SK Hynix Locks In $750 Billion Nvidia Memory Deal as South Korea's AI Chip Diplomacy Hits $950B

South Korea has just executed the largest coordinated semiconductor diplomacy in its history. South Korean President Lee Jae Myung flew to San Francisco last weekend and sat down with Jensen Huang of Nvidia, Sam Altman of OpenAI, Dario Amodei of Anthropic, and Hock Tan of Broadcom — a gathering that produced two multi-hundred-billion-dollar AI chip supply agreements signed simultaneously.

The headline numbers: SK Hynix committed to supply Nvidia and unnamed US partners with $750 billion worth of high-bandwidth memory chips over five years. Samsung’s concurrent agreement with Broadcom, covering advanced memory and foundry services, adds another $200 billion. Combined: $950 billion in Korean AI chip commitments, announced within days of each other as part of a single state-level diplomatic push.

HBM Is the Bottleneck That Explains Everything

To understand why these numbers are so large, it helps to understand what high-bandwidth memory actually does in an AI system. Every modern AI accelerator — Nvidia’s H100, H200, and Blackwell family; AMD’s MI300 series; Google’s TPUs — is architecturally limited not by compute but by memory bandwidth. The GPU or specialized chip can execute billions of operations per second, but it is constantly waiting for data to arrive from memory. HBM is the solution: stacked layers of DRAM connected to the processor with extremely short electrical paths and extraordinarily wide buses, delivering bandwidth of several terabytes per second compared to the few hundred gigabytes per second of conventional GDDR memory.

Without HBM, there is no frontier AI accelerator. And without SK Hynix and Samsung, there is essentially no HBM. The two Korean companies manufacture the overwhelming majority of the world’s HBM output. SK Hynix holds roughly 50 percent of the global market; Samsung has been working to recover lost ground after early yield problems with its HBM3E generation. Micron is the only other meaningful producer, with a market share well below the Korean duopoly.

The $750 billion SK Hynix-Nvidia commitment is therefore less an ordinary supply contract and more a declaration of shared destiny. Nvidia’s GPU roadmap cannot be executed without SK Hynix’s memory roadmap. SK Hynix’s memory roadmap cannot be validated without Nvidia’s architectural requirements. The long-term deal formalizes that interdependence at a scale that makes both companies’ futures structurally linked for the better part of a decade.

The Diplomatic Architecture

President Lee Jae Myung’s San Francisco summit was not improvised. The visits to Altman, Huang, Amodei, and Tan were coordinated in advance, and the deal announcements were timed to coincide with Lee’s presence — giving South Korea’s head of government the ability to claim credit for a diplomatic outcome visible to both domestic and international audiences.

This reflects a deliberate shift in South Korea’s industrial policy approach. Rather than leaving semiconductor relationships to companies to manage independently, the Lee government has positioned itself as an active intermediary between Korean chip makers and their most important customers. The logic is twofold. First, at the scale of hundreds of billions of dollars, deals benefit from government-level signals of stability and commitment — US companies have more confidence signing long-term supply agreements when both parties can point to a state-level imprimatur. Second, AI has become sufficiently intertwined with geopolitics and national security that technology deals of this scale are increasingly quasi-diplomatic rather than purely commercial.

The meetings with Altman and Amodei are particularly notable because neither OpenAI nor Anthropic are direct counterparties to the chip supply agreements. Their inclusion in the summit suggests the South Korean government was mapping the full AI ecosystem rather than narrowly targeting memory customers — understanding where AI is heading, not just where Korean chips are going today.

SK Telecom’s 2-Gigawatt Data Center

Embedded within the SK Hynix-Nvidia deal is a related commitment from SK Telecom, South Korea’s largest wireless carrier and a member of the SK Group conglomerate. SK Telecom plans to build a 2-gigawatt AI data center in the United States, powered by Nvidia’s Vera Rubin GPU clusters and populated with SK Hynix’s HBM4 memory. The facility is expected to come online in 2027.

A 2-gigawatt data center is a substantial undertaking — 20 times the scale of a conventional hyperscale facility. This places it at the frontier of planned AI infrastructure builds, smaller than the 10-gigawatt OpenAI campus proposed for Piketon, Ohio (the Nvidia backstop deal announced simultaneously), but large enough to rank among the largest AI facilities in the US market.

The SK Telecom build is strategically significant for reasons beyond its size. It represents a South Korean conglomerate making a direct investment in US AI compute infrastructure, rather than merely supplying components to US-based customers. If completed, it would give SK Group a direct stake in the AI cloud market alongside established players like AWS, Azure, and Google Cloud — an unusual position for a company whose primary business is Korean mobile telecommunications.

The Samsung-Broadcom Context

The Samsung-Broadcom $200 billion commitment (already reported and separately analyzed) covers both advanced HBM memory supply and 2-nanometer foundry production for Broadcom’s AI chip designs. Broadcom is the dominant supplier of custom AI accelerators — ASICs — to Google, Meta, and ByteDance, among others. Samsung’s 2nm process is expected to enter production in 2026 for early customers, making the timing of the Broadcom deal particularly relevant: Samsung is essentially pre-selling capacity on a production line not yet fully operational at scale.

Together, the two deals represent something new: South Korea positioning itself as the indispensable supply-chain partner for the dominant AI chip architectures of the next generation. Nvidia’s GPU approach requires SK Hynix’s HBM. Broadcom’s ASIC approach requires Samsung’s foundry. South Korea now has binding, multi-year commitments from the leading representatives of both approaches.

Supply Chain Concentration Risk

The very scale of these commitments creates a risk that deserves acknowledgment. The global AI chip supply chain is already extraordinarily concentrated — in compute, at TSMC; in memory, at SK Hynix and Samsung. Long-term exclusive or semi-exclusive commitments at this scale make that concentration more durable rather than less. If either company faces a natural disaster, a geopolitical disruption, a major manufacturing incident, or a technology setback, the entire global AI infrastructure build-out faces simultaneous constraints.

US policymakers have been explicitly concerned about this concentration, which is part of the rationale for CHIPS Act investments in domestic semiconductor manufacturing. The South Korean deals, by cementing Korean companies’ centrality to US AI infrastructure, arguably cut against that diversification goal — though they do so by binding allied nations more tightly rather than creating dependencies on adversarial suppliers.

For AI companies and data center operators, the deals offer near-term certainty at the cost of long-term optionality. The HBM and foundry capacity locked in by these agreements will be unavailable to competitors for the term of the contracts, creating bottlenecks that will be felt by companies that delayed securing their own supply commitments.

What $950 Billion Tells Us About AI Capex

The combined figure is worth sitting with. $950 billion in chip supply commitments from a single country, announced in a single diplomatic visit, timed to coincide with simultaneous deals totaling hundreds of billions more from Nvidia’s own financing discussions with OpenAI, means that the total announced AI infrastructure commitments in the final week of July 2026 alone may exceed $1.5 trillion.

Whatever the ultimate execution rate on these announced deals — and many large commitments in this space have historically been revised downward or extended over longer time horizons — the trajectory is clear. The AI infrastructure cycle is not slowing. If anything, the scale of commitments suggests that the key bottleneck has shifted from capital availability to physical execution capacity: the ability to actually build, power, cool, and operate facilities at this scale is becoming the binding constraint, not the financing to pay for them.