The letters of intent were signed at an AI summit in San Francisco on July 24, but the groundwork took months of quiet meetings between two families of companies that already sit inside each other’s supply chains. Nvidia and South Korea’s SK Group announced a partnership valued at more than $500 billion, a package that stretches from gigawatt-scale data centers to the high-bandwidth memory chips that have become the most contested component in modern computing.
CNBC described the agreement as the largest single collaboration in AI infrastructure history. The scope is hard to argue with. SK Telecom, the group’s carrier arm, will build a 2-gigawatt AI cloud in South Korea on Nvidia’s DSX full-stack platform, running the chip maker’s Vera Rubin accelerated computing line with SK Hynix HBM4 memory. The first phase of that AI factory is scheduled to come online in 2027.
For Nvidia, the memory terms may carry more weight than the construction plans. The company has locked in long-term supply of next-generation high-bandwidth memory from SK Hynix, the dominant producer in a segment where capacity, not demand, sets the pace. Nvidia and SK Hynix will also co-develop memory stacks tuned to the bandwidth, power and thermal demands of large language model training, agentic AI and physical AI systems.
The arrangement addresses the industry’s sharpest bottleneck. Every advanced accelerator sold in the past two years has required a stack of HBM, and memory makers have struggled to add capacity fast enough to match the data center build-out. SK Hynix began shipping HBM4 this year, and the supply agreement gives Nvidia a claim on that output for the Rubin generation and beyond.
“South Korea has all the ingredients to become a global AI powerhouse — world-class networks and data centers, leadership in chip technology and vast industrial scale,” Jensen Huang, Nvidia’s chief executive, said in a statement. SK Group Chairman Chey Tae-won framed the deal as industrial strategy: “In the AI era, competitiveness depends not just on how effectively AI is utilized, but on how much intelligence we can produce.”
The pact also extends Nvidia’s push into sovereign AI, the effort by national governments to build computing capacity inside their own borders. SK Telecom’s cloud is aimed at demand across the Asia-Pacific region, and Nvidia separately said it and partners including Naver and Brookfield plan to expand Naver’s data center in South Korea.
Analysts said the structure suits both sides. Nvidia secures memory at a moment when HBM pricing has climbed and allocation has become a boardroom topic at every large data center operator. SK Hynix gains the revenue certainty of a long-term contract and a development partner for its next memory generations, a cushion in a famously cyclical business.
The deal deepens a relationship that already ran decades deep. SK Group’s conglomerate structure — chips, telecom, energy — gives Nvidia a single counterparty that can build, power and supply a data center. Few companies in the world can make that claim, which is why the two sides have kept expanding the scope of their cooperation as the AI build-out accelerates.
For the broader industry, the agreement signals that the AI expansion has entered a procurement phase: not whether to build, but how to secure the components. Samsung Electronics and Micron Technology are racing to expand HBM capacity alongside SK Hynix, and memory prices have recovered sharply after a brutal downturn in 2023. SK Hynix’s shares have climbed this year as investors priced in a multiyear upcycle.
Nvidia said the letters of intent formalize work that has been underway for years, including the recently announced plan for SK Telecom to build an AI cloud in Korea. The companies declined to disclose financing details for the individual projects, a person familiar with the matter said.
The scale of the announcement raises questions that will play out over years. Gigawatt-scale data centers require power deals, grid upgrades and environmental approvals that no letter of intent can guarantee. Analysts said the risk is real but manageable: the projects will be built in phases, and the memory portion of the deal will move fastest because it relies on existing factories.
Chey offered the clearest statement of intent. “By leveraging SK Hynix’s AI memory and SK Telecom’s AI infrastructure capabilities, SK will collaborate with NVIDIA to build a world-class AI factory, helping Korea transcend its role as a leading adopter of AI and become a global hub that drives AI innovation,” he said.
The phrase “global hub” is doing heavy work. South Korea already hosts two of the world’s three HBM suppliers and the largest memory fabs on earth. What it lacks is a homegrown market for that output. The Nvidia deal, if it delivers on its numbers, gives Korea something it has never had: a buyer of national scale committed to building there.
For Nvidia, the agreement is part of a pattern of supplier diplomacy. The company has struck similar infrastructure deals across the Middle East, Southeast Asia and Europe, tying its chips to national AI plans. Each deal locks in demand for Rubin and its successors while spreading the cost of the build-out across governments and conglomerates.
None of the previous pacts carried a $500 billion price tag. If the number holds, it will define the scale of the current cycle. The first test comes in 2027, when the first gigawatt of SK Telecom’s cloud is supposed to switch on. Between now and then, the hardest work — securing power, hiring thousands of engineers, ramping HBM4 yields — has only just begun.
People familiar with the discussions said the two sides expect to sign binding agreements for individual projects in the coming months. Investors will be watching the same dates, because in the memory business, a committed buyer changes everything.
