Nvidia and SK Group have signed letters of intent to develop an alliance valued at over $500 billion, combining data centers, Vera Rubin systems, and high-bandwidth memory from SK Hynix. This figure does not represent a simple chip purchase or a finalized contract with the memory manufacturer; it encompasses a long-term initiative that will still need to be solidified through definitive agreements.
The key points of the Nvidia and SK agreement in 30 seconds
- Nvidia and SK Group propose a collaboration exceeding $500 billion through letters of intent.
- SK Telecom plans to build a AI infrastructure of up to 2 gigawatts based on Nvidia Vera Rubin.
- SK Hynix will supply and jointly develop future generations of HBM memory with Nvidia.
- Samsung and Broadcom separately announced a collaboration estimated at over $200 billion through 2030.
- Naver, Brookfield, and Nvidia will finance another $10 billion project to expand a data center in South Korea.
The announcements were made during an AI summit held in San Francisco, led by South Korean President Lee Jae-myung. The event gathered representatives from Nvidia, SK Group, Samsung Electronics, Broadcom, Naver, Hyundai Motor, and various academic institutions.
These series of agreements generated headlines with projected projects surpassing $700 billion. However, these figures should be interpreted with caution: they include memoranda of understanding, letters of intent, future investments, and estimates accumulated over several years.
Not all this money will be spent immediately nor does it guarantee revenues for a single company.
The $500 billion agreement isn’t just about buying HBM
The larger announcement unites Nvidia with SK Group, a conglomerate that includes SK Hynix and SK Telecom. Both parties describe this as a comprehensive initiative of over $500 billion covering AI infrastructure development and advanced memory supply.
The collaboration has initially been formalized through letters of intent. Such documents indicate a commitment to negotiate and develop a project but are not necessarily binding contracts for the full announced amount.
Part of the initiative involves SK Telecom, which aims to build a cloud infrastructure in South Korea dedicated to AI with a capacity of up to 2 gigawatts. This project will utilize Nvidia’s DSX architecture and Vera Rubin-based systems, which are successors to Blackwell accelerators and platforms.
The first “AI factory” is scheduled for 2027. Nvidia describes these data centers as being specifically designed for training models, running inference, and generating tokens used by generative applications and agents.
The scale of energy consumption illustrates the project’s magnitude. Two gigawatts involve more than just powering servers; they require substations, electrical grids, cooling systems, backup infrastructure, buildings, connectivity, and a supply chain capable of delivering large quantities of accelerators and memory modules.
The agreement also enables SK Telecom to offer these resources as a service to companies, public organizations, and developers throughout the Asia-Pacific region. The operator aims to become not just a GPU buyer but a Nvidia cloud service provider.
Nvidia aims to secure the memory needed for Vera Rubin
SK Hynix will cover the other major part of the alliance. The company will expand its cooperation with Nvidia to supply and develop future generations of memory for AI systems.
High-bandwidth memory, known as HBM, has become one of the most critical, scarce components of modern accelerators. It is integrated alongside the GPU using advanced packaging and enables data transfers at speeds vastly exceeding conventional memory.
Nvidia and SK Hynix will collaborate on optimizing new solutions, including HBM4. This partnership will allow accelerator manufacturers to adapt their platforms more swiftly to memory characteristics and help SK Hynix plan production capacity based on anticipated demand.
Supply isn’t just about delivering chips once they’re produced. Designing HBM involves coordinating interfaces, power consumption, thermal management, stacking of layers, and packaging that connect memory with the processor.
This close relationship can help Nvidia reduce one of its key growth risks: ensuring the supply not only of GPUs manufactured by TSMC but also of memory, advanced encapsulation, boards, networks, and power capacity.
In return, SK Hynix gains a high-volume customer and the opportunity to participate early in developing upcoming platforms. The company benefits from strong demand for HBM but must make substantial investments to expand factories and keep pace with technological advancements.
Samsung and Broadcom add another $200 billion
The summit also announced an extension of the relationship between Samsung Electronics and Broadcom. Both companies signed a memorandum of understanding to collaborate on memory, chip manufacturing, and advanced packaging.
Samsung estimates this collaboration could surpass $200 billion over the next five years, until 2030.
In memory, the deal includes supplying HBM solutions for future accelerators developed by Broadcom. The American firm designs custom chips and connectivity systems for major cloud operators, a market that has grown thanks to internal projects by companies like Google and other hyperscalers.
In manufacturing, Broadcom will explore using Samsung Foundry’s 2 nanometer and below processes. The collaboration also covers 2.3D and 2.5D integration technologies for connecting processors, chiplets, and memory within a single package.
For Samsung, the agreement carries importance beyond the figures. It aims to regain ground against SK Hynix in HBM and TSMC in advanced manufacturing. Securing Broadcom’s designs could help fill new production lines and showcase the maturity of its 2-nanometer process.
Nonetheless, this remains a memorandum of understanding. Actual volumes will depend on process validation, industrial performance, Broadcom’s needs, and how the AI infrastructure market evolves.
Nvidia to invest $1 billion in Naver project
Naver, Brookfield, and Nvidia announced another project to expand South Korea’s sovereign AI infrastructure.
The plan will raise Naver’s GAK Sejong data center capacity from 55 to 200 megawatts. The facility will use Nvidia DSX and combine Blackwell systems with the upcoming Vera Rubin platform.
The three companies value the project at $10 billion. Brookfield will contribute up to $9 billion as the exclusive financial partner, Nvidia will invest $1 billion, and Naver will cover the remaining necessary amount.
This expansion equates to approximately 100,000 GPUs, though the final count will depend on the configuration and hardware generations used, according to the companies.
Naver intends to provide computing capacity to companies, government agencies, and developers in South Korea and the United States. They will also use the infrastructure for HyperCLOVA X, AI agents, and models applied to robotics and physical environments.
This agreement reinforces a trend changing the relationships between chip manufacturers and customers. Nvidia is no longer just selling accelerators; it’s involved in financing, designing, and operating the infrastructure that runs on its platforms.
South Korea aims to connect memory, manufacturing, and data centers
These announcements are part of a broader industrial strategy. South Korea already holds a dominant position in memory due to Samsung and SK Hynix but seeks to capture a larger share of the AI-related business.
The country needs to transform its expertise in DRAM and NAND into offerings that include HBM, logic manufacturing, packaging, data centers, cloud services, and proprietary models.
The partnership with Nvidia helps connect these components. SK Hynix supplies memory, SK Telecom develops infrastructure, Naver operates cloud services, and Samsung provides chips, foundry, and encapsulation.
However, the scale of these initiatives increases risks. Building multi-gigawatt data centers requires securing energy, water, permits, networks, and clients willing to contract capacity. Additionally, some projects will not commence operations until 2027 or later.
The evolution of accelerators over that period could also alter plans. Vera Rubin still needs to reach the market, and future architectures will demand new versions of HBM, faster networks, and more advanced cooling systems.
The summit concluded with a casual dinner among policymakers and business leaders. Informal images and comments — including Jensen Huang’s interest in extending the evening with another beer — added anecdotes. Yet, the most significant development occurred beforehand: South Korea presented a bid to turn its manufacturing and memory capacity into a globally scaled AI infrastructure.
Frequently Asked Questions
Has Nvidia bought $500 billion worth of memory from SK Hynix?
No. The figure relates to a broader initiative between Nvidia and SK Group that includes data centers, computing systems, and memory. It was announced via letters of intent and has yet to be formalized in definitive agreements.
What will SK Telecom build with Nvidia?
SK Telecom plans to develop a AI infrastructure of up to 2 gigawatts based on Nvidia DSX and Vera Rubin systems. The first facility is scheduled for 2027.
What role will SK Hynix play?
SK Hynix will supply advanced memory and collaborate with Nvidia on developing and optimizing future HBM generations, including HBM4.
How much will Nvidia invest in Naver?
Nvidia has committed to a $1 billion investment in a $10 billion project. Brookfield will fund up to $9 billion, with Naver covering the rest.
