The Biggest AI Number This Week Was Not About Chips. It Was About Memory.

The largest AI commitment announced this week did not come from a frontier lab or a hyperscaler. It came from Seoul, and the bulk of it is memory.

South Korea outlined about $950 billion in AI commitments over five years. SK Group is pursuing roughly $750 billion in advanced memory supply deals. Samsung Electronics signed a memorandum of understanding with Broadcom worth up to $200 billion. The remainder covers AI data centers and physical AI work with firms including Nvidia, Broadcom, Brookfield, Anthropic and Waymo.

Read that split again. The single biggest line item in the single biggest AI announcement of the week is memory supply.

## The market prices the accelerator. The contracts price what sits beside it.

Almost every AI investment story gets told through GPUs: which lab bought how many, which vendor won the rack, how many exaflops it rates. But an accelerator is increasingly a delivery vehicle for high-bandwidth memory. AMD's Helios platform, unveiled on July 23, packs 72 MI455X accelerators into a single rack. Every one of those parts is a claim on an HBM production line that only a handful of firms on earth can operate.

That is why the Korean number is structured the way it is. You cannot buy your way out of a memory constraint in a quarter. Fabs take years. So the parties who actually have to deliver capacity are signing multi-year supply commitments now, while the public conversation is still about model releases.

## Demand stopped being the uncertainty

The demand side is no longer speculative. OpenAI's Project Camellia is a reported $20 billion campus near Savannah drawing at least 3.2 gigawatts. Korea is building domestic capacity at the same time, with LG CNS backing the Namyangju data center project on a first tranche of 30 billion won.

Both ends of that pipeline are now committed. Power is being contracted, as we argued in the case for owning your own electrons. Land is being bought. What stays contested is the middle: the components that turn a building into usable inference.

## The counter-argument, and why it is weaker than it looks

The obvious objection is that memorandums are not purchase orders. A $200 billion MoU is a statement of intent, and intent is cheap. Korea also has every incentive to announce large round numbers in the days after a summit.

But the direction survives the discount. Even if you haircut these figures hard, the composition still says the same thing about where scarcity sits. And it is corroborated by behavior elsewhere. When remarks from DeepSeek's founder about the lab's reliance on Nvidia chips leaked, the fallout was severe enough that the company paused a raise of at least 10 billion yuan. Hardware dependency is not a technical footnote in this market. It moves valuations.

## What to do with this

Two concrete moves.

First, when you read the next AI infrastructure announcement, separate the compute number from the supply number. A gigawatt figure tells you ambition. A multi-year component supply agreement tells you someone has reserved scarce capacity they cannot cheaply walk away from. Only the second is costly to reverse.

Second, watch whether these Korean agreements convert into disclosed, dated contracts over the next two quarters. If that $750 billion in memory supply starts appearing as specific signed deals, the AI trade has genuinely moved a layer down, and Korea is a primary beneficiary rather than a supplier footnote. If it stays at the MoU stage, it was positioning. That conversion rate is the number worth tracking, and unlike a valuation it is checkable.

We log each agreement as it lands on the AI deal tracker.