Late on Friday in San Francisco, Nvidia and South Korea's SK Group signed a set of letters of intent that, added together, they valued at more than 500 billion dollars. It is one of the largest single commitments to AI infrastructure anyone has announced, and it was unveiled with South Korean President Lee Jae Myung in the room. The number is the headline. What sits underneath it is the more interesting part: a plan to turn most of a country into a single, vertically integrated AI factory.
The centrepiece is a two-gigawatt data centre that SK Telecom will build on Nvidia's DSX "AI factory" platform, running the company's next-generation Vera Rubin systems fed by SK Hynix's HBM4 memory. Two gigawatts is not a figure of speech; it is roughly the output of two large nuclear reactors, enough to power hundreds of thousands of GPUs. The first facility is due online in 2027. Around it sit a long-term memory supply and co-development pact with SK Hynix, a separate billion-dollar Nvidia investment in the Korean cloud firm Naver, and a further Naver data-centre expansion backed by Canada's Brookfield.
The half-trillion figure deserves a footnote. As Tom's Hardware points out, nobody is writing a 500 billion dollar cheque. The number bundles years of expected commercial activity under one banner: SK Telecom buying Nvidia hardware, SK Hynix selling memory back to Nvidia, the makers of cooling, networking and power that fill an AI factory, and vaguely defined future expansion. Read plainly, it is less a deal than a decade-long relationship handed a round, attention-grabbing price tag.
Strip the number away and the strategy is clear. South Korea already houses two of the world's three big memory makers, Samsung and SK Hynix. Bolt multi-gigawatt compute onto that base and the country has something very few can match: memory, chips and cloud capacity stitched into one national stack, aimed at what the companies call sovereign, enterprise, physical and agentic AI. Nvidia has run this play before, signing comparable AI-factory deals in Japan, the United Arab Emirates and Saudi Arabia. Each one ties a country's ambitions to Nvidia's roadmap.
There is a neat symmetry in it for Nvidia. In SK Telecom it gains a giant customer; in SK Hynix it secures a critical supplier, at a moment when high-bandwidth memory is the scarcest ingredient in the industry. The same weekend, Samsung answered with a 200 billion dollar memory and foundry pact with Broadcom, a sign that everyone can now see where the bottleneck sits. For Seoul, the prize is genuine: a shot at being a compute superpower rather than just a components exporter. The open question is how sovereign a national AI stack can really be when its beating heart, the accelerators and the software that runs them, is still designed 9,000 kilometres away in Santa Clara.