Nvidia's $500B SK deal is a letter of intent, not a contract

Two enormous AI numbers landed inside 24 hours, and both of them are non-binding. Nvidia and SK Group announced a $500 billion-plus partnership across AI factories and next-generation memory. Samsung and Broadcom announced a $200 billion-plus collaboration across memory and foundry. Neither is a signed supply contract. If you're doing capacity or cost planning off headline AI numbers, the contract type is the part you actually need to read.

What actually happened

Per Nvidia's own announcement on July 24, SK Telecom will build an AI cloud in Korea targeting 2 gigawatts of compute on Nvidia's DSX platform, with the first AI factory aimed at coming online in 2027. SK hynix will supply HBM4 memory for Nvidia's Vera Rubin accelerators and codevelop next-generation AI memory. The parties signed letters of intent to formalize it. CNBC pegged the multi-year value at roughly $500 billion.

The day before, Samsung and Broadcom announced an MOU covering HBM supply for Broadcom's AI accelerators plus foundry work at 2nm and below, including 2.3D and 2.5D packaging. CNBC reported the estimate at more than $200 billion across five years through 2030.

An LOI and an MOU are statements of intent. They set direction. They are not purchase orders, and the dollar figures attached to them are estimates of a relationship, not revenue anyone has booked.

Why it matters for your business

Most of these deals will probably happen in some form — that's not the point. The point is what you do with the number. We see operators read "$500 billion of AI compute coming" and conclude their inference bill is about to fall off a cliff in 2027, so they defer the architecture work. That's a bet on a non-binding agreement, denominated in gigawatts that don't exist yet, for capacity already spoken for by hyperscalers.

Here's the more useful read: HBM4 supply through 2030 is being allocated right now, in advance, by the four or five companies large enough to sign letters with two commas in them. You are not in that negotiation. Your token price in 2028 is being set by contracts you'll never see, on hardware that hasn't been fabbed.

So plan like the price is flat. Build the routing layer, the caching layer, and the model-swap seam now, while it's cheap to do. If tokens do get cheaper, you keep the savings. If they don't — or if a memory crunch pushes them the other way — you already have the dials.

Sources: NVIDIA Newsroom, Samsung Global Newsroom, CNBC.