South Korean President Lee Jae Myung unveiled three sweeping public-private mega-projects on Sunday, anchored by a commitment from Samsung Group to invest 1,000 trillion won — roughly $649 billion over 10 years — in semiconductors, artificial intelligence data centers, and physical AI including robotics. The announcement, delivered at the presidential compound in Seoul, is the largest single corporate investment pledge in South Korean history and positions the country’s two dominant memory chipmakers, Samsung Electronics and SK Hynix, as the primary engines of what Lee’s government calls a national “great leap” into the AI era.

Samsung Group Executive Chairman Lee Jae-yong and SK Group Chairman Chey Tae-won both attended the presidential briefing and presented their companies’ respective investment plans. The two chairmen had met separately with President Lee in the weeks before the announcement to coordinate the public-private framework — Lee Jae-yong on June 25 and Chey Tae-won on June 19.

Why Samsung’s $649 Billion Is About More Than One Company’s Balance Sheet

The figure deserves context before it becomes a headline. Samsung Group’s 1,000-trillion-won commitment is spread across Samsung Electronics, Samsung SDI, Samsung Display, and Samsung Electro-Mechanics, covering chip factories, AI data centers, next-generation battery production, and advanced substrates for AI semiconductors. More than 350 trillion won is earmarked specifically for AI infrastructure, primarily data centers. SK Group presented separate investment commitments alongside Samsung’s at Sunday’s briefing.

But the investment’s strategic weight does not rest on its dollar amount alone. It rests on what Samsung Electronics and SK Hynix actually make: high-bandwidth memory, or HBM — the specialized chip architecture that feeds data to AI accelerators fast enough to keep them running at capacity. Together, Samsung and SK Hynix supply roughly 80% of global HBM production. SK Hynix alone held a 58% share of the HBM market in the first quarter of 2026, according to Counterpoint Research, and is the primary memory supplier for every NVIDIA GPU generation from the H100 through the forthcoming Rubin platform.

That concentration is not accidental. HBM is the kind of technology where the accumulated manufacturing advantage of the market leaders compounds with each generation. SK Group Chairman Chey Tae-won stated the implication plainly: “What used to be a peripheral component has become a core component. If SK Hynix’s HBM is replaced with another product, the AI system may not function properly.” That is not a marketing claim — it describes the architecture. HBM stacks are physically integrated onto the same silicon interposer as the GPU die, which means switching to a different memory supplier requires redesigning the chip package from scratch.

How HBM Actually Works: The Memory Architecture Behind AI’s Critical Bottleneck

High-bandwidth memory exists to solve a problem called the memory wall. Large language models and other AI workloads stream enormous tensors between memory and processor cores. Conventional DRAM lays memory chips flat across a circuit board, connecting them to the processor through a relatively narrow data bus — a standard DDR5 memory channel provides a 64-bit-wide pathway. That is inadequate for AI training at scale: the math operations are fast; memory access is the bottleneck.

HBM solves this by stacking multiple DRAM dies vertically, connecting them through thousands of microscopic copper pillars called through-silicon vias, or TSVs. Each HBM die is thinned to approximately 30 micrometers — about a third the diameter of a human hair — so that 12 or 16 layers can be stacked within JEDEC’s mandated height limit of 775 micrometers. The completed stack sits on a silicon interposer directly adjacent to the GPU, eliminating the long circuit board traces that limit data transfer speed in conventional memory architectures.

The architectural result is a memory bus 32 times wider than DDR5. HBM3E, the generation currently in production for NVIDIA’s Blackwell GPUs, provides a 1,024-bit interface and delivers approximately 1.15 terabytes per second of bandwidth per stack. HBM4, whose JEDEC standard was published in April 2025 and whose mass production began in early 2026, doubles the interface width to 2,048 bits and delivers up to 2 terabytes per second. Each NVIDIA Rubin GPU will carry eight such stacks, providing an aggregate bandwidth exceeding 16 terabytes per second — enabling real-time inference on AI models at scales that were physically impossible with any prior memory architecture.

SK Hynix shipped samples of HBM4E to major customers on June 18, ahead of its previously stated second-half timeline. HBM4E uses a denser 1c-generation DRAM core die and raises per-pin data rates to as much as 16 gigabits per second, delivering approximately 4 terabytes per second of bandwidth per stack — roughly 3.5 times what HBM3E provides. The base die at the bottom of the HBM4E stack, which controls reads, writes, and error correction for the layers above it, is understood by industry analysts to be manufactured on TSMC’s 3nm-class process, a significant engineering advance over the 12nm-class base die used through the prior generation.

Micron projects the HBM total addressable market will reach approximately $100 billion by 2028, up from roughly $35 billion in 2025 — a compound annual growth rate near 40%. The $100 billion figure, Micron noted in its December 2025 investor presentation, would arrive two years earlier than the company’s prior forecast and would exceed the size of the entire DRAM market in 2024.

👉 Read more: 318633 SK Hynix HBM4E sample specs and AI implications

Three Pillars: Chips, Data Centers, and Physical AI

The program is structured around three distinct investment areas. The centerpiece is a semiconductor cluster in the southwestern Honam region, anchored by Samsung Electronics and SK Hynix. The government also plans to expand AI data center infrastructure across multiple regions, and to invest in physical AI — the convergence of robotics, edge computing, and on-device inference systems, an area where South Korea’s existing industrial base in automotive and precision manufacturing provides a natural competitive foundation.

The data center dimension already has a significant private-sector anchor. The Jeollanam-do provincial government signed a memorandum of understanding with Stock Farm Road — a venture co-founded by LG heir Brian Koo and Amin Badr-El-Din — for a facility designed to reach 3 gigawatts of computing capacity upon projected completion in 2028, with an initial investment exceeding $10 billion and potential scale of $35 billion. If it reaches those specifications, it would be the largest AI data center in the world by capacity.

The physical AI pillar is already in motion through a separate partnership: NVIDIA and Hyundai Motor Group have committed approximately $3 billion to advance physical AI, including an NVIDIA AI Technology Center and a Hyundai Motor Group Physical AI Application Center. NAVER Cloud has separately committed to deploying over 60,000 NVIDIA Blackwell GPUs for sovereign and physical AI workloads.

Why the Geography of Chips Is Also the Politics of Chips

The most politically charged element of Sunday’s announcement is the location of the semiconductor cluster. The Honam region — including Gwangju and South Jeolla Province — is a Democratic Party stronghold where President Lee received approximately 85% of the vote in the June 2025 presidential election, according to election data. Lee won 49.42% of the national vote overall.

The opposition People Power Party argued before the announcement that semiconductor investment decisions should be driven by infrastructure considerations — power supply, water access, logistics, and supplier networks — rather than electoral geography. PPP spokesperson Park Sung-hoon said: “Where semiconductor factories are built should be decided by companies, not by the president.”

President Lee defended the choice, noting in posts on X over the weekend that Gwangju and South Jeolla Province received the highest evaluation in the semiconductor category during a government competition to select specialized national high-tech industry zones conducted under the previous administration. “The creation of a semiconductor industrial ecosystem in the southwest is not a special favour for a particular region,” Lee wrote. “It is the additional creation of the most rational semiconductor industrial centre through the decisions of relevant companies under full government support.”

South Korea’s Regulatory Foundation and U.S. Partnership

The announcement builds on a regulatory and diplomatic foundation constructed since early 2025. South Korea’s AI Framework Act took effect on January 22, 2026, making the country one of the first in Asia to enact comprehensive AI legislation. The law provides regulatory clarity around high-risk AI applications while establishing government mandates to support AI data centers and research infrastructure.

South Korea’s designation as a Tier 1 partner by the United States, granting Korean companies unrestricted access to the most advanced AI chips under U.S. export control rules, has further strengthened the country’s position relative to other AI-ambitious nations in Asia. That partnership has a visible infrastructure expression: at last October’s APEC Summit, NVIDIA CEO Jensen Huang announced a sweeping deployment of more than 260,000 NVIDIA GPUs across Korean government agencies, Samsung, SK Group, Hyundai Motor Group, and NAVER Cloud. “Just as Korea’s physical factories have inspired the world with sophisticated ships, cars, chips and electronics,” Huang said at the event, “the nation can now produce intelligence as a new export that will drive global transformation.”

The Infrastructure Constraint No Announcement Can Waive

Sunday’s announcement establishes a financial and political commitment. The engineering reality is more demanding. Building a leading-edge semiconductor fabrication plant requires stable power of between 200 and 400 megawatts per facility — before accounting for the data center buildout the adjacent AI infrastructure pillar requires. It also requires ultrapure water in industrial volumes, deep supplier networks for specialty process gases and photoresist, and a workforce trained in nanometer-scale fabrication processes.

These are not elements that can be incentivized into existence quickly. McKinsey has documented that greenfield semiconductor ecosystems face the same baseline requirements as established ones, and that building a capable workforce and adequate infrastructure in a new region is one of the hardest questions to answer in advance of a multi-decade investment. Industry experts cited by Reuters noted that diversifying chip investment beyond Seoul could ease infrastructure bottlenecks over time but warned that the elements required for cutting-edge fabs may not scale quickly enough in a new region to meet surging AI demand.

The timeline pressure is real. South Korea’s semiconductor exports reached $37.2 billion in May 2026, a 169% year-on-year increase, setting an all-time monthly record. The memory supercycle driving those numbers is a present phenomenon — driven by AI infrastructure spending already underway and by HBM supply already sold out through the year. Investments announced today will not produce volume output from a new Honam cluster until the late 2020s at earliest.

What Sunday’s announcement does accomplish is equally significant: it establishes the formal framework under which Samsung and SK Hynix will be expected to invest not just at their existing sites in the greater Seoul area but across new regional centers — expanding the total capacity of South Korea’s AI hardware supply chain precisely when that supply chain is functioning as a strategic national asset.

Frequently Asked Questions

What does South Korea’s $649 billion AI investment actually cover?

The figure comes primarily from Samsung Group’s 10-year pledge of 1,000 trillion won (approximately $649 billion), which spans Samsung Electronics’ semiconductor and AI factory investments, Samsung Display’s display manufacturing, Samsung SDI’s battery production, and Samsung Electro-Mechanics’ advanced packaging operations. SK Group presented separate investment commitments alongside Samsung’s at Sunday’s presidential briefing. The government’s own direct spending, through vehicles like the National Growth Fund, is an additional figure on top of these private corporate commitments. The three-program structure covers semiconductor manufacturing in the Honam cluster, AI data center infrastructure across multiple regions, and physical AI including robotics and industrial automation.

What is high-bandwidth memory and why does it matter for AI?

High-bandwidth memory is a 3D-stacked DRAM architecture that delivers data to AI processors at speeds no conventional memory can match. Where a standard DDR5 memory module uses a 64-bit data bus, an HBM4 stack uses a 2,048-bit bus — 32 times wider — and delivers up to 2 terabytes per second of bandwidth. AI training workloads require continuous high-speed access to the parameters of models with hundreds of billions or trillions of values; HBM is the only memory architecture that can feed those parameters to the GPU fast enough to keep the processor fully utilized. South Korea’s Samsung Electronics and SK Hynix supply roughly 80% of global HBM production, giving the country structural leverage over the AI compute supply chain that no other nation currently replicates.

Why is the location of South Korea’s new semiconductor cluster controversial?

The Honam region in southwestern South Korea — the designated site for the new chip cluster — is a stronghold for President Lee Jae Myung’s Democratic Party, where Lee received approximately 85% of the vote in the 2025 presidential election. The opposition People Power Party has argued that semiconductor investment decisions should be driven by infrastructure factors such as power availability, water access, and supplier networks rather than electoral geography. President Lee has defended the location by citing a government competition under the prior administration that rated the region highly and by arguing that decentralizing advanced industry investment is both a national economic strategy and a legitimate public policy goal.

Can South Korea actually build a new semiconductor fab cluster outside Seoul on this timeline?

Building a leading-edge semiconductor fab requires 200 to 400 megawatts of stable power per facility, ultrapure water in industrial volumes, an existing supplier ecosystem, and a trained technical workforce — none of which can be incentivized into existence quickly. Industry experts have noted that these requirements may not scale fast enough in a greenfield region like Honam to meet the current AI demand surge. The investments announced Sunday will realistically produce volume chip output from a new cluster in the late 2020s at earliest, making them most relevant to meeting AI demand in the 2030s rather than the current HBM supercycle.