NVIDIA and SK Group announced a $500 billion-plus strategic partnership on July 25, 2026, covering AI factory construction, accelerated computing, and next-generation memory co-development. The announcement was made during NVIDIA CEO Jensen Huang’s visit to South Korea and represents one of the largest single AI infrastructure commitments in history.
The deal is not just a chip order. It is a full-stack agreement that ties together SK Telecom building a massive AI cloud, SK hynix supplying the memory that will power the next generation of NVIDIA GPUs, and both companies co-developing AI infrastructure across Asia-Pacific.
What Was Actually Announced
The partnership has two main pillars.
The AI Factory: SK Telecom will build a two-gigawatt AI factory in South Korea built on NVIDIA’s DSX platform. DSX is NVIDIA’s full-stack AI factory architecture that integrates accelerated computing, systems, and software to deliver the lowest token cost at maximum energy efficiency. The factory will deploy NVIDIA Vera Rubin accelerated computing powered by SK hynix HBM4 memory, with the first facility planned to come online in 2027.
The Memory Partnership: NVIDIA and SK hynix signed a multi-year agreement to co-develop and optimize next-generation AI memory solutions, including HBM (high-bandwidth memory), to support evolving infrastructure needs from large language model training to agentic and physical AI. SK hynix is estimated to supply roughly 60 to 70 percent of HBM4 volume for NVIDIA’s Vera Rubin platform.
Together, the letters of intent spanning both pillars represent over $500 billion in commitments.
Why a 2-Gigawatt AI Factory Matters
For context, most current large-scale data centers consume between 50 and 200 megawatts. A two-gigawatt AI factory is an order of magnitude larger, built specifically to run AI inference and training at a scale that no current facility can match.
NVIDIA’s DSX architecture is designed around a single goal: reducing the cost per token. Every AI output a model generates costs compute. At current scale, token costs are one of the primary constraints limiting how broadly businesses can deploy AI agents. A factory like this, running next-gen Vera Rubin GPUs with purpose-built HBM4 memory, is built to push that cost down significantly.
Jensen Huang put the scale of the ambition plainly: “AI factories are the engines of the next industrial revolution, and advanced memory is essential to their performance. Together, we will codevelop the next generation of memory for AI factories and support the accelerating global expansion of AI infrastructure.”
HBM4 and Why the Memory Race Matters
High-bandwidth memory is the bottleneck in modern AI hardware. GPUs can only process as fast as data can move in and out of memory. HBM4 is the next standard, designed to move data faster and more efficiently than the current HBM3 used in NVIDIA’s H100 and H200 chips.
The multi-year co-development agreement between NVIDIA and SK hynix signals something important: the memory supply chain for next-generation AI compute is now being locked in years in advance. NVIDIA reportedly asked SK hynix to bring forward HBM4 chip supply by six months, a sign of how urgently the infrastructure race is moving.
The Sovereign AI Angle
The partnership has an explicit focus on sovereign, physical, agentic, and enterprise AI services across Asia-Pacific. The South Korean government and technology sector have made AI sovereignty a national priority, and this deal positions Korea as a significant node in global AI compute rather than a consumer of US or Chinese infrastructure.
Huang also signaled a broader bet on physical AI, noting that “the age of physical AI has finally arrived and no country is better prepared for robotics than Korea.” This points to the factory not just serving cloud AI workloads but also supplying the compute needed for robotic and agentic physical systems.
What This Means for Business
Deals of this scale tend to be viewed as abstract infrastructure news. But the consequences for businesses running on AI are direct.
Token costs will continue to fall. Every major AI factory being built at scale is competing to produce cheaper compute. When SK Telecom’s 2GW AI factory comes online in 2027, it adds significant supply-side pressure that makes inference cheaper for everyone building on models that run in the cloud.
Next-gen models are coming faster than the timeline suggests. The Vera Rubin platform plus HBM4 memory is not the current generation. It is two generations ahead of what most enterprises are running on today. The fact that supply chains for this hardware are being locked in now means the next step-change in model capability is closer than it appears.
Geographic diversification of AI compute is real. A two-gigawatt AI factory in South Korea adds a significant non-US compute hub to the global AI supply chain. For businesses worried about US regulatory exposure, data sovereignty, or geopolitical risk in their AI infrastructure, this is meaningful.
What Enterprise DNA Sees Here
The AI infrastructure race is not slowing down. If anything, the scale of commitments like this one shows that the major players are betting on sustained exponential demand for AI compute through at least the end of the decade.
For businesses still evaluating whether to commit to AI tools and workflows, this should settle one concern: the infrastructure will be there. The compute needed to run powerful AI agents across a business is being built at a scale that will make it broadly available and affordable within a few years.
The question is not whether AI infrastructure will be there. It is whether your business has the skills and workflows in place to take advantage of it when the costs reach the point where every team can run agents continuously.
That is the gap Enterprise DNA exists to close. Whether that means upskilling your data and analytics team, deploying an AI agent workforce across your operations, or getting a fractional AI advisor to map your roadmap before the compute economics tip further in your favor. Start that conversation here.