Nvidia's $500B Memory Deal Is Still an LOI

Nvidia and SK Group signed a $500 billion-plus AI-factory and memory initiative whose planned 2-gigawatt cloud represents 40% of the 5GW data-center program Korean officials described this week. The operator signal is real—HBM supply is moving into long contracts—but the parties signed letters of intent and disclosed no stack volume, wafer allocation, price, or take-or-pay commitment.

That gap between strategic intent and executable supply is the story. The Nvidia announcement says SK Telecom will build an AI cloud of up to 2GW using Vera Rubin systems and SK hynix HBM4, with the first factory targeted for 2027. The SK hynix version promises stable next-generation memory supply and deeper codevelopment. Neither document says how much of the $500 billion belongs to servers, power, HBM, or services.

Scarcity is moving into the design room

The pact matters because Nvidia is no longer waiting for memory to arrive as a qualified component. A June partnership announcement already tied SK hynix memory to Vera Rubin supercomputers, Vera CPUs, RTX Spark PCs, and Jetson Thor robotics while putting Nvidia simulation and digital-twin tools inside SK’s fabs. The July LOIs extend that relationship from purchasing into joint design and manufacturing. Memory becomes part of the platform roadmap, not a line item chosen after the GPU.

The scale around it explains the urgency. Korean officials said SK’s semiconductor arrangements with Nvidia and other US companies total $750 billion, while Samsung’s Broadcom collaboration adds more than $200 billion, taking the five-year headline above $950 billion. The same briefing described 5GW and two million GPUs across AI-data-center projects, according to Yonhap’s account of the government announcement and Reuters reporting syndicated by CNA.

Set SK Telecom’s 2GW against that 5GW envelope and one Nvidia-aligned project equals 40% of the announced power. That derived ratio is more useful than pretending $500 billion is an HBM order. It says one platform relationship could absorb an extraordinary share of the region’s grid connections, liquid cooling, networking, packaging, and engineering talent. A proportional allocation would imply 800,000 of the two million GPUs, but rack power varies too much to call that a forecast. Power share is the defensible point.

Samsung’s response shows the same bundling logic. Its official Broadcom collaboration spans HBM, sub-2nm foundry services, and 2.3D/2.5D packaging with an estimated value above $200 billion through 2030. Customers are not reserving a chip; they are reserving the stack that makes a chip deliverable. The dynamic extends the compute reservation race behind Alphabet’s $811 billion commitment pile and AMD’s $5 billion entry onto Anthropic’s roadmap.

SK hynix is backing the thesis with long-dated capacity, although not all of it is committed. Its investment plan totals KRW 1,100 trillion across Yongin, Cheongju, and a new southwestern cluster, phased according to demand and approvals. A more concrete Cheongju plan assigns KRW 20 trillion to the P&T7 advanced-packaging facility, with initial line completions targeted around late 2027 and early 2028. Counterpoint estimates SK hynix held roughly 58% of Q1 2026 HBM shipments, with Samsung and Micron near 21% each. Nvidia is making the leader harder for rivals to dislodge by collaborating before qualification begins.

The LOI is a signal, not inventory

The strongest counterpoint is printed in the documents: these are letters of intent. There is no minimum purchase, price, prepayment, annual delivery schedule, cancellation penalty, or disclosed allocation of HBM4 stacks. The $500 billion-plus number is an estimated commercial opportunity across categories, not committed capex and certainly not a single memory purchase order. Operators should not translate a summit headline into an available-capacity forecast.

Execution risk is equally physical. Rubin qualification could slip; HBM4 yields and advanced packaging could miss targets; the grid may not energize 2GW on schedule; cooling and construction costs can outrun financing. SK’s biggest fabs arrive over a long horizon—Yongin’s fourth-fab cleanroom is targeted for 2033—after today’s shortage may have changed shape. The same supply ceiling that constrained Intel’s 2026 outlook can appear at packaging, power, or networking even if wafers arrive.

Who should change decisions now? Independent GPU clouds and sovereign-AI programs planning 2027 Rubin capacity should negotiate reservations this quarter, with explicit HBM generation, delivery remedies, and cancellation terms. Enterprises too small to secure factory-level commitments should diversify across clouds and accelerator families, then use routing to avoid being hostage to one allocation queue. The economic logic resembles Kimi K3’s API-first deployment verdict: buy the service experiment before owning the giant serving topology.

The verdict changes with contract evidence. Take-or-pay quantities, dedicated wafer or packaging allocations, customer prepayments, grid permits, and energized megawatts would turn scarcity from inference into obligation. It weakens if Samsung or Micron wins meaningful HBM4 qualification, Rubin slips, or market lead times and prices fall for several quarters.

For now, Nvidia and SK have revealed the direction of travel: memory supply is being negotiated years before the model run. The 40% power share shows how much industrial capacity one alliance can command. The missing volume terms show why smaller buyers should treat the announcement as a warning to reserve, diversify, and demand contractual specificity—not as proof that two gigawatts of usable AI capacity will arrive on time.