SK Group and Nvidia signed letters of intent for a planned $500 billion-plus initiative spanning AI factories and next-generation memory. The signatures are complete. The factories, financing, procurement, memory deliveries, customers and returns are not. [2] [3]
South Korea describes a $950 billion umbrella as a $750 billion SK scope that includes, rather than adds to, the $500 billion-plus Nvidia/SK initiative, plus an approximately $200 billion Samsung/Broadcom memorandum of understanding. [4] [5] The corporate releases call the Nvidia/SK instruments letters of intent. [2] [3] The umbrella broadens the national context without changing their completion status.
That distinction extends the paper's July 24 account of Alphabet's planned AI spending, which kept a $205 billion capital plan separate from completed capacity, use, revenue and return. SK and Nvidia have now put names on a larger ambition, but not receipts beneath it.
The companies describe two connected plans. SK Telecom intends to build a 2-gigawatt-scale AI Cloud in Korea using Nvidia's DSX architecture and Vera Rubin accelerated-computing systems. SK hynix and Nvidia intend to form a long-term memory partnership covering supply, codevelopment and optimization of next-generation high-bandwidth memory. The releases say the planned cloud would use SK hynix HBM4 and that its first AI factory is planned to come online in 2027. [2] [3]
Those are consequential commitments. Nvidia designs systems whose accelerators need high-bandwidth memory; SK hynix makes that memory; SK Telecom can assemble the computing service around both. The proposed structure puts the system designer, memory supplier and cloud operator inside one program. The letters themselves disclose none of the orders, sites, power agreements or delivery schedules needed to execute it.
CNBC supplies the more muscular version. Its headline says Nvidia has "locked down" memory as part of a $500 billion AI deal, and its account says the agreement could be worth that amount over several years. CNBC also reports that Nvidia is seeking supply while the memory market faces a global shortage and quotes an Nvidia executive saying the expansion would help secure a stable supply of HBM. [1] That language explains the commercial objective. It does not publish quantities, prices, delivery dates, qualification tests or exclusivity.
The primary announcements use both the language of partnership and the caution of corporate counsel. They say the letters formalize plans ranging from factory construction to memory supply, then devote extensive notices to forward-looking statements, risks and outcomes that may differ from expectations. SK hynix's release goes further: no part of it should form the basis of a contract, commitment or investment decision. [2] A press release can celebrate a commitment stage while warning readers not to mistake it for performance.
The 2-gigawatt description demands the same care. It names the scale the companies intend to build, not electricity already contracted, connected or consumed. Neither primary announcement identifies the sites, permits, utility agreements, construction contracts, system count or commissioning tests behind the figure. The announced 2027 date belongs to the first planned factory, not to an operating facility available today. [2] [3]
Memory is equally specific in product and unspecific in commerce. The releases name HBM, including HBM4 in the planned Vera Rubin deployment, and describe codevelopment and optimization across training, agentic AI and physical AI workloads. [2] [3] They do not report delivered units, reserved wafer capacity, price, duration, remedies for shortfalls or supply left for other customers. "Stable supply" is the purpose of the relationship, not a completed inventory count.
Nvidia's official X post removes most of those verbs from the future tense. It presents the factories, memory work and Korea's industrial base as one national ascent, saying SK Telecom "is building" the 2-gigawatt factory and that the partners "will codevelop" memory. The post is real and useful evidence of how the partnership wants to be seen. It is not financing, a purchase order, a power agreement or a delivery receipt.
The signed letters matter because they align three companies around an unusually large infrastructure program. They also create a clear test. Financing should lead to budgets. Budgets should lead to orders, sites and power. Those should lead to construction, installed systems, qualified HBM4, commissioned capacity, customers and utilization. Revenue and return come after all of that, if they come at all.
For now, SK and Nvidia have completed the first step and advertised the last. Calling the space between them a $500 billion deal makes the promise easy to rank against every other AI spending announcement. Calling it a planned $500 billion-plus initiative under letters of intent makes the work still missing impossible to overlook. The next meaningful headline needs a financing instrument, construction milestone or delivery record, not another rendering of the same ambition.
-- THEO KAPLAN, San Francisco