Introduction
The headline figure is hard to ignore: more than $500 billion committed between NVIDIA and South Korean conglomerate SK Group, under a deal announced on July 24, 2026. Behind the eye-watering sum and the promise of a massive AI factory complex lies a far more technical — and far more consequential — story: high-bandwidth memory, specifically HBM4, is fast becoming the single component that determines the pace and scale of the entire global AI infrastructure buildout.
The GPU Alone Is No Longer Enough
For years, raw GPU compute dominated the conversation. But a hard physical constraint has moved to center stage: a GPU can only deliver on its performance potential if memory can feed it data fast enough. With NVIDIA's Vera Rubin architecture — with first deliveries to major cloud providers scheduled for Q3 2026 — that constraint has reached a critical inflection point.
Each Rubin GPU packs up to 288 GB of HBM4 memory, capable of delivering bandwidth approaching 22 terabytes per second. That is roughly 2.75 times more than the previous Blackwell generation. To hit those speeds, SK Hynix engineered a doubled memory interface — 2,048 bits versus 1,024 bits for HBM3E — running at pin speeds exceeding 6.4 Gbps. Without memory at this performance tier, the world's most powerful GPU is running on empty.
SK Hynix at the Center of the Value Chain
The NVIDIA–SK Hynix agreement goes well beyond a standard supply contract: it is a long-term co-development partnership spanning the next generation of AI memory, from large model training through to physical and agentic AI applications. SK Hynix currently supplies the majority of HBM4 volumes integrated into Vera Rubin systems, with Samsung and Micron rounding out the supplier landscape — all three qualified and in simultaneous production.
This supply chain concentration carries direct implications for hardware buyers and CIOs planning their infrastructure. South Korea's HBM4 production capacity is a hard constraint on delivery timelines for next-generation AI servers across every market worldwide.
A 2-Gigawatt Factory by 2027
The most striking element of the deal involves SK Telecom, which has committed to building an AI factory in South Korea with a capacity of 2 gigawatts, built on NVIDIA's DSX platform — the architecture that brings together Vera Rubin accelerators, NVLink networking, and the associated software stack into a unified system. The facility is scheduled to come online in 2027.
Two gigawatts is roughly equivalent to the electricity consumption of a mid-sized city. That single figure captures the shift in scale underway in the infrastructure race, and underscores a harder truth: the compute capacity required for next-generation AI is being locked in right now, through commitments that stretch well beyond the planning horizons most organizations currently have in place.
What This Means for IT Buyers
For IT teams and CIOs, the practical stakes are twofold. First, the concentration of HBM4 production across a small number of South Korean sites creates a geographic dependency on a component with no short-term alternative — a supply chain risk factor that belongs in business continuity planning. Second, the Vera Rubin systems being deployed at hyperscalers in 2026 and 2027 will define the cloud compute capacity available to enterprises that have no intention of running their own AI infrastructure.
Memory is no longer a passive commodity ordered from a catalog. It has become a strategic asset that industry giants are locking up through multi-year partnerships and reserved gigawatts. If you are mapping out your hardware roadmap for 2027–2028, the question of compute access is already being negotiated several layers upstream — and the window to influence your position is narrowing.