NVIDIA AI infrastructure partnerships across South Korea

The most striking feature of NVIDIA's June 2026 Seoul announcements is that they form a coordinated division of labor across Korea's largest conglomerates rather than a set of unrelated GPU sales ^[1]. SK Group takes the foundational layer, signing a multi-year partnership to codevelop next-generation memory across four NVIDIA platforms while SK Telecom stands up a gigawatt-scale AI cloud on the NVIDIA DSX platform with its first data center online in early 2027 ^[4]. NAVER occupies the sovereign-AI layer, building a full-stack AI factory on DSX, joining the NVIDIA Nemotron Coalition as the first Korean member, and expanding its GAK Sejong site toward gigawatt scale ^[1]. LG, Doosan, and Hyundai own the physical-AI layer: LG Electronics co-develops humanoid robots on Isaac GR00T while LG Innotek and LG CNS contribute sensing and autonomous-manufacturing software, Doosan Robotics feeds industrial robots into the platform, and Hyundai's Euisun Chung proposed NVIDIA co-invest in the roughly 9 trillion won Saemangeum data center and robot plant ^[3][5]. Read together, the deals stack from memory to compute to models to robots, with each chaebol slotting into a distinct rung of the same national architecture.

These agreements are easy to read as Korea buying its way into AI, but the supply chain runs both directions. SK Group's multi-year deal to jointly develop next-generation memory for AI factories is not a customer relationship — it is codevelopment of the high-bandwidth memory that NVIDIA's accelerators physically cannot ship without ^[4]. NVIDIA framed advanced memory as being at the core of AI-factory performance and tied the expansion of global AI infrastructure to jointly developing it with Korea ^[1]. The Oct 2025 APEC baseline of over 260,000 committed GPUs — Samsung 50,000+, SK 50,000+, Hyundai 50,000 Blackwell, NAVER Cloud 60,000+, and the government up to 50,000 — gives Korea the demand side of the leverage ^[2]. But because SK hynix and Samsung dominate HBM supply, NVIDIA's ability to build accelerators at all depends on Korean fabs, which is why the relationship looks more like a mutual dependency than a one-way procurement. Community discussion of the deals picked up exactly this thread, noting that Korea's HBM position makes NVIDIA reliant on Korea even as Korea leans on NVIDIA's near-monopoly on advanced GPUs.

The headline numbers are GPU counts, but the partnerships reveal that energy and thermal capacity are the real bottleneck for sovereign-scale AI. NAVER's blueprint calls for 55 MW operational in Sejong in the first half of 2027, expanding to 200 MW overseas in 2028, and ultimately a 1 GW-class facility housing hundreds of thousands of GPUs ^[3]. Sustaining a gigawatt of compute is a power-generation problem, which is precisely why Doosan's role pairs robotics with large-scale low-carbon power from small modular reactors and gas turbines, and why LG Electronics is certifying liquid cooling solutions for AI-factory heat management ^[3][5]. NVIDIA itself has framed accelerated-computing infrastructure as becoming as vital as power grids and broadband ^[2]— a framing that is literally true here, since the buildout consumes utility-scale electricity. The deals that look like robotics and cooling sidebars are in fact the load-bearing pieces: without SMRs, turbines, and liquid cooling, the gigawatt projections are aspirational.

Community reception of the announcements has been skeptical-but-impressed rather than euphoric. The market context partly explains the caution — NVIDIA ties had already helped push the KOSPI up roughly 76% year-to-date in late 2025, its best stretch in over two decades, so much of the upside was arguably priced in before June ^[7]. Discussion threads raised three durable critiques: that buying chips is not the same as building capability, that NVIDIA's dominant share of advanced GPUs creates CUDA-style vendor lock-in for Korea, and that GPU allocation among Korean players is itself a competitive battleground where NAVER's large reserved allotment could hand it an edge over rivals. There is also genuine execution risk on the official side: the value of the June 2026 deals was not disclosed, and NAVER's gigawatt expansion is contingent on procurement and power availability ^[4]. The national computing initiative — roughly 2 to 2.9 trillion won across the support project and the National AI Computing Center — anchors the public commitment ^[6][8], but the gap between committed GPUs and operational sovereign capability is where the skeptics are watching.