South Korea’s AGI Crossroads: Leapfrogging or Dependency

January 2026, the World Economic Forum in Davos. Google DeepMind CEO Demis Hassabis floated a proposal.

“We need to build an institution for AGI — one that does collaborative research like CERN and international oversight like the IAEA.” He invoked the names of two organizations humanity created to govern itself during the nuclear age, and called it a ‘Technical UN’ (Fortune).

On the same stage, Yuval Noah Harari put it more bluntly. “Most countries have no say in the design of AI systems. AI is made in a few places in the US and China, and then spreads to the rest of the world.” (WEF Davos 2026) Countries without a voice become dependent.

Is South Korea one of those voiceless “rest of the world” countries, or the exception? Answering that requires looking at three cards and three barriers simultaneously.

Today, we dissect the structural strengths South Korea holds in the AGI value chain — and the structural weaknesses eating away at those very strengths. This is also the final installment in the series — connecting the dots from the Hassabis vs. Amodei debate covered in The Day After AGI to what it means for South Korea specifically.

The Cards South Korea Holds

Semiconductors: The HBM Dominance

If you had to name one critical component for AGI training, it would be HBM — High Bandwidth Memory. It’s the semiconductor that enables AI to process massive datasets at extreme speeds. If the GPU is the AI’s brain, HBM is its short-term memory — the bigger and faster it is, the better AI performs.

South Korea’s share of this market is overwhelming. As of 2025, SK Hynix holds roughly 62% and Samsung Electronics about 17%. Combined: 79%. US-based Micron takes the remainder (Global Economic). NVIDIA builds the engines, but the memory inside those engines doesn’t exist without South Korea.

On February 12, 2026, Samsung Electronics officially announced the world’s first mass production of HBM4, achieving data transfer speeds of 11.7Gbps — exceeding NVIDIA’s 11Gbps requirement. UBS estimates SK Hynix will capture roughly 70% of the HBM4 market in 2026 (SK Hynix Newsroom). Bank of America projects the total HBM market to grow 58% year-over-year to $54.6B in 2026.

The numbers alone don’t tell the full story. The massive clusters required for AGI training — tens of thousands of GPUs powering something like GPT-5 — all run on HBM. Without Korean companies, neither the US, China, nor any country can advance toward AGI. This isn’t just market share. It’s strategic supply leverage.

As TheByteDive previously analyzed in AI-Era Leapfrogging, South Korea has the DNA of ‘path-creating’ strategies — it achieved the world’s first CDMA commercialization and executed ‘path-skipping’ in D-RAM by leaping over intermediate generations (Lee & Lim, 2001). HBM is an extension of that DNA.

The risks are real, though. Samsung Electronics’ HBM4 yield rates reportedly haven’t reached sufficient levels yet. SK Hynix leads on cost, Samsung on performance — the race is still on. And Micron, the American competitor, keeps closing the gap.

Energy: The Nuclear Renaissance

The second bottleneck in the AGI infrastructure war is power. Data centers are energy gluttons. We’ve entered an era where training a single AI model consumes as much electricity as a small city uses in a year.

The numbers first. South Korea’s data center power demand is projected to increase 3.7x, from 8.2 TWh in 2025 to 30 TWh by 2038. Experts estimate that meeting this demand requires building three additional large-scale nuclear reactors (Korea Power Supply Expert Estimate). 2038 may seem distant, but considering that building a single reactor takes over 10 years, this is a decision that needs to be made now.

This is where South Korea’s hidden card emerges. Despite going through a phase of nuclear phase-out policy, the country still maintains world-class nuclear design, construction, and operations capabilities. The UAE Barakah and Czech Dukovany reactor contracts prove it. While the US and Europe let these capabilities atrophy by not building reactors for decades, South Korea kept accumulating real-world experience.

Energy is really the core issue in the competition for AI data center locations. Currently, 75% of domestic data centers are concentrated in the Seoul metropolitan area, where power self-sufficiency is just 11.6% — the lowest nationwide (Electimes). If this structure persists, large-scale AI data centers will inevitably migrate outside the capital or overseas. That’s why a strategy of building AI infrastructure clusters near nuclear plants makes sense.

Comparing with global competitors makes South Korea’s position clearer.

South Korea’s most strategic move on the energy front is clear: SMR (Small Modular Reactor) development + data center power package exports — bundling nuclear exports with AI infrastructure exports. There are already signs of movement in this direction.

Digital Government: The #1 E-Government Asset

The third card is the most underrated. South Korea scored 0.94 on the OECD Digital Government Index (DGI), ranking #1 for two consecutive terms. Its government AI maturity score stands at 0.89 — nearly double the OECD average of 0.53 (OECD Digital Government Review). Even second-place France can’t keep up.

Why does this matter in the AGI era? Government services are among the fastest domains for AI adoption — tax processing, permit issuance, welfare distribution. When all of these are automated by AI, countries with existing digital foundations hold an enormous advantage. Just as Estonia leapfrogged with digital governance, South Korea’s e-government infrastructure dramatically lowers the cost of transitioning to AI-driven government.

The corporate numbers back it up too. Korean companies’ AI adoption rate stands at 30.28%, the highest among OECD nations (second-place Denmark: 27.58%). Microsoft ranked South Korea as the “#1 country in AI adoption speed” in 2025 (AI Times).

The real value of being #1 in digital government is the data infrastructure. AI trains on data. Data comes from digitized administrative systems. South Korea already has this pipeline running. Not leveraging it is like sitting on a gold mine and refusing to dig.

The Barriers

The AI Governance Vacuum

South Korea enacted its Framework Act on AI on January 22, 2026 — becoming the second country after the EU to establish a comprehensive AI regulatory framework. It created the National AI Strategy Committee, allocated a 2026 AI budget of ~$7B (10.1 trillion won), and finalized an AI Action Plan with 99 implementation tasks and 326 policy recommendations (MoneyToday).

The numbers are impressive. The problem is execution.

The Information Technology and Innovation Foundation (ITIF) identified structural flaws in South Korea’s AI Framework Act: “Broad definitions, prescriptive R&D obligations, SME bias, and inefficient regulatory mechanisms.” (ITIF Report) Because the law is designed as a monolithic package, effective provisions get locked in alongside inefficient ones.

There’s a more fundamental issue. South Korea’s governance debate is still trapped in the “promotion vs. regulation” binary. The Ministry of Science and ICT says it prioritizes “promotion over regulation,” but there is zero strategy for participating in international governance discussions around AGI-level risks. The law exists, but there’s no roadmap connecting it to the kind of international AI institution Hassabis proposed.

Think of it this way: the country has installed fire extinguishers inside the house, but has no communication line to the fire department. It can handle small fires, but when a major blaze breaks out, it lacks the capacity to mobilize international coordination.

The Rigidity of Chaebol SI Structures

The second barrier choking South Korea’s AI transformation is the chaebol-linked SI (Systems Integration) structure. Group IT service companies like Samsung SDS, LG CNS, and SK C&C dominate a large share of Korean enterprise IT.

The contrast with Palantir‘s FDE (Forward Deployed Engineer) model makes the gap obvious. Palantir FDEs embed at client sites, directly solve on-the-ground problems, and feed those experiences back into the product to evolve the entire platform. Korean SI firms operate on a ‘contract labor’ model — they receive requirements and execute projects. The goal isn’t deep understanding of the client’s problem; it’s delivering to spec, on time.

Under this structure, AI gets deployed as a ‘cost-cutting tool,’ not an ‘innovation engine.’ The data shows it. 82.3% of Korean manufacturers don’t use AI at all. 74% of companies say “the cost of AI transformation is prohibitive.” 61% believe the impact would be “minimal” (Electronic Times).

The high dependence on intra-group transactions also stifles innovation. When captive group contracts are guaranteed, there’s little incentive to build genuine market competitiveness. By the time global AI platform companies start directly capturing Korean enterprise IT, it’ll be too late. Large SI firms are declaring AI transformation pivots — but declarations don’t change outcomes if the underlying philosophy stays the same.

Brain Drain

The third barrier is the quietest and most lethal. South Korea’s AI talent is leaving.

The numbers first — AI talent net inflow reversed from +0.23 (ranked 14th) in 2020 to -0.36 (ranked 35th) in 2024. South Korea has become one of the worst brain-drain countries in the OECD (Hankook Ilbo). The IMD brain drain index plummeted from 28th in 2020 to 48th in 2025. Around 11,000 Korean AI professionals currently work overseas — 16% of the total AI workforce.

Why are they leaving? A STEM PhD with 10 years of experience earns ~$70K domestically. The same person earns ~$280K abroad (Seoul Shinmun). A 4x gap. And 76.9% of the top 1% of science students are choosing medical school. The best minds capable of AI research are picking a different path from the start.

By 2029, South Korea is projected to face a shortage of 580,000 workers in AI and emerging tech fields (NewsSpace). The HBM fabs exist, but the brains to design next-generation HBM won’t be there.

The paradox is stark. South Korea holds the #1 digital government ranking and 79% of the HBM market — yet the talent needed to sustain and advance those strengths is walking out the door. The country is building infrastructure while losing the people who would run it.

Hassabis’ International AI Governance and South Korea

Back to Davos. The CERN + IAEA model Hassabis proposed demands two things simultaneously. First, collaborative research — the world’s best minds working together on the final steps toward AGI. Second, international oversight — agreeing on minimum standards so no single actor develops AGI unilaterally, much like the nuclear regime (New Savanna Blog).

Hassabis himself acknowledged the limitation: “Even if one company, one country, or the West chooses this direction, it won’t work unless the entire world agrees on minimum standards.” Making this proposal real requires the participation of key countries beyond the US and China.

This is where South Korea’s position becomes critical. South Korea can be more than a mere ‘participant’ in this discussion — it has the conditions to be a player with genuine bargaining power.

• Semiconductor supplier: AGI training is impossible without HBM. South Korea sits in a position to influence supply coordination.
• Non-aligned nation: It trades with both the US and China without being fully tilted toward either side.
• Digital governance leader: It has the credentials to propose international standards in AI-driven government.

Harari’s warning deserves a re-read. “AI immigrants travel at the speed of light, and their loyalty lies with the country that created them.” If AI enters Korean society designed in the US or China, then that AI’s value system, language priorities, and cultural assumptions belong to somewhere else — not Korea. This isn’t just technological dependency. It’s cultural dependency (Singjupost).

So South Korea has two reasons to join the governance table Hassabis proposed. First, practical interest — as an HBM supplier, a seat at the international AI governance table means a voice in chip export controls, AI safety standards, and semiconductor regulation. Second, survival — if you’re not at the table, the rules made at that table come for your companies.

The question is whether South Korea is building the governance capacity to participate. Right now, it isn’t. The AI Framework Act exists, but an international AI governance diplomacy strategy does not.

Conclusion: Time to Choose

The historical formula for leapfrogging has three ingredients: absence of legacy systems, access to new technology, and people who know the direction. Kenya leapfrogged to mobile finance because it lacked banking infrastructure. Estonia went straight to digital governance because it had no landlines.

South Korea’s situation is different. It has too many legacy systems. Chaebol SI firms dominate IT services. STEM talent drains to medical school and overseas. AI governance discussions are stuck in the promotion-vs.-regulation binary. What the country already has is paradoxically what’s blocking its leap.

Yet at the same time, it holds stronger cards than at any previous point in history. 79% HBM share is bargaining power. The #1 digital government ranking slashes the cost of transitioning to AI-driven governance. Nuclear capabilities offer a real path to solving the AI data center power problem.

Having cards doesn’t win the game. Knowing when to play them does.

Bottom Line. South Korea holds real cards in the AGI value chain. The problem is it doesn’t yet have a strategy for playing them.

What This Means for You. You don’t need to wait for national strategy. Ask right now: “Is AI an innovation engine or a cost-cutting tool at my company?” If your organization only deploys AI to cut costs, your position may end up on the list of costs AI eliminates. The rigidity of Korea’s SI structure is a career risk that hits individuals directly. Move like an FDE before your organization changes — deeply understand the customer’s problem, build solutions yourself, evolve through that experience. That’s individual-level leapfrogging in the AGI era.

References

1. Fortune — AI luminaries at Davos clash over how close human-level intelligence really is (2026.01.23)
2. WEF / Singjupost — Yuval Noah Harari’s Remarks at WEF Davos 2026
3. New Savanna Blog — Hassabis on the future of AI: A CERN for AI
4. SK Hynix Newsroom — 2026 Market Outlook: The HBM-Led Memory Supercycle
5. Global Economic — Samsung · SK Hynix Projected for $180B Jackpot in 2026
6. EconMingle — SK Hynix HBM Market Share 57%, Tech Gap Analysis
7. Electimes — AI Power War: Data Centers as Core National Infrastructure
8. Daum News — Data Centers Need Three More Nuclear Reactors
9. OECD — Digital Government Review of Korea
10. Global Government Forum — South Korea tops OECD Digital Government Index
11. AI Times — Microsoft: “South Korea ranked #1 in AI adoption speed in 2025”
12. MoneyToday — Leap to AI Top-3: AI Action Plan Finalized
13. ITIF — South Korea’s AI Framework Act: Promotion, Strategy, and Regulatory Risk (2025.09)
14. Electronic Times — 82.3% of Korean Manufacturers: “We Don’t Use AI”
15. ZDNet Korea — Chaebol SI Firms Bet Their Survival on AI
16. Hankook Ilbo — Korean AI Talent Net Outflow Hits OECD Bottom
17. Seoul Shinmun — $70K vs $280K — Top 1% STEM Brains Flock to Med School
18. NewsSpace — K-Brain Drain Alert: 580K AI Talent Shortage by 2029
19. Introl Blog — South Korea’s $735B Sovereign AI Initiative
20. Chatham House — A ‘CERN for AI’ — what might an international AI research organization address?
21. Lee, K. & Lim, C. (2001) — Technological regimes, catching-up and leapfrogging (Research Policy) — cited in AI-Era Leapfrogging
22. The Day After AGI — Hassabis CERN+IAEA proposal in detail, Korean semiconductor analysis (Items 33, 45-46)