South Korea bets $880 billion on AI chips: Samsung and SK Hynix to lock down the global supply chain

South Korea bets $880 billion on AI chips: Samsung and SK Hynix to lock down the global supply chain

🔎 $880 billion, a single obsession: avoid becoming an anonymous supplier

On June 29, 2026, the South Korean president unveiled the most ambitious industrial investment plan in the country's history: 1,350 trillion won, or approximately $880 billion, entirely dedicated to the AI chip value chain. This plan resembles nothing Seoul has done before. It is no longer about subsidizing R&D or offering targeted tax breaks. It is a brutal geopolitical repositioning, calibrated so that South Korea ceases to be a subordinate link in the supply chain and becomes the essential node between American design and Chinese manufacturing.

The announcement comes amid extreme tension. The United States is restricting advanced chip exports to Beijing. China is accelerating its own sector, as shown by the $2 billion raise by Moonshot AI whose model Kimi K2.6 dominates the open-weight. And American AI giants are beginning to bypass traditional foundries: OpenAI is developing its custom infrastructure chip "Jalapeno" with Broadcom, potentially marking the end of exclusive reliance on TSMC and Samsung for design.

Faced with this existential risk, Seoul is responding with money. A lot of money. With a clear goal: 10 GW of data center capacity by 2035, equivalent to the electricity consumption of several European countries combined, entirely dedicated to AI.

The essentials

• South Korea announces a 1,350 trillion won (~$880 billion) plan to dominate the AI chip supply chain, according to Reuters.
• Samsung and SK Hynix are the two pillars of the plan, with an ambition of 10 GW of data centers by 2035.
• The positioning is deliberately geopolitical: Seoul wants to be the indispensable buffer between American technology and Asian manufacturing capabilities.
• The plan reacts to OpenAI's verticalization threats (Jalapeno chip) and the rise of China (Kimi K2.6), which challenge the pure foundry model.
• South Korea is no longer content with just manufacturing chips: it is building the complete ecosystem, from HBM memory to computing infrastructure.

Recommended tools

The figures behind the plan: $880 billion broken down

$880 billion is approximately 40% of South Korea's 2025 GDP. It is more than the annual US defense budget. It is an amount that, in itself, redefines what an average state can mobilize for an industrial sector.

According to Al Jazeera, the plan is broken down into three major pillars. The first concerns chip manufacturing itself, with massive investments in Samsung Foundry's advanced fabs and SK Hynix's HBM lines. The second pillar targets computing infrastructure: data center construction, power grid connection, and cooling. The third pillar is the most often underestimated — it involves R&D on materials, lithography equipment, and a skilled workforce.

The target of 10 GW of data centers by 2035 is the most striking figure. For comparison, France consumes around 60 GW at peak times. South Korea plans to add the equivalent of a sixth of France's electricity consumption just for AI. This implies dedicated nuclear power plants, redesigned power transmission networks, and international energy agreements that the country had never envisioned on this scale.

The role of the state: subsidizing, not directing

Unlike China, where the state plans and assigns quotas, the South Korean model relies on massive tax incentives, loan guarantees, and special economic zones. The BBC notes that the government acts more like an investment banker than a central planner. Private companies — Samsung, SK Hynix, but also secondary suppliers — remain the decision-makers regarding the precise allocation of capital.

Samsung and SK Hynix: the HBM memory duopoly under pressure

Samsung and SK Hynix are not starting from scratch. Together, they control over 90% of the global HBM (High Bandwidth Memory) market, the essential component that accompanies every AI chip from Nvidia, AMD, and soon custom chips. It is this duopoly that allowed Micron to quadruple its Q3 2026 revenue with an 81% gross margin on HBM4 — a market so lucrative that even the third American player is raking in historic margins.

But the duopoly's comfort is starting to crack. HBM4, which enters mass production in 2026, requires such advanced packaging processes (hybrid bonding, 12-Hi and 16-Hi stacking) that barriers to entry are lowering for players like Micron, while clients like Nvidia are considering developing their own packaging. The $880 billion plan is precisely aimed at preventing this erosion.

Samsung Foundry: catching up with TSMC before clients leave

Samsung's foundry is the plan's weak point. The company has fallen behind TSMC in advanced nodes (3 nm, 2 nm), losing strategic clients to the Taiwanese giant. The arrival of custom chips like OpenAI's Jalapeno chip with Broadcom illustrates the danger: if major AI clients internalize design and freely choose their foundry, Samsung Foundry has no guarantee of contracts.

The government plan therefore includes specific subsidies for the foundry, with an undeclared but obvious goal: offering aggressive enough prices to convince AI chip developers not to go all-in on TSMC. It is a price war disguised as industrial policy.

SK Hynix: consolidating the HBM fortress

SK Hynix is in a more comfortable position. The company has been Nvidia's main HBM supplier since the HBM3 generation, and its technological lead in packaging remains significant. For the company, the government plan mostly means money to accelerate the transition to HBM4 and HBM4E, with production capacities tripled by 2028.

10 GW of data centers: the energy gamble that defies physics

The goal of 10 GW of data center capacity by 2035 is the boldest — and riskiest — step in the plan. To understand the scale, a modern 100 MW AI data center can host around 50,000 next-generation GPUs. Ten gigawatts is the capacity to host 5 million GPUs. That is more than the entire estimated global AI GPU fleet in early 2025.

The power grid problem

South Korea currently produces about 130 GW of electricity. Adding 10 GW of constant and growing demand within nine years requires not only new generation capacity, but above all a completely redesigned distribution grid. AI data centers can't just be plugged in anywhere: they require ultra-high-voltage substations, very high-bandwidth fiber optic connections, and access to water for cooling.

The plan calls for the construction of "AI clusters" along the west coast, near existing nuclear power plants and submarine cable landing points. It is an unyielding geographical logic: putting the chips as close as possible to energy and data.

Cooling as a bottleneck

Even with 10 GW of electricity, the cooling issue could derail everything. Next-generation GPUs each dissipate over 1,000 watts under load. Concentrating tens of thousands of these chips in a single building demands liquid cooling systems whose mechanical complexity rivals that of the chips themselves. No country has yet demonstrated the ability to deploy this level of infrastructure at this speed.

Geopolitics of chips: South Korea as the third pole

The South Korean plan can only be fully understood through a geopolitical prism. The semiconductor industry is currently structured around three poles: the United States for design (Nvidia, AMD, OpenAI, Google), Taiwan for foundry (TSMC), and China for assembly and domestic demand. South Korea occupies a hybrid position — it designs (Samsung), it manufactures (Samsung Foundry), and it produces critical components (SK Hynix). But it is not the absolute leader in any of these categories.

The $880 billion plan aims to change this equation. By investing simultaneously in foundry, memory, and infrastructure, Seoul hopes to become the only country capable of delivering a complete AI stack — from chip to data center — without depending on any other player.

Facing the United States: the reluctant ally

Relations with Washington are complex. The United States needs South Korea to diversify the supply chain away from Taiwan (the "Taiwan + 1" strategy). But at the same time, export restrictions imposed by Washington prevent Samsung and SK Hynix from selling their best chips in China, their largest market. The South Korean plan is partly a signal sent to Washington: if you block our access to the Chinese market, we must build alternative demand — hence the 10 GW of domestic data centers.

Facing China: the rising rival

China is no longer just a copycat. With models like Kimi K2.6 reaching 84 on the LLM benchmark and 88.1 in agentic self-host, Chinese companies are beginning to rival the best Western models. Moonshot AI, which raised $2 billion, illustrates this dynamic: China is building its complete AI ecosystem, from model design to infrastructure.

For South Korea, the threat is twofold. On the one hand, China could develop its own HBM supply chain and eliminate its dependence on SK Hynix. On the other hand, Chinese open-weight models like Kimi K2.6 reduce the competitive advantage of proprietary American models, which could ultimately decrease the demand for the most advanced chips produced by South Korea.

Facing Taiwan: the competitor to overtake

TSMC remains the undisputed leader of the foundry. But Taiwan's geopolitical vulnerability — threatened by a potential Chinese invasion — is pushing customers to seek alternatives. This is Samsung Foundry's historic opportunity. If the South Korean plan succeeds in closing the technological gap with TSMC on 2 nm and sub-2 nm nodes, Taiwan's "china risk" could become the driving force behind Samsung's reconquest.

The threat of custom chips: when clients become competitors

The South Korean plan is a reaction to a structural trend that directly threatens the business model of foundries: the verticalization of AI. The largest semiconductor clients no longer want to depend on third-party foundry roadmaps. They are designing their own chips.

OpenAI's Jalapeno chip, developed with Broadcom is the most striking example of this trend. OpenAI, which spends billions on compute using Nvidia chips, is now building its own silicon for inference. The promise is simple: a 50% reduction in cost per request. If this promise materializes, the entire margin model of traditional foundries will collapse.

Anthropic and OpenAI also invest on the enterprise side

Verticalization doesn't stop at silicon. Anthropic and OpenAI are each launching their own enterprise joint venture, mobilizing $10 billion to deploy AI in SMEs and large corporations. This movement means that model creators are also becoming infrastructure integrators, choosing which chips to use, where to host them, and at what price. The balance of bargaining power shifts definitively to the software side.

Groq and the neocloud model: another intermediary breaking free

Even players specializing in hardware acceleration are breaking free. Groq, after raising $650 million, pivots to the neocloud offering turnkey inference infrastructure. Every player in the chain is trying to capture more value, and the pure-play foundry — Samsung Foundry, TSMC — risks finding itself squeezed between increasingly powerful clients and equipment suppliers (ASML, Applied Materials) that are raising their prices.

This is the catastrophic scenario that the South Korean plan is attempting to avoid. By controlling data center infrastructure in addition to the foundry, South Korea offers an integrated package that no client can easily replicate on their own.

Global AI Infrastructure in 2026: Where Does South Korea Stand?

To measure the ambition of the South Korean plan, it must be compared to other global investments in AI infrastructure. The United States remains the undisputed leader, with hyperscalers (Microsoft, Google, Meta, Amazon) each investing between $50 and $80 billion per year in AI compute in 2026. China is catching up with massive state investments in data centers and local chips. Europe, with its GAIA-X project and national initiatives, is lagging behind.

South Korea is positioning itself differently. Rather than relying on software giants like the United States, it is betting on physical infrastructure as a strategic product. The idea is to become the "docks" of global AI — the place where models, whether American, Chinese, or European, come to run on South Korean chips and in South Korean data centers.

The models that will power these data centers

There will be no shortage of demand for these 10 GW. Current models illustrate a race toward complexity that demands ever more compute. Google's Gemini 3.1 Pro scores 92, OpenAI's GPT-5.5 reaches 98.2 in agentic, and Anthropic's Claude Opus 4.7 scores 94.3 in the same category. Each additional benchmark point translates into exponential infrastructure needs.

The trend toward adaptive models like Claude Opus 4.7 (Adaptive), which dynamically adjusts its compute usage based on task complexity, could slightly moderate demand. However, deep reasoning models like Gemini 3 Pro Deep Think (95.4 in agentic) compensate for this efficiency with much higher compute requirements during "thinking" phases.

The agentic model market: the real driver of demand

The June 2025 agentic ranking reveals that models are increasingly specializing in the autonomous execution of tasks. GPT-5.5 at 98.2, Gemini 3 Pro Deep Think at 95.4, Claude Opus 4.7 at 94.3 — these scores are not benchmark artifacts. They represent models capable of planning, executing, and correcting complex task chains without human supervision. And each agentic execution consumes significantly more tokens — and therefore more GPU-seconds — than a simple question-and-answer exchange.

South Korea is betting that this explosion in agentic demand will create an infrastructure market so vast that even 10 GW will not be enough. It is a risky but rational bet: if agentic becomes the dominant mode of interaction with AI by 2030, compute demand could be 10 to 100 times higher than in 2025.

The financing: who is really paying the 880 billion?

An $880 billion plan doesn't just spring out of the ground by magic. The financing structure reveals the true priorities of the South Korean government.

Private companies: 70% of the total

According to details reported by Reuters, about 70% of the investment will come from the companies themselves — Samsung, SK Hynix, and their supplier ecosystems. This represents about $616 billion over four to five years. For Samsung, whose annual revenue is around $200 billion in 2025, this is a considerable commitment but not impossible, especially when spread over several years.

The State: 30% in subsidies and guarantees

The remaining $264 billion comes from the State, in the form of direct subsidies, tax cuts, and loan guarantees. The mechanism is crucial: rather than financing factories directly, the government guarantees corporate loans, lowering their cost of capital. This is a model that already proved itself in the 1980s when South Korea built its semiconductor industry from scratch.

The risk of overcapacity

The obvious danger of this plan is overcapacity. If AI demand slows down — whether due to technical limits, regulations, or an "AI winter" — South Korea will find itself with underutilized factories and data centers, and a heavier public debt. This is exactly what happened with the LCD display industry in the early 2010s, when Korean and Chinese overcapacity caused prices to collapse.

Jobs and training: talent as the ultimate bottleneck

The most often overlooked aspect of industrial plans of this scale is the workforce. Building advanced semiconductor factories and data centers spanning several hundred megawatts requires thousands of specialized engineers—profiles that South Korea does not produce in sufficient numbers.

The plan therefore includes a massive education component: creation of new university programs, partnerships with companies for work-study training, and targeted immigration programs to attract foreign engineers. The government aims to train 150,000 semiconductor specialists by 2030, a figure that remains ambitious given South Korea's birth rate—the lowest in the world in 2025, at 0.18 children per woman according to World Bank data.

Global competition for talent

South Korea is not the only country looking for semiconductor engineers. The United States, Taiwan, Japan, and China are all competing for the same profiles. The salaries of engineers in lithography, advanced packaging, and chip design have increased by 30 to 50% between 2023 and 2026 according to the industry. The South Korean plan will have to include globally competitive compensation, otherwise the factories will remain empty even if they are built on time.

HBM memory: why it's the true strategic lever

Among all segments of the supply chain, HBM memory is the one where South Korea holds the greatest power. HBM is not a standard component — it's the memory that directly powers AI chips, and its performance determines the energy efficiency and speed of the entire system.

The evolution of HBM in 2026

The HBM4 generation, which enters mass production in 2026, marks a significant technological leap. With bandwidth throughput exceeding 1.5 TB/s per stack and capacities reaching 48 GB per module, HBM4 allows AI chips to operate at speeds impossible with the previous generation. Micron illustrated this potential with gross margins of 81% in Q3 2026, proving that demand far exceeds supply.

SK Hynix and Samsung share the bulk of this windfall. The $880 billion plan aims to accentuate this dominance by investing heavily in HBM4E (the next generation) and developing exclusive packaging techniques that will bind HBM memory directly to the logic chip, making replacement by a competitor much more difficult.

The risk of commoditization

The danger for South Korea is that HBM becomes a commodity — a standardized product where price takes precedence over technical differentiation. This is what happened with standard DRAM in the 2010s, when prices collapsed due to Chinese overcapacity. The South Korean plan attempts to avoid this scenario by maintaining a two-generation technological lead over competitors, rendering any attempt to catch up economically unrealistic.

❌ Common mistakes

Mistake 1: Confusing this plan with a simple grant plan

This is not a classic grant plan where the government hands out money and hopes for the best. It is a leveraging plan: the government mobilizes 264 billion to trigger 616 billion from the private sector. The difference is fundamental. If private companies do not follow suit, the plan collapses. The risk is borne primarily by Samsung and SK Hynix, not by the taxpayer.

Mistake 2: Thinking that 10 GW of data centers is realistic in 9 years

The target of 10 GW by 2035 is likely a political PR figure. No country has ever added 10 GW of data centers in nine years. South Korea itself only has 130 GW of total electrical capacity. The real target is probably between 3 and 5 GW, with the rest serving as a signal to markets and competitors.

Mistake 3: Ignoring the Taiwan risk

Many commentators present this plan as an offensive against China. In reality, the main competitor targeted by South Korea is Taiwan, and more specifically TSMC. If TSMC maintains its 18- to 24-month lead on advanced nodes, no amount of investment will be enough for Samsung Foundry to regain market share. The South Korean plan is a bet on Taiwan's geopolitical vulnerability as much as on technology.

Mistake 4: Underestimating the energy issue

Building data centers is one thing. Powering them is another. South Korea still relies massively on fossil fuels for its electricity. Deploying 10 GW of data centers powered by coal and gas would be a climate disaster and incompatible with the country's carbon neutrality commitments. The plan therefore implies a massive nuclear pivot, with small modular reactors (SMRs) that do not yet exist on a commercial scale.

❓ Frequently Asked Questions

Why $880 billion and not a rounder number?

The plan is denominated in wons (1,350 trillion wons), which equates to approximately $880 billion at the June 2026 exchange rate. The amount is not rounded because it results from the sum of specific commitments from various companies and government agencies.

Can South Korea really compete with TSMC in foundry?

On paper, Samsung Foundry's technological lag behind TSMC is around 18 months on advanced nodes. Money can accelerate the catch-up, but TSMC is also investing massively. Seoul's real hope lies in Taiwan's "china risk," which pushes customers to diversify their foundries, even at the cost of a slight loss in performance.

What happens if AI demand slows down?

This is the main risk of the plan. Semiconductor and data center overcapacity would lead to a price collapse, massive losses for Samsung and SK Hynix, and a waste of public funds. The government is betting on the fact that agentic demand — which is exploding with models like GPT-5.5 scoring 98.2 on the agentic score — will continue to grow exponentially.

Do Chinese models like Kimi K2.6 directly threaten this plan?

Indirectly, yes. If Chinese open-weight models like Kimi K2.6 become performant enough to replace American proprietary models, demand for the most advanced chips (the ones South Korea manufactures) could decrease in favor of mid-range chips produced in China. This is the technological "de-moat" scenario that Seoul fears the most.

Is this plan enough to ensure South Korea's technological sovereignty?

No. Total sovereignty is illusory in semiconductors. South Korea depends on ASML (Netherlands) for EUV lithography, Applied Materials (United States) for deposition equipment, and JSR (Japan) for chemicals. The plan reduces dependencies, but does not eliminate them.

✅ Conclusion

South Korea is not bluffing: $880 billion is a commitment that structurally transforms the country's economy for the decades to come. The success or failure of this plan will determine whether Seoul becomes the world's third AI hub — between American design and Chinese manufacturing — or whether it remains a component supplier whose added value will gradually be compressed by the vertical integration of software players. To follow the evolution of these positions, check out our ranking of the best AI tools and our analysis of current AI trends.