South Korea Secures 260,000 Nvidia Blackwell Chips in 10 Billion Dollar Push for Sovereign AI

A Record Allocation That Recasts Korea’s AI Capacity

South Korea has secured more than 260,000 of Nvidia’s latest Blackwell AI processors, a haul that vaults the country into the top tier of computing nations. The allocation, announced during the APEC gatherings in Gyeongju, where President Lee Jae Myung met Nvidia CEO Jensen Huang and corporate leaders, will raise the national stock of high end AI chips from roughly 65,000 to over 300,000. Government officials describe the plan as the foundation of sovereign AI infrastructure, a way to equip Korean researchers, startups and manufacturers with domestic computing power rather than relying on foreign cloud services. Industry estimates place the total cost at up to 10 billion dollars over several years, though the parties did not disclose a final price or delivery schedule.

GPUs, once best known for gaming graphics, have become the engines of modern artificial intelligence. Training a large language model or an autonomous driving stack requires thousands of accelerators working in parallel inside specialized data centers. Nvidia holds about 92 percent of the data center GPU market by sales, and its Blackwell family sits at the top of the performance ladder for training and inference. The chips pair massive compute throughput with stacks of high bandwidth memory (HBM3E), allowing models to process huge datasets quickly and at lower energy per operation than earlier hardware.

The scale of the order puts Korea in the same conversation as the biggest buyers of AI compute. Executives and policymakers cited a target of third place worldwide for available compute, behind the United States and China. They also framed the deal as a start, not an endpoint. Deliveries are set to start next year and expand in phases, giving operators time to build facilities, train engineers, and create applications that turn raw compute into products and services that can be exported.

Who Gets the Chips and What They Will Build

The allocation spans public infrastructure and private AI factories. The Ministry of Science and ICT plans to deploy more than 50,000 Blackwell GPUs across the National AI Computing Center and sovereign cloud providers such as Naver Cloud, Kakao and NHN Cloud. Naver Cloud alone plans to add over 60,000 GPUs, while Samsung Electronics, SK Group and Hyundai Motor Group each intend to install roughly 50,000 units in advanced data centers to support manufacturing, mobility and robotics.

Sovereign compute for researchers and startups

The National AI Computing Center will serve universities, research institutes and startups with shared access to top tier compute. That pool will support training large language models in the Korean language, building domain specific models for science and engineering, and running generative AI workloads for healthcare, finance and public services. Officials say the center will lower barriers for small teams that cannot afford private clusters, speeding up experimentation and leveling access to modern tools.

Naver Cloud is slated to lead the first wave of installations with an initial tranche of about 13,000 Blackwell and other Nvidia GPUs, then scale toward more than 60,000. Its sovereign cloud will host enterprise applications, digital twins for heavy industry, and what Nvidia calls physical AI, a category that blends simulation, robotics and real world deployment. Domestic providers will keep sensitive data inside national borders while offering performance on par with the largest foreign clouds.

AI factories at Korea’s industrial champions

Samsung plans a semiconductor AI factory that applies Blackwell GPUs, Nvidia’s cuLitho library and Omniverse simulation to improve chip yield, optimize lithography steps and model entire production lines. Executives say they will embed trained models into products from smartphones to appliances and use digital twins to make factory changes in software before touching real equipment.

SK Group is designing an AI factory that can host more than 50,000 Nvidia GPUs for semiconductor research, development and production. Its telecom arm plans to provide a sovereign industrial cloud using Nvidia RTX Pro 6000 Blackwell Server Edition GPUs for robotics, letting domestic manufacturers test and deploy robot workflows, machine vision and safety systems with shorter cycles.

Hyundai Motor Group is deepening its collaboration with Nvidia across mobility and smart factories. The company plans to train and validate models for autonomous driving, in vehicle AI and industrial automation using about 50,000 Blackwell GPUs. Hyundai and Nvidia also outlined a plan for new centers focused on physical AI and regional data infrastructure, with investment commitments of approximately 3 billion dollars.

Why Blackwell Chips Matter

Blackwell represents Nvidia’s flagship platform for training and serving very large models. The architecture combines high core counts, fast interconnects that link GPUs into a single logical accelerator, and multiple stacks of HBM3E memory that sit close to the processor. In practice, this lets engineers fit larger models on a node, move data across GPUs with less delay, and finish training runs in days instead of weeks.

For industry, the headline is speed and scale. A carmaker that once retrained perception models every few weeks can update them far more often. A factory can spin up a full digital twin of a line to test a layout or a new tool recipe before any physical change. Engineers can run physics informed AI that blends simulation and data, then push the best policies back to robots or process controllers.

These chips also help with inference, the phase where models serve answers. Blackwell includes features that reduce energy per token or per image processed, which matters when a service must respond to millions of users or when robots run AI at the edge. The focus on both training and inference gives buyers more flexibility as workloads shift from building models to running them at scale.

What Sovereign AI Means for Korea’s Economy

Sovereign AI means the ability to build, train and run critical models on infrastructure governed by domestic institutions, using local data under local law. The goal is to keep strategic capabilities inside the country while still connecting to global markets where Korean firms sell products, parts and services.

For South Korea, sovereignty has a practical business angle. Manufacturers can train models on sensitive production data without sending it overseas. Hospitals and research labs can develop medical models with stricter privacy control. Public agencies can use generative systems while applying local rules on safety and auditing.

Officials also see spillovers. Shared compute lowers startup costs, which tends to raise the number of experiments and new companies. Large firms gain new tools for yield improvement and supply chain planning. Export revenue can come from software, models and services that encode Korean know how, turning decades of manufacturing expertise into digital products.

Energy, Water and Land: The Infrastructure Challenge

The benefits come with substantial infrastructure demands. A campus that houses tens of thousands of accelerators requires large plots of land, high voltage grid connections and specialized cooling. Power draw depends on the configuration, but the totals add up quickly when multiple sites come online. Water use rises as operators adopt liquid cooling to manage heat. Communities expect careful planning to keep data centers efficient and resilient.

South Korea already faces energy security questions, balancing nuclear plants, liquefied natural gas imports and a growing share of renewables. Building AI capacity at this scale will require grid upgrades, long term electricity contracts and advanced cooling such as warm water liquid loops and heat reuse. Officials say Nvidia is working with Korean authorities and operators on designs that cut energy per computation so that capacity growth does not overwhelm local resources.

Operators are also exploring seawater cooled facilities on the coast, recycled water systems inland, and partnerships that reuse waste heat in district heating. Regulators can push for tight power usage effectiveness targets, clear permitting, and incentives for load shifting to off peak hours. These steps turn a potential bottleneck into a test bed for greener computing that still delivers the performance labs and factories need.

Geopolitics and Export Rules Shape the Deal

The timing reflects a strategic pivot in global chip flows. The United States restricts exports of the most advanced AI processors to China on national security grounds, which pulled down Nvidia’s sales in that market. The company has responded by deepening partnerships in allied countries, from Japan to India and now South Korea. Seoul, a treaty ally, is not subject to China focused bans for this class of chips, though licenses and technical specifications are still governed by Washington.

US President Donald Trump signaled a harder line on access to the top end processors in a recent interview. He said the most advanced Blackwell chips should stay with American buyers, while leaving open the possibility of lower capability versions in China.

We do not give the Blackwell chip to other people.

The debate over what can ship to China explains part of Nvidia’s intense focus on markets such as South Korea. Jensen Huang, Nvidia’s chief executive, framed the Korea deal in broader terms during meetings in Gyeongju, saying the country could become a net exporter of intelligence, created in its data centers and embedded across products and services.

Korea can now export intelligence as a new driver of global transformation.

At the same time, Nvidia has expressed interest in selling to China if permitted, citing the need to fund research and serve global customers. For now, the bulk of its growth is coming from projects in countries that face fewer controls and have the industrial capacity to stand up large AI sites.

Supply Chain Effects and Winners Across Korean Industry

The order ripples through the supply chain. Each Blackwell unit can include up to eight stacks of HBM3E, which points to more than two million high bandwidth memory modules across the full deployment. Samsung Electronics and SK hynix are the leading producers of HBM and stand to book large orders. TSMC will fabricate the core chips and perform advanced packaging, a step that remains in tight supply, so coordination across firms is essential to hit delivery targets.

Korean telecom and research organizations are linking the new compute to next generation networks and science. Teams at Samsung, SK Telecom, ETRI, KT, LGU+ and Yonsei University are working with Nvidia on AI radio access networks and 6G features aimed at lower power and higher reliability. The Korea Institute of Science and Technology Information plans a Center of Excellence tied to the HANGANG national supercomputer and new work on connecting quantum processors and GPU systems with the NVQLink open architecture.

Construction, power equipment and cooling vendors also get a boost as sites expand. Data center builders will need local technicians trained in liquid cooling, high density power distribution and GPU cluster orchestration. Public programs through the Ministry of SMEs and Startups and Nvidia’s Inception initiative are lining up to help startups rent time on clusters, use toolkits such as NeMo and CUDA, and move proofs of concept into production.

From Compute to Capability in Language, Robotics and Manufacturing

Compute is necessary, yet it is only the starting point. Korea is backing a consortium that includes Naver Cloud, LG AI Research, SK Telecom, NC AI and Upstage to build foundation language models tailored for Korean users and industries, using Nvidia NeMo and open Nemotron datasets. The aim is not just translation or chat. Teams want better reasoning, retrieval and tool use, with models that speak technical vocabulary for shipbuilding, chipmaking, finance and healthcare.

Physical AI sits close to the Korean industrial base. In this approach, engineers train models in rich simulations that mirror the physics of factories, warehouses and roads, then deploy them on robots and vehicles. SK Telecom’s RTX Pro powered industrial cloud and Hyundai’s mobility stack are aligned to that flow. Digital twins running on Blackwell clusters feed insights to the shop floor, where robots and controllers keep learning from the real world.

The business case centers on exportable capability. If Korean firms can convert their unique manufacturing and design expertise into AI models and services, they sell not only cars or chips but also intelligence that upgrades plants and products in other countries. The government wants small and medium companies to benefit as well by renting compute at fair prices, so the AI dividend spreads beyond the largest conglomerates.

What Could Go Wrong and How Korea Plans to Mitigate Risk

Execution will decide how much of the promise turns into real outcomes. Delivery schedules can slip if packaging capacity tightens or if components such as HBM run short. Export policy can shift. Concentration on a single vendor creates exposure, even if that vendor currently dominates the market. Operators also need to budget for rapid hardware cycles as new generations arrive.

Power and water are the most immediate constraints. Local communities will judge projects by grid impact, water stewardship and carbon intensity. The country has climate goals that will press data centers to reach low power usage effectiveness, recover heat, and source cleaner electricity. Cybersecurity and model safety standards must keep pace with the rush to deploy.

Korea is trying to address these risks with phased rollouts through the late decade, more efficient rack designs, and a wider set of partners across hardware and software. Officials encourage a competitive ecosystem that includes future accelerators, support for open source software, and vouchers that let startups test ideas on sovereign clusters. The public private alliances formed around this deal provide a structure to coordinate training programs, security baselines and best practices.

At a Glance

• South Korea will receive more than 260,000 Nvidia Blackwell GPUs, lifting its national stock of AI chips above 300,000.
• The price was not disclosed, with industry estimates near 10 billion dollars over several years.
• Government plans to deploy over 50,000 chips for sovereign infrastructure; Samsung, SK Group and Hyundai will install about 50,000 each; Naver Cloud is adding more than 60,000.
• The National AI Computing Center will provide shared access for universities, labs and startups.
• AI factories will support digital twins, semiconductor yield improvement, robotics and autonomous driving.
• Data center growth brings energy and water challenges, with efforts focusing on efficient cooling, heat reuse and grid upgrades.
• US export controls limit sales to China; the Korea deal aligns with partnerships in allied countries.
• HBM3E demand will benefit Samsung and SK hynix, while research and telecom teams advance AI RAN and 6G.
• Rollout begins next year and expands in phases across the decade.
• The strategic aim is to turn compute into exportable AI models, services and smarter manufacturing.