The Deal the Press Release Wants You to See
Here is the headline number: $500 billion. That is what NVIDIA and South Korea's SK Group announced at an AI summit in San Francisco on July 24, 2026 — a "mega alliance" covering high-bandwidth memory supply and a 2-gigawatt AI factory. The coverage has been breathless, and the word "historic" appears in almost every headline.
Read the fine print, though, and the first thing you find is a phrase that rarely shows up in the celebration: letters of intent. Not purchase orders. Not binding contracts. Letters of intent — the corporate equivalent of a handshake photographed for the front page.
Here is what the two companies actually said they will do. SK Telecom will build a 2-gigawatt AI cloud facility in South Korea on NVIDIA's DSX AI factory architecture, running Vera Rubin platforms paired with SK hynix's HBM4 memory. The first phase comes online in 2027. In return, NVIDIA gets a "stable supply pipeline" of next-generation HBM for Vera Rubin and future architectures, plus a joint effort to engineer custom memory stacks tuned for LLM training and real-time inference.
SK Group Chairman Chey Tae-won framed it as national destiny, declaring Korea will "transcend its role as a leading adopter of AI and become a global hub that drives AI innovation." Jensen Huang, never one for understatement, said South Korea is positioned to become a "global AI powerhouse."
Who Actually Holds the Leverage?
Now the uncomfortable part. NVIDIA is the most valuable company on the planet, and it just signed a $500 billion piece of paper to guarantee it can buy memory chips. Sit with that for a second.
The AI boom has produced a global memory shortage, and SK hynix is the leader in HBM production. Compute was supposed to be the bottleneck of the AI era. It turns out the real constraint is a memory oligopoly that can name its price. NVIDIA's own enterprise vice president, Raj Mirpuri, said the agreement "will help us secure a stable supply of HBM memory." Translation: the chipmaker that dominates AI accelerators still has to negotiate for allocation like a startup.
That inversion matters, because this is not really a partnership. It is a supply agreement between a company that needs memory and a company that controls it. When one side owns the bottleneck, the other side pays the memory tax — and that cost does not stop at NVIDIA. It flows straight into every GPU rental, every cloud API price, and every inference bill you will see for the next five years.
- SK Group gets a guaranteed marquee customer plus the prestige of building a national AI hub — politics and business in a single signing ceremony.
- NVIDIA gets allocation and co-engineered HBM stacks, but at a price set by someone else's fab capacity.
- SK Telecom gets a 2-gigawatt flagship and a 2027 deadline that will test whether any company can actually deliver a data center that size.
- Everyone else gets higher memory prices, because locking up supply for one customer does not create more HBM — it just reallocates it.
The Fine Print Nobody Read
Then there is the question of what $500 billion actually buys. CNBC, reporting the announcement, was careful to say the deal "could be worth" $500 billion "over a number of years." Could be. That is the language of aspiration, not obligation.
Notice, too, what else happened that week. Samsung signed a $200 billion memorandum of understanding with Broadcom on the same day. NVIDIA separately committed $1 billion to Naver, a Korean cloud provider. South Korea's president attended the summit personally. This is not merely corporate dealmaking; it is national industrial policy wearing a business suit. Every AI summit now produces a fresh trillion dollars of announcements, and none of it is binding until the shovels are in the ground.
The memory industry also has a history the press releases leave out. It is one of the most cyclical businesses in technology — boom, bust, oversupply, repeat. The last two downturns wiped out billions in shareholder value and left fabs idle. Signing a $500 billion supply agreement at the top of a memory supercycle is either genius timing or the industry's oldest mistake, and nobody at the summit was in a position to tell you which.
- What happens to these gigawatt promises if token demand plateaus?
- Why letters of intent instead of binding contracts, if the deal is really so certain?
- When the cycle turns — as memory cycles always do — who eats the $500 billion?
None of this means the deal is fake. NVIDIA genuinely needs HBM, SK hynix genuinely has it, and the 2027 timeline is ambitious but plausible. The skepticism is about the framing, not the fundamentals. A supply chain agreement is being sold as a world-historic alliance, and the difference between those two things is exactly what the press release leaves out.
So the real question is not whether NVIDIA secured its memory. It clearly did. The question is what happens when the AI buildout slows down, and whether the company that just promised $500 billion of handshakes has a plan for that day. Korea is betting its industrial future on AI factories. The memory giants are betting on a supercycle that never ends. One of those bets is probably wrong — we just do not know which one yet.
Comments