The AI Stack, Mapped
The global AI industry runs on 5 layers — from physical hardware to end-user applications. Like the OSI model, each layer builds on the one below. This map shows who leads at each layer, how competitive Korea is in each, and what it would take to close the gaps — the Gap Report now lives here.
2
layers
2
layers
1
layer
Avg. Competitiveness
56%
across 5 layers
Est. Funding Gap
$2B–$7.5B
sum of layer estimates
Vertical AI Applications
The application layer. Industry-specific AI products — healthcare, legal, manufacturing, media, education, fintech, customer service.
Global Leaders
Tempus One
Tempus · Public (NASDAQ)
Harvey
Harvey · Legal AI leader
AXON
AppLovin · Public (NASDAQ)
Synthesia STUDIO
Synthesia · $2.1B valuation
Duolingo Max
Duolingo · Public (NASDAQ)
Charlotte AI
CrowdStrike · Public (NASDAQ)
China
SenseNova
SenseTime (商汤科技) · Public (HK)
iFlytek Spark
iFlytek (科大讯飞) · Public (SHE)
uAI
United Imaging (联影医疗) · Public (SSE)
Congrong LLM
CloudWalk (云从科技) · Public (SSE)
Face++
Megvii (旷视) · Enterprise
Assessment
Korea's second strongest layer. Global leader in healthcare AI (Lunit). Strong in legal (Law&Company), companion AI (Scatter Lab), and manufacturing (MakinaRocks).
Opportunity
Scale successes internationally. Build agent-ready APIs so vertical products become "skills" for agent platforms.
Policy Recommendation
Scale winners internationally. Help vertical AI companies build agent-ready APIs so their products become 'skills' for agent platforms. Fund underserved verticals: Korean fintech AI, construction AI, agriculture AI.
Estimated Funding Gap
$300M–$1B
Agent Platforms & Tooling
The orchestration layer. Coding agents, AI search, workflow automation, no-code agent builders — where AI work happens.
Global Leaders
Claude Code
Anthropic · Dominant coding agent
Codex CLI
OpenAI · GPT-powered
Cursor
Cursor · $60B announced deal
LangGraph
LangChain · Industry standard
China
Manus
Manus · Chinese agent platform
AutoCoder
AutoCoder · Chinese Cursor
Assessment
Emerging Korean players: Liner (self-reported est. 1M users; ~300K Chrome installs; AI research/search), Allganize (Alli, enterprise agents), Sendbird (Delight.ai, CX agents). No Korean coding agent (Cursor/Claude Code equivalent).
Opportunity
No Korean coding agent yet — the biggest remaining gap. MCP server ecosystems, developer tooling, and enterprise workflow agents are all opportunities.
Policy Recommendation
Support Liner and Allganize's global expansion. The real question: should Korea build a Korean Cursor, or help Korean developers be the best users of global tools? Fund agent-adjacent tooling (Korean MCP servers, Korean enterprise integrations).
Estimated Funding Gap
$500M–$2B
Inference & Infrastructure
The serving layer. Inference platforms, GPU cloud, model optimization, MLOps — making models run at scale.
Global Leaders
Together Inference
Together AI · $3.3B valuation
Pinecone DB
Pinecone · Industry standard
OpenRouter
OpenRouter · Unified API
Modal
Modal · GPU cloud
China
SiliconCloud
SiliconFlow · Chinese Together AI
BigModel (GLM)
Zhipu AI (智谱AI) · Enterprise platform
Assessment
Emerging Korean players in inference optimization (FriendliAI) and data tooling (Superb AI, Nota AI, Lablup). But no Korean vector DB or model router yet.
Opportunity
Vector database and model routing are still open. Korean inference optimization is strong but needs scale. Sovereign AI cloud opportunity via Korean telcos.
Policy Recommendation
Accelerate FriendliAI and Nota AI scale-up. Fund Korean vector DB and model routing startups. Sovereign AI initiative should include software infrastructure.
Estimated Funding Gap
$500M–$2B
Foundation Models
The intelligence layer. LLMs, multimodal models, speech/voice AI — the "brains" that power everything above.
Global Leaders
GPT-4o / o1
OpenAI · $852B reported valuation
Claude 4.6
Anthropic · $965B reported valuation
Gemini 2.5
Google · Integrated across Google
Llama 4
Meta · Open source leader
China
DeepSeek V4
DeepSeek (深度求索) · $52B reported valuation
Qwen
Alibaba (阿里巴巴) · Top China model family
Doubao (豆包)
ByteDance (字节跳动) · 60M+ users
What free, open distribution would reveal
For a competitive model
For a model that performs well, broad free distribution is a meaningful advantage. Distribution, rather than training, is often the binding constraint in a market where developers tend to adopt the strongest available tool, and serving on a neutral platform addresses it directly. A free tier can act as the top of the funnel, while reliability, privacy, fine-tuning, and enterprise support remain paid offerings — an approach used by Mistral, Llama, and DeepSeek.
For a less established model
For a model with limited differentiation, the same transparency provides an early, objective signal: low adoption at no cost indicates where further development or investment may be needed.
Participation is itself informative — a willingness to compete openly reflects confidence in a model's performance.
Assessment
Korea produces competent LLMs (HyperCLOVA X, EXAONE, Solar), but global adoption remains limited, making this the least internationally established layer. The gap relates more to distribution and funding structure than to capability: DeepSeek and Mistral were built largely with private capital and concentrated talent, while much of Korea's model work is supported through sovereign-AI programs and large corporate research divisions whose goals extend beyond near-term global adoption.
Opportunity
There are better-aligned uses of public capital than a 'Korean GPT' built for parity. One worth considering: let private capital develop the models while the public sector serves them at no cost on public GPU and through OpenRouter, so that adoption — not grant milestones — becomes the test. A free model that attracts no users indicates a real performance gap. Support is more effective behind defensible positions (Korean-language enterprise, on-device and edge, open weights, efficiency) and orchestration over lower-cost models.
Policy Recommendation
This report aims to surface options rather than recommend specific policy. One approach that has been discussed is to have models developed primarily by private labs, then served at low or no cost on public GPU infrastructure and distributed through aggregators such as OpenRouter, so that real-world adoption provides an objective measure of performance. Adoption data of this kind can help direct support toward approaches that demonstrate traction.
Estimated Funding Gap
$500M–$2B
Hardware & Chips
The physical layer. AI accelerator chips, HBM memory, inference hardware — everything runs on this.
Global Leaders
H100 / B200
NVIDIA · $5.03T mcap
MI300X
AMD · NVIDIA challenger
Gaudi 3
Intel · Emerging AI chip
China
Ascend 910B
Huawei (华为) · China's NVIDIA alternative
MLU accelerators
Cambricon (寒武纪) · Chinese AI silicon
Assessment
Korea's strongest position. Samsung and SK Hynix dominate HBM memory. Rebellions and FuriosaAI are building Korean AI inference chips.
Opportunity
Korea builds the memory and increasingly the inference chips. The gap is in the software stack running on those chips.
Policy Recommendation
Continue supporting chip startups. Fund the software stack that runs on Korean hardware — chips alone are necessary but not sufficient.
Estimated Funding Gap
$200M–$500M
What China did vs. what Korea could do
Two very different paths to sovereign AI capability
China's approach: Forced localization
- - Great Firewall blocks ChatGPT, Claude, Google services
- - US chip export controls forced Huawei to build Ascend chips
- - Domestic demand creates viable market for local alternatives
- - Result: DeepSeek, Qwen, SiliconFlow — competitive stack top to bottom
Korea's opportunity: Open ecosystem advantage
- - No firewall — Korean companies can use AND build AI
- - World-class developer talent
- - Hardware advantage as foundation
- - Strategy: Build infrastructure that works globally, not just domestically
- - Align support with adoption: a mix of private capital and talent, focused where it adds the most value rather than on frontier parity alone
The picture
Korea is strong at the top and bottom of the stack — hardware (Layer 1) and vertical applications (Layer 5). Korean companies lead globally in HBM memory and have world-class AI products in healthcare, legal, and companion AI.
The middle layers have players but no global winners: infrastructure (FriendliAI, Nota AI) and agent platforms (Liner, Allganize) qualify as "some players" rather than empty. Only one layer is genuinely empty at global scale — foundation models with meaningful global adoption. Korea produces models, but almost no one outside Korea uses them.
The difference is more about distribution and funding structure than capability. DeepSeek and Mistral were built largely with private capital and concentrated talent around a clear point of differentiation, while much of Korea's foundation-model work is supported through sovereign-AI programs and large corporate research divisions. A productive question is less how to outspend frontier labs than whether that is necessary — and where public support aligns best with adoption: orchestration over lower-cost models, and the layers where local data, regulation, and hardware provide durable advantages.
This analysis is maintained by Korea AI. For corrections, additions, or to discuss findings: [email protected]