage in 2027, with demand outstripping supply well into the next decade. This deal is NVIDIA's attempt to buy certainty.

For teams building AI products, this matters in three ways:

• Memory pricing and availability: Long-term supply agreements could stabilize HBM pricing for NVIDIA's customers, but they may also tighten supply for other GPU makers and downstream markets like smartphones.
• Infrastructure diversification: The deal signals that AI infrastructure is expanding beyond hyperscalers like AWS, Azure, and GCP. Foreign governments and large conglomerates are now directly involved, which could create new regional capacity options.
• Vera Rubin timeline: The data centers are expected online in 2027, giving builders a clearer timeline for next-generation NVIDIA hardware availability.

Practical implications

The co-development component means SK Hynix and NVIDIA will work together on next-generation AI memory, likely HBM4 and beyond. This could lead to memory that is more tightly integrated with NVIDIA's architecture, potentially improving performance per watt for inference and training workloads.

The 2 GW power footprint is enormous. For context, a single large data center typically uses 50-100 MW. This implies a multi-site deployment that could reshape power availability in South Korea and influence where AI compute capacity is located.

The Naver investment provides a path for customers in South Korea and globally to secure AI computing capacity before the full 200 MW of capacity is completed. This could be a model for other regional cloud providers looking to offer NVIDIA GPU capacity.

Caveats

The $500 billion figure is a strategic target, not a fixed contract value. Timelines and execution details may change. Many specifics, including pricing, contract lengths, and precise capacity allocations, have not been disclosed. The deal is also subject to regulatory review in multiple jurisdictions.

SK Hynix listed on the Nasdaq earlier in July 2026 as part of an effort to finance infrastructure developments, which adds a public-market dimension to the partnership's execution risk.