Nvidia and SK Group Announce $500 Billion Partnership for HBM and AI Factory Construction

Nvidia and South Korea's SK Group have signed letters of intent for a $500 billion-plus partnership to build massive artificial intelligence infrastructure and co-develop next-generation memory. It is the largest single AI hardware commitment announced to date, signaling a fundamental shift in how frontier compute is financed and deployed globally. But the gargantuan headline number requires careful unpacking to understand what is actually being built, what is merely projected, and how the semiconductor industry is moving beyond simply selling chips to underwriting entire ecosystems.[1][2]

The strategic collaboration rests on two primary physical pillars that address the most pressing bottlenecks in AI development. First, SK Telecom will construct a 2-gigawatt "AI factory" in South Korea, serving as a massive cloud compute hub for the Asia-Pacific region. Second, SK Hynix—which already dominates the global supply of high-bandwidth memory—is locked in as Nvidia's long-term supplier and co-developer for the next generation of memory chips, ensuring that Nvidia's future architectures will not be starved for data.[1][6]

The $500 billion figure heavily promoted in the announcement is not a single upfront check, nor is it a finalized commercial contract. Instead, it represents the estimated aggregate commercial value across the multi-year partnership. It bundles together Nvidia's massive projected purchases of memory chips from SK Hynix, SK Group's acquisitions of Nvidia supercomputers and networking gear, and the broader revenue generated by ecosystem partners involved in constructing these massive data centers over the next decade.[1][2]

The hardware at the center of the SK Telecom facility is Nvidia's upcoming Vera Rubin DSX platform. Succeeding the highly anticipated Blackwell generation, the Rubin architecture is designed specifically for data center-scale deployments. It integrates accelerated computing, advanced networking systems, and proprietary software to maximize energy efficiency and deliver what Nvidia describes as the lowest cost per token generated. The first phase of this deployment is targeted to come online in 2027.[1][3]

To truly understand the scale of the SK Telecom project, one must look at the staggering power requirements. A typical large enterprise data center today consumes between 50 and 100 megawatts of electricity. The planned South Korean facility is rated for 2 gigawatts—roughly equivalent to the total continuous output of a commercial nuclear power plant, dedicated entirely to powering and cooling racks of AI accelerators.[2][4]

This 2-gigawatt facility serves as the flagship anchor for SK Telecom's broader national ambition. The telecommunications giant plans to build up to 15 gigawatts of AI data center capacity across South Korea by 2029. This massive infrastructure play is designed to position the country as a sovereign AI hub, ensuring that domestic enterprises and government agencies can train and deploy models locally rather than relying entirely on foreign-hosted cloud infrastructure.[2]

The memory component of the deal highlights the most critical physical constraint in modern artificial intelligence hardware. A graphics processing unit is only as fast as the memory feeding it data; if the processor has to wait for information, compute cycles are wasted. High-bandwidth memory solves this by stacking DRAM chips vertically and placing them extremely close to the processor, drastically increasing the speed and volume of data transfer.[4][8]

SK Hynix currently commands the majority of the global high-bandwidth memory market, making it an indispensable partner for any company building AI accelerators. By securing a long-term, locked-in supply agreement, Nvidia is proactively insulating itself against the kind of severe component shortages and supply chain shocks that have historically plagued previous graphics processor boom cycles.[2][4]

Crucially, the relationship between the two tech giants is shifting from a standard buyer-seller dynamic to deep, integrated co-development. Nvidia and SK Hynix are co-designing the next generation of memory—dubbed HBM4—to integrate directly and seamlessly with the Vera Rubin architecture. This ensures that the memory architecture perfectly matches the specific compute and thermal requirements of Nvidia's future silicon.[1][6]

The partnership also illustrates a broader, structural trend in the semiconductor industry: chipmakers are increasingly acting as infrastructure orchestrators and financiers. Nvidia is no longer content to simply design and sell silicon; the company is actively involved in the blueprinting, scaling, and sometimes the financing of the physical data centers that house its products, ensuring that the physical world can keep pace with its hardware roadmap.[4]

This shift toward infrastructure orchestration is evident in other recent moves across the sector. For instance, Nvidia is reportedly negotiating a $250 billion financing guarantee to support a separate massive data center project in Ohio, spearheaded by SoftBank and OpenAI. While chipmakers traditionally sell components to cloud providers and move on, the sheer capital intensity of modern AI is forcing them to underwrite the ecosystem to guarantee future sales.[4]

Despite the massive financial commitments and ambitious timelines, physical constraints remain the ultimate governor of artificial intelligence progress. A 2-gigawatt facility requires not just billions of dollars in specialized hardware, but unprecedented access to stable power grids, advanced liquid cooling infrastructure, and vast amounts of real estate—logistical hurdles that cannot be solved simply by signing a letter of intent.[1][3]

The agreements signed by Nvidia CEO Jensen Huang and SK Group Chairman Chey Tae-won formalize a deep symbiotic dependency. Nvidia absolutely needs guaranteed access to advanced memory to maintain its hardware dominance and performance leads, while SK Hynix requires massive, guaranteed volume commitments to justify the staggering capital expense of building next-generation semiconductor fabrication plants.[2][6]

Ultimately, the $500 billion umbrella agreement serves as a roadmap for the physical reality of artificial intelligence over the next decade. As models grow exponentially larger and move toward complex physical and agentic tasks, the industry is moving rapidly past individual server racks and into the era of gigawatt-scale, purpose-built AI factories.[6][7]