On July 24, 2026, NVIDIA and SK Group signed a letter of intent (LOI) for a comprehensive partnership exceeding $500 billion, combining AI factory construction with next-generation AI memory. SK Telecom will build a computing base of up to 2GW in Korea using NVIDIA's systems, while SK hynix will cooperate on long-term supply of memory including HBM4 to NVIDIA. In other words, the SK side is simultaneously acting as both a buyer of AI computing resources and a seller of their core components. The meaning behind the $500 billion figure becomes clear once this bidirectional commercial flow is separated.
$500 billion is not a one-directional investment amount
What the two companies signed at the K-AI Summit in San Francisco was an LOI intended to move future cooperation toward formal contracts. The announcement states the total as "exceeding $500 billion," but does not disclose the target period, the allocation per project, or the amount already committed. It is not positioned as NVIDIA's confirmed order value, nor as a lump-sum capital expenditure figure that SK Group will contribute.
Within this cooperation, products and funds do not move in the same direction. SK Telecom will receive next-generation GPUs from NVIDIA and advance investment cooperation toward AI factories. Meanwhile, NVIDIA will receive stable supply of next-generation AI memory from SK hynix. Furthermore, the two companies will jointly develop HBM, optimizing it for NVIDIA's computing infrastructure.
| Cooperation Area | Main Executing Parties | Announced Content | Undisclosed Items |
|---|---|---|---|
| AI Factory | SK Telecom, NVIDIA | Up to 2GW, adopting DSX and Vera Rubin | Construction site, capacity per facility, number of GPUs, capital expenditure |
| AI Memory Supply | SK hynix, NVIDIA | Long-term stable supply of next-generation memory including HBM | Duration, quantity, unit price, exclusivity |
| Joint Development | SK Group, NVIDIA | Joint optimization of AI memory and AI factories | Product-specific roadmap, cost sharing, handling of deliverables |
As this table shows, the $500 billion figure is an explanation of comprehensive scale covering multiple projects of different natures. At this stage, treating it as a single sales contract would create the mistake of conflating NVIDIA's GPU orders with SK hynix's HBM orders as if they were the revenue of a single company. What needs to be confirmed before the massive headline figure is how many formal contracts this will be divided into, and who pays whom over what period.
From GW-class to up to 2GW, figures entered via the LOI
The concrete progress this time concerns the scale of the AI factory. NVIDIA and SK Telecom had announced on June 8 a plan to build a "GW-class" AI cloud in Korea based on NVIDIA DSX, with the first AI factory to begin operation in 2027. The July LOI specified this as "up to 2GW within Korea," and presented a plan to progressively introduce Vera Rubin systems equipped with SK hynix-made HBM4 starting in 2027.
Here, 2GW is not a figure representing the number of GPUs or computational performance. It indicates the scale of a large-scale facility in terms of power capacity, and the actual computing resources that can be housed will vary depending on rack configuration and cooling equipment. Power consumed by networking and shared facilities must also be subtracted. Since this announcement does not include power allocation per facility or PUE (Power Usage Effectiveness), it cannot be converted into the number of Rubin GPUs or NVL72 racks.
DSX is also not the name of a specific server product. NVIDIA describes it as a platform for integrated design spanning chips and systems, infrastructure software, facility design including power and cooling, and operations. SK Telecom plans to use this design framework to provide training and inference infrastructure for sovereign AI, enterprise AI, physical AI, and agentic AI, expanding its customer base from Korea to the Asia-Pacific region.
The manufacturing-oriented AI factory announced by SK Group and NVIDIA in October 2025 involved deploying over 50,000 GPUs, with an initial batch using over 2,000 RTX PRO 6000 Blackwell Server Editions. In contrast, the 2GW plan is centered on Vera Rubin, with the number of GPUs undisclosed. Connecting these two figures, which differ in both generation and use case, cannot be used to estimate the current deployment count.
Vera Rubin and HBM4, two directly linked roadmaps
The memory cooperation between SK hynix and NVIDIA did not suddenly begin in July. The two companies announced on June 7 a multi-year technology partnership to jointly develop next-generation memory aligned with NVIDIA's AI infrastructure plans and support the supply of advanced memory. The scope includes memory for the Vera Rubin AI supercomputer, as well as the Vera CPU, RTX Spark-equipped PCs, and the Jetson Thor robotics platform. The July LOI serves as a follow-up measure, incorporating long-term supply and stable procurement into the cooperation framework exceeding $500 billion.
HBM4 is built directly into the performance design of Vera Rubin itself. According to NVIDIA's preliminary specifications, a single Rubin GPU will feature 288GB of HBM4 and 22TB/s of memory bandwidth. The Vera Rubin NVL72 connects 72 Rubin GPUs, providing a combined 20.7TB of HBM4 and 1580TB/s of memory bandwidth. In large language model training and inference handling long contexts, if data cannot be fed from memory even when the GPU is capable of computation, utilization drops—making HBM capacity and bandwidth a determining factor for overall system throughput.
This cooperation moves a step beyond the relationship where NVIDIA procures memory for completed GPUs, with SK hynix now jointly optimizing next-generation AI memory from the development stage. For SK hynix, this allows earlier insight into NVIDIA's product plans, making it easier to structure mass-production investment; for NVIDIA, it enables extending supply plans beyond Rubin into the longer term. However, the supply volume, price, and contract duration for HBM4 have not been disclosed, nor is this an agreement for SK hynix to exclusively supply HBM4 to NVIDIA.
Before the GPUs, the land and power to drive the 2GW
Ahead of the LOI with NVIDIA, SK Telecom's board decided on July 23 to establish an AI data center development company, "SK Hyper." It will contribute 750 billion won in stages by 2030, tasking the company with securing land, constructing and operating substations, acquiring customers, and commercializing the business. This is not funding disclosed as part of the $500 billion breakdown, but rather a capital contribution framework for business infrastructure that SK Telecom disclosed independently.
SK Hyper's roadmap calls for sequentially opening the first phase of 5GW starting in 2029, expanding to a total of 15GW by 2035 depending on demand. In its July 5 announcement, SK Telecom also outlined a concept to build over 2GW in the southeastern region of Korea and 1GW in the southwestern region. However, it has not clarified which region or existing project the up to 2GW agreed upon with NVIDIA will be allocated to.
Turning the plan into actual facilities requires securing land and grid power before procuring Vera Rubin systems, building substations and cooling equipment, and securing anchor tenants for long-term use. SK Telecom itself has explained that it will consider location, power supply, and customer acquisition in an integrated manner, citing funding, semiconductors, and network supply as risks. This is not an announcement of bringing the full 2GW online at once in 2027, but rather a plan to operate the first AI factory that year and build up capacity afterward.
Whether the LOI exceeding $500 billion has substance can be measured by whether the formal contract fixes the duration and purchase volume, and whether the first AI factory with NVIDIA begins operation in 2027. SK Hyper's domestic roadmap has separate checkpoints, with 5GW in 2029 and 15GW in 2035 as milestones. Where the up to 2GW NVIDIA project fits within that has not yet been indicated. The first checkpoint will be confirming, in 2027, whether HBM4 mass production and the initial operation of the AI factory align.