Korea’s Trillion-Dollar Sovereign AI Investment: Nvidia Wins, Hynix Loses

Max Kan · SemiAnalysis · September 01, 2026 at 20:14 · ⏱ 18 min read  | Read on Substack ↗
Summary
The article argues that South Korea's massive sovereign AI infrastructure buildout is a strategic lifeline for Nvidia to diversify its customer base away from US hyperscalers. However, this dynamic means Nvidia is leveraging these sovereign compute deals to extract favorable high-bandwidth memory (HBM) pricing and volume commitments from Korean memory giants, ultimately hurting Samsung and SK Hynix shareholders.
  • •South Korea hosted a $350M 'Independent AI Foundation Model' tournament, initially providing ~3,000 H100 equivalents and $45M in data subsidies to domestic consortiums.
  • •Motif Technologies, a 30-person startup with just $17M in funding, trained a 2.6B parameter model (Motif 3) for ~$15M that outperformed US open-source models like Inkling and Nemotron 3 Ultra.
  • •Despite scoring the highest on benchmarks, the Korean government eliminated Motif from the tournament due to opaque expert reviews and concerns over foreign investment.
  • •South Korea announced a $919B investment to build 8.4GW of datacenter capacity by 2029 and 18.4GW by 2035, with SK Group, GS Group, and Naver leading Phase 1.
  • •Nvidia desperately needs sovereign AI and open-source ecosystems to succeed, as it is currently on track for a world where Anthropic and OpenAI dominate frontier token volumes, leaving Nvidia with only 2-7 real customers.
  • •SK Group has committed to building 2GW of Nvidia Rubin systems, while Naver has signed on for 200MW, and Samsung plans a 50,000 GPU AI factory.
  • •Nvidia is using Korea's sovereign compute buildout to deepen ties with SK Hynix, likely securing favorable HBM pricing and large volume commitments for 2027 at the expense of memory margins.
Read time 18 min
Length 18,712 chars
Category finance
Ideas
Max Kan Substack author, SemiAnalysis
The author argues sovereign AI is critical for Nvidia to avoid customer concentration risk with US hyperscalers. 'Sovereign AI is another huge opportunity for Nvidia to diversify their customer base.
The author argues sovereign AI is critical for Nvidia to avoid customer concentration risk with US hyperscalers. 'Sovereign AI is another huge opportunity for Nvidia to diversify their customer base. SK Group has already committed to building 2 GWs of Rubin... and Naver has signed on for another 200MW.' Risk: If sovereign AI initiatives fail to produce competitive models, token volume and compute demand will consolidate back to US hyperscalers who are building custom silicon (XPUs) to cannibalize Nvidia's margins.
NVDA
Max Kan Substack author, SemiAnalysis
The author explicitly warns that Korea's sovereign AI ambitions 'may not be aligned with Samsung and Hynix shareholders.' Nvidia is leveraging the infrastructure buildout to secure 'favorable HBM pric
The author explicitly warns that Korea's sovereign AI ambitions 'may not be aligned with Samsung and Hynix shareholders.' Nvidia is leveraging the infrastructure buildout to secure 'favorable HBM pricing and large volume commitments for 2027,' which will pressure SK Hynix's memory margins. Risk: If HBM supply remains structurally constrained globally, SK Hynix may retain pricing power despite Nvidia's negotiation leverage.
000660.KS
Max Kan Substack author, SemiAnalysis
Similar to SK Hynix, Samsung is caught in Nvidia's leverage play. 'Samsung reportedly plans to build an Nvidia-powered AI factory using more than 50,000 GPUs,' which Nvidia is using to deepen strategi
Similar to SK Hynix, Samsung is caught in Nvidia's leverage play. 'Samsung reportedly plans to build an Nvidia-powered AI factory using more than 50,000 GPUs,' which Nvidia is using to deepen strategic alignment and extract favorable terms across the AI infrastructure stack at the cost of memory margins. Risk: Samsung's broader foundry and consumer electronics businesses could offset any margin compression in their HBM/memory division.
005930.KS
Max Kan Substack author, SemiAnalysis
SK Telecom is a major beneficiary of the government's AI subsidies and infrastructure push. They survived the initial tournament rounds and their 'announced 2GW DSX AI factory is expected to deploy Ve
SK Telecom is a major beneficiary of the government's AI subsidies and infrastructure push. They survived the initial tournament rounds and their 'announced 2GW DSX AI factory is expected to deploy Vera Rubin systems powered by SK Hynix HBM4.' Risk: Massive capital expenditures for the 2GW datacenter buildout could strain the company's balance sheet if domestic AI demand does not materialize.
SKM
Max Kan Substack author, SemiAnalysis
Naver is positioned as a key infrastructure player in Korea's sovereign AI push. They are 'building a 1GW site in Sejong, South Korea' and have 'signed on for another 200MW' of Nvidia compute, giving
Naver is positioned as a key infrastructure player in Korea's sovereign AI push. They are 'building a 1GW site in Sejong, South Korea' and have 'signed on for another 200MW' of Nvidia compute, giving them massive domestic scale. Risk: Naver was disqualified from the national AI tournament for using Alibaba's Qwen vision/audio encoders, which cuts them off from direct government model-training subsidies.
035420.KS
More from SemiAnalysis
Most Neoclouds Suck At Security
Aug 30, 15:46
OpenAI Jalapeño: Better Than Nvidia Blackwell
Aug 25, 14:00
AgentX - InferenceXv3: Does CUDA Moat Hold up in Agentic Inferencing?
Aug 24, 00:19
Are Open Models Catching Up?
Aug 21, 16:40
Cerebras's Next Generation CS-4: Fast Just Got Faster
Aug 19, 01:32

This newsletter, published September 01, 2026, features Max Kan discussing NVDA, 000660.KS, 005930.KS, SKM, 035420.KS. 5 trade ideas extracted by AI with direction and confidence scoring.

Speakers: Max Kan  · Tickers: NVDA, 000660.KS, 005930.KS, SKM, 035420.KS