The Nvidia SK Hynix deal announced late Friday goes well beyond a memory purchase: it hands Nvidia the ability to decide who else gets memory at all.

The agreement, unveiled at an AI summit in San Francisco alongside South Korean President Lee Jae Myung, could be worth up to $500 billion over a number of years and includes new data centers coming online in 2027 that will need roughly 2 gigawatts of power, enough to imply hundreds of thousands of GPUs. SK Hynix’s telecom affiliate will build a cloud business on Nvidia’s upcoming Vera Rubin systems, and Nvidia is separately putting $1 billion into Naver, the Korean cloud company, continuing the same sovereign AI roadshow that had Nvidia cutting checks to SK Hynix, Naver, and Doosan back in June, to secure 200 megawatts of AI computing capacity for itself and its customers.

I keep coming back to the language Nvidia’s own people used on the call. Raj Mirpuri, the company’s enterprise vice president, said the deal “will help us secure a stable supply of HBM memory” through a co-development arrangement on SK Hynix’s next-generation AI memory. That reads like ordinary procurement language. It isn’t. SK Hynix is the leading producer of high-bandwidth memory, ahead of Samsung and Micron on both volume and yield at the leading edge, and HBM isn’t a part you swap between suppliers the way you’d source a resistor. Every generation gets co-engineered between the GPU maker and the memory maker, with stacking tolerances measured in microns. Locking that relationship in before HBM4 ramps to real volume isn’t Nvidia hedging against a shortage. It’s Nvidia deciding how much of that shortage everyone else has to absorb.

That matters because Google, Amazon, and Meta are all racing to ship their own AI accelerators: TPU, Trainium, MTIA, and every one of those chips needs HBM too. Nvidia spent the last decade building a moat out of CUDA and developer lock-in at the software layer. This is the same instinct applied one layer down, into the physical memory stack that every AI accelerator on the planet depends on, whether it runs Nvidia’s software or not.

The scale of what happened in Seoul this weekend is easy to lose inside any single headline. The $500 billion Nvidia figure was one piece of a $950 billion set of announcements a South Korean presidential adviser laid out at a briefing on Saturday, which also included $750 billion in long-term memory supply from SK Hynix to US companies broadly and a separate $200 billion foundry pact between Samsung and Broadcom. Read together, it’s the clearest sign yet that AI infrastructure spending has stopped being a story about a handful of American hyperscalers writing checks to themselves, and started being a story about entire national industrial bases getting wired directly into one customer’s supply chain.

I’ll admit I’m still working out how to feel about the antitrust angle here. Samsung, SK Hynix, and Micron are already facing a lawsuit over DRAM prices that climbed 700 percent this year, and Nvidia pre-buying half a trillion dollars of future HBM capacity from the market leader is not going to make memory cheaper for anyone standing outside this deal. Whether regulators treat securing supply and cornering supply as different things is a question for lawyers, not me. What I do know is that the vendor financing circularity I flagged when AI capex crossed $700 billion looks different now. Nvidia used to just invest in its own customers to keep the growth story alive. Now it’s writing checks to its suppliers too, sitting on both sides of the same trade, and I don’t know yet what to call that.

Sources