The Infrastructure Is Laid, But There Are No Players — What Is Missing From Korea’s AI Policy

A 10.1 trillion won budget, 260,000 GPUs, an AI Framework Act now in force. The numbers look impressive. So where is the Korean frontier AI model? The Korean AI unicorn? The Korean AI service used worldwide? “The foundations are laid, but there is no team on the field.”

1. What Korea Is Putting in Place — The Facts

Start with the numbers.

The AI Framework Act

Passed by the National Assembly in December 2024 and in force from January 2026. Its philosophy is promotion-first, minimum regulation. It imposes five obligations — transparency, safety assessment, explainability, human oversight, and non-discrimination — on high-impact AI (medicine, judiciary, employment, and the like) and leaves the rest to self-regulation. It is a deliberate contrast to the comprehensive regime of the EU AI Act.

Infrastructure

  • 2026 AI budget of 10.1 trillion won — roughly triple the prior year (Ministry of Science and ICT).
  • Goal of securing 260,000 GPUs for national AI compute.
  • Haenam National AI Computing Center: a 2.5 trillion won investment, targeted as Korea’s version of ABCI.
  • A plan for 76 AI data centers by 2028.

Talent

  • 1.4 trillion won for AI talent development.
  • A new AI School at KAIST.
  • A target of one million people trained in AI over five years.

Visas

  • K-STAR visa: 400+ people per year, drawing overseas talent in AI, semiconductors, and other frontier fields.
  • Expanded Top-Tier visa and a new E-7-M category for specialists in advanced industries.

Semiconductors

  • HBM (High-Bandwidth Memory): Samsung Electronics and SK Hynix hold 80–90% of the global market.
  • Domestic NPUs (Neural Processing Units): 1.27 trillion won invested to build the K-NPU software ecosystem.

By the numbers, it does not look bad. A pro-innovation AI Framework Act, a budget in the 10-trillion-won range, HBM leadership in semiconductors. So why has no OpenAI, no DeepSeek, not even a Sakana AI emerged from Korea?

2. So What Is Missing — The Core Analysis

The thesis: Korea has “infrastructure projects,” not an “ecosystem.”

The infrastructure is laid, but there are no players competing on top of it. Five gaps stand out.

(1) No Frontier Models

OpenAI (GPT-5), Anthropic (Claude), Google (Gemini), Meta (Llama), DeepSeek (V3/R1), Mistral — the global AI race is fundamentally about who builds frontier models. Korea does not have one.

Naver’s HyperCLOVA X and Samsung Gauss exist, but neither is competitive on global benchmarks. Japan, by contrast, has SoftBank’s Sarashina, NTT’s Tsuzumi, and research models out of Sakana AI, and has even stood up a domestic foundation-model joint venture across ten companies.

Without models, infrastructure is just machinery that consumes electricity.

(2) A Shallow Venture Capital Ecosystem

US AI VC investment in 2025 was **159billion∗∗,accountingfor79159 billion**, accounting for 79% of global AI investment (Crunchbase, 2025). OpenAI's valuation: 300 billion. Anthropic’s: $183 billion. Mega-deals like these are what spin the US AI ecosystem.

Korea has stood up a 150 trillion won national growth fund, but GPs (fund managers) keep reporting the same thing: “there is not enough to invest in.” Government capital is present, but the depth and scale of AI startups that could absorb it is not. This is not a capital problem; it is an ecosystem problem.

(3) A Broken Research-to-Startup Pipeline

The strongest engine in the US AI ecosystem is the researcher-to-founder pipeline.

  • Researchers from Google Brain → co-founded OpenAI.
  • Researchers from Meta FAIR → spun out into Anthropic, Cohere, Adept, and others.
  • A researcher from Google DeepMind → founded Sakana AI in Tokyo.

How many startup founders have spun out of Samsung Research, Naver AI Lab, or ETRI? A culture in which joining a big-company research lab is a one-way door, the lack of institutional support for spinouts, and the social stigma around failure all clog this pipeline.

(4) Low Enterprise AI Adoption

OECD data (2024) lays out the reality.

  • Korean large enterprise AI adoption: 9.2%.
  • Korean mid-sized enterprise AI adoption: 2.9%.
  • OECD average: 20.2%.

Korea has bought GPUs, trained workers, and even passed laws, but the share of companies actually using AI is not even half the OECD average. The gap between infrastructure and on-the-ground adoption is that wide. It is like building a highway nobody drives on.

(5) The “Indiscriminate Training” Problem

One million people trained in AI over five years. Impressive in volume. But what actually produces AI leaders is not training programs — it is real projects and real startup experience.

Ilya Sutskever (OpenAI co-founder) came up through a graduate lab, Dario Amodei (Anthropic CEO) through Google Brain, David Ha (Sakana AI co-founder) through Google Japan, each of them founding companies after accruing operational experience at the highest level. What made them was not a certificate of completion but world-class research environments and real projects. Teaching a million people Python and TensorFlow is a fundamentally different problem from producing a single frontier AI researcher.

3. How Other Countries Are Playing — A Comparison

The United States: The Power of the Ecosystem

In the US, the government lays the infrastructure and the private sector plays the game.

  • $159 billion in 2025 AI VC investment — 79% of global AI investment (Crunchbase).
  • OpenAI: $300 billion valuation; GPT-4/o series setting global AI standards.
  • Anthropic: $183 billion valuation; Claude attacking the enterprise AI market.
  • CHIPS Act: 52.7billionofpublicmoney→triggering∗∗over52.7 billion of public money → triggering **over 1 trillion in private investment.**
  • Regulation: not the EU’s ex ante regime but a flexible executive-order approach.

The key point: the government’s role stays at “infrastructure plus minimum regulation,” while actual AI innovation is led by the private sector. OpenAI, Anthropic, Google, Meta, xAI — every frontier-model developer is private.

China: A State-Enterprise Integrated Model

China is pushing through algorithmic efficiency even under US chip export controls.

  • DeepSeek: developed V3/R1 under chip restrictions, achieving competitive performance at $5.57 million in training cost — roughly 1/20 of GPT-4.
  • 15 “national AI teams” designated — Baidu, Alibaba, Tencent, Huawei and others groomed as champions in their respective domains.
  • 535 universities with AI majors and 3.57 million STEM graduates per year — five times the US scale.
  • Super-app ecosystem: WeChat, Alipay, and others put AI services in front of 1.4 billion people instantly.

The core lesson is the capacity to turn constraint into innovation. With chips scarce, algorithmic efficiency was pushed to the limit; with a vast domestic market, services can be validated at home long before going global.

Japan: A Full-Stack AI Sovereignty Strategy

Japan is the most useful comparison for Korea. The two countries passed AI legislation at similar times and announced government investments of similar scale. But Japan has something Korea lacks: corporate-level AI strategy.

Law and Budget

  • AI Promotion Act (passed May 2025) and AI Basic Plan (issued December 2025) — roughly the same timing as Korea.
  • Infrastructure investment: 10 trillion yen (65billion)fromthegovernmentplusover65 billion) from the government plus over 70 billion from the private sector — about 10x Korea’s level.
    • Among others: AWS’s 15.2billionJapaninvestment,Microsoft′s15.2 billion Japan investment, Microsoft's 2.9 billion, SoftBank’s Sakai data center.
    • ABCI 3.0: 6,128 H200 GPUs, 6.22 exaflops — a national AI supercomputer.

The Key Difference — Companies Are Executing Full-Stack AI Strategies

SoftBank — going all in for the AI era

  • Project Izanagi: a $100 billion AI chip venture, targeting custom AI silicon shipments by late 2026.
  • Stargate JV: a $500 billion AI data-center partnership with OpenAI and Oracle.
  • SB OAI Japan: a JV with OpenAI for Japanese-language-tuned AI services.
  • Sarashina LLM: an in-house large Japanese-language model.
  • Masayoshi Son’s declaration: “SoftBank will become an AI company.”

NTT — lightweight AI married to an optical network

  • Tsuzumi 2: a lightweight LLM that runs on a single GPU, aimed at cost-efficient enterprise deployment.
  • IOWN: next-generation optical-network communications infrastructure, targeting a 100x improvement in power efficiency for AI workloads.

Fujitsu — sovereign AI servers built domestically

  • Domestic production of AI servers begins March 2026.
  • Takane LLM: an in-house large language model aimed at enterprise AI.
  • The Fujitsu Kozuchi AI platform accelerates enterprise adoption.

Sakana AI — Japan’s first AI unicorn

  • Founded in Tokyo by David Ha (formerly Google) and Llion Jones (co-author of the Transformer paper).
  • 479millionraised,479 million raised, 2.65 billion valuation — the first AI unicorn in Japanese history.
  • Differentiated research grounded in an “Evolutionary AI” approach.

A National AI Company

SoftBank, Preferred Networks, and others — ten companies have jointly established a corporate entity to develop a domestic foundation model. The government is backing it with 1 trillion yen over five years. Korea has no equivalent corporate consortium plus government support model.

GENIAC

METI’s (Ministry of Economy, Trade and Industry) AI accelerator program. It selects 30 projects, gives them access to ABCI 3.0 infrastructure, and supports commercialization. The point is that it doesn’t stop at laying infrastructure — it actively cultivates the players who will compete on top of it.

Physical AI — Differentiating Through Robotics

38% of the world’s industrial robots are Japanese-made (IFR, 2024). Japan is pursuing a “Physical AI” strategy that fuses this hardware advantage with AI. If beating the US and China at software AI is hard, robotics + AI fusion becomes the angle of differentiation.

Semiconductors

Rapidus: targeting 2nm mass production by 2027. Technology partnerships with IBM and IMEC, foundry under construction in Chitose, Hokkaido. A clear intent to secure cutting-edge AI silicon process technology on Japanese soil.

4. What Korea Should Learn — From “Infrastructure” to “Ecosystem”

The formula is simple.

Real AI competitiveness = Infrastructure × Frontier models × Venture ecosystem × Enterprise adoption.

It is multiplication. A zero anywhere zeroes out the whole product.

ElementUSChinaJapanKorea
Infrastructure◎○◎○
Frontier models◎◎△→○×
Venture ecosystem◎○△→○△
Enterprise AI adoption◎○○×
Regulatory environment◎○○○

Korea keeps pace on infrastructure (○) and regulation (○) but has critical gaps in frontier models (×) and enterprise AI adoption (×).

Four Concrete Proposals

1. A National Frontier-Model Project

Korea should borrow from Japan’s “national AI company” model. Samsung, Naver, Kakao, and SKT should establish a joint K-Foundation Model entity, with the government supplying infrastructure (GPUs, data centers) and capital. Pooling resources to produce one globally competitive model is more realistic than each company building HyperCLOVA and Samsung Gauss in isolation.

2. Incentives for Big-Company-Research-Lab Spinouts

Sakana AI became a $2.65 billion unicorn after a Google researcher founded a company in Japan. Korea needs institutional support so that researchers from Samsung Research, Naver AI Lab, and ETRI can do the same. Specifically:

  • Loosen non-compete clauses for founders coming out of big-company labs.
  • Offer incentives (such as tax breaks) for parent companies that invest in and partner with spinouts.
  • Reinstatement guarantees for founders whose spinouts fail (already in place at some Japanese firms).

3. Mandates and Incentives for Enterprise AI Adoption

The 9.2% large-enterprise and 2.9% SME adoption rates must be raised to the OECD average of 20% within five years. How:

  • Expand SME AI adoption subsidies (cover 50–70% of adoption costs).
  • Award preferential points to AI-using firms in public procurement.
  • Build sector-by-sector AI adoption roadmaps with consulting support.
  • Operate a platform for sharing success cases.

4. Foster Mega-Deal AI VC Funds

Korea needs specialized AI venture funds at the 10-billion-won-plus scale. The government participates as LP, while GPs are private. Korea’s problem today is not a shortage of money; it is that professional GPs capable of running mega-deals and the startups worth those deals are simultaneously in short supply. A Yozma-style model (government matching + private operations), specialized for AI, like Israel’s, is worth adapting.

5. From AI Developer to AI Venture Founder — A Personal Growth Roadmap

If you cannot wait for policy, you have to move yourself. And right now is the lowest-barrier moment in history to start an AI company.

The Rise of the Solo-Founder AI Startup

The numbers tell the story.

  • Solo-founder share: 17% (2017) → 36% (2024). One-person startups in AI have more than doubled (PitchBook).
  • Productivity from AI tools: a 55% boost to developer productivity (GitHub Copilot Research, 2024) — small teams can now ship products that once required large ones.
  • 19% of early-stage AI funding is going to solo founders (Crunchbase).

In an era when AI tools turn one developer into the productive equivalent of ten, all you need is the idea and the execution to build a global AI product on your own.

A Practical Growth Path

Step 1: Technical depth — find domain-specific problems

Building general-purpose AI models is OpenAI or Anthropic territory. The opportunity for individuals lies in specific domains (healthcare, finance, manufacturing, law) where AI can deliver a 10x improvement. Opportunity sits at the intersection of domain knowledge and AI craft.

Step 2: Product building — minimum-headcount MVPs

Use AI to build an MVP with a 1–3 person team and validate the market fast. With Claude, GPT-4, Cursor, Replit Agent and the like, work that once required a team can now be done alone.

Step 3: Global markets — look outward from day one

Korea’s domestic market is small. For an AI SaaS, target Japan, Southeast Asia, and the English-speaking world from the beginning. Japan’s demand for AI adoption is exploding while its AI talent is short — an opening for Korean AI developers. Southeast Asia is in the middle of a fast digital transformation.

Step 4: Build a global network

  • Japan: Sakana AI, the GENIAC program, the SoftBank Vision Fund ecosystem.
  • United States: Y Combinator, a16z, Sequoia and other top-tier VCs and accelerators.
  • Singapore: Southeast Asia’s AI hub, with proactive government policies for attracting AI startups.

Step 5: Reach global capital

Use Korean government programs (K-Startup Grand Challenge, TIPS) as well — but also apply directly to global accelerators. Y Combinator, Japan’s GENIAC, Singapore’s IMDA AI Verify program. Given that Korean VC’s AI investment volume is less than 1% of the US figure, accessing global capital is not optional. It is required.

6. Closing: A Country That Builds Highways But No Cars

This is not a denial of the Korean government’s effort. The pro-innovation AI Framework Act, the 10-trillion-won budget, HBM leadership — all of it is necessary. But not sufficient.

Look at Japan. As the government laid infrastructure, SoftBank pushed a $100 billion AI chip project, NTT shipped a lightweight LLM, Sakana AI became a unicorn, and ten companies formed a domestic foundation-model entity. Government and industry are building the full stack (chips → models → services) together.

Look at the US. As the government laid infrastructure through the CHIPS Act, the private sector poured over $1 trillion in, OpenAI/Anthropic/xAI built frontier models, and thousands of AI startups built services on top. An ecosystem is running on top of the infrastructure.

Look at China. Even under chip constraints, DeepSeek built a frontier model through algorithmic efficiency. They possess the capacity to convert constraint into innovation.

Korea? It has laid the infrastructure, written the laws, allocated the budget. But there are no players on the field.

The measure of AI-era competitiveness is not the size of the infrastructure budget but the number of AI products from your country used globally. For Korea to become a genuine AI power, it has to switch paradigms — from “infrastructure projects” to “ecosystem building.”

And for AI developers: don’t wait for policy. Throw yourselves directly into the global ecosystem. Thanks to AI tools, an individual can ship a global product. If Korea’s AI ecosystem isn’t enough yet, make the world your ecosystem.

References