// cosmogenic.org — semiconductor intelligence — 2026.02
Geopolitical risk assessment · Unclassified

Global Fab
Sovereignty vs Competition

// semiconductor foundry roadmap · 2024–2030 · US CHIPS Act · China Made in China 2025

The current TSMC concentration risk is a transitional condition, not a permanent state. New foundries across the US, China, Japan, Europe, and Korea will reshape the semiconductor landscape by 2030 — but the transition creates its own risks. This document maps who is building what, when they'll produce, and what it means for edge compute.

All Nations 24
USA 7
China 6
Taiwan 4
Korea 4
Japan 2
Europe 2
// advanced node capacity share — select year to project forward
Concentration Risk
by Year
// sub-5nm advanced node wafer capacity
// 2024 — current state
92%
TSMC's share of sub-5nm advanced node production. A single company in a single geopolitical flashpoint controls the fabrication of every H100, MI300X, and Apple M-chip on earth.
// key transition
2027–2028
The inflection point. TSMC Arizona 2nm, Intel Ohio 18A, Samsung Texas 2nm, and Rapidus Japan 2nm all target this window. The first meaningful diversification of advanced node production since TSMC's dominance was established.
// china trajectory
7nm Wall
SMIC reached 7nm via DUV multi-patterning (slower, expensive, lower yield). Without EUV access, China is structurally limited to ~7nm until domestic EUV is viable — currently targeted 2027–2030 by CXMT/SMEE.
// fab construction & production timeline — click any event to inspect
2024
2026
NOW
2027
2028
2029
2030
2031+
// national semiconductor strategy — geopolitical risk assessment
National Strategies
🇺🇸
United States
CHIPS and Science Act · $52.7B
The CHIPS Act is the largest industrial policy intervention in US semiconductor history. Goal: domestic advanced node production by 2030 to reduce Taiwan Strait dependency. Intel, TSMC Arizona, Samsung Texas, and Micron all receiving grants.
CHIPS Act funding$52.7B
Intel Ohio grant$8.5B
TSMC Arizona grant$6.6B
First advanced node2027
🇨🇳
China
Made in China 2025 · $150B+
China's semiconductor self-sufficiency drive accelerated dramatically after US export controls on advanced chips and equipment. SMIC is pushing 7nm via DUV multi-patterning. CXMT targets DRAM independence. The EUV wall remains the critical barrier — without it, China is structurally capped at ~7nm.
State investment$150B+
SMIC best node7nm (DUV)
EUV accessBlocked (NL)
Domestic EUV target2028–2030
🇹🇼
Taiwan
TSMC global expansion · Geopolitical hedge
TSMC is simultaneously the world's most strategically important and most geopolitically exposed company. Its response: geographic diversification — Arizona, Japan (Kumamoto), Germany (Dresden), and future sites in Singapore. Taiwan remains the leading edge node home (2nm, 1.6nm A16) while exports serve geopolitical stability.
TSMC global sites6+ by 2030
Arizona Fab 213nm → 2nm
Japan Fab 2312nm (2024)
Leading edge homeTaiwan (2nm)
🇰🇷
South Korea
Samsung + SK Hynix · HBM dominance
Korea plays both sides: Samsung Foundry competes with TSMC for logic chips while SK Hynix dominates HBM memory critical for AI accelerators. Samsung's 3nm GAA yield issues have been a drag, but the 2nm roadmap (2025–2026) is competitive. HBM4 race between Hynix and Samsung will determine AI memory supply through 2028.
Samsung 2nm target2025–2026
SK Hynix HBM share~50%
HBM4 race2025–2026
Texas fab (Samsung)2026
// what this means for edge compute — cosmogenic strategic implications
Edge Compute Implications

The fab roadmap changes the GPU supply picture materially by 2028–2030. The strategic question is how to build edge infrastructure during the transition window — using what's available now — so that when new fabs increase supply, the network is already operational.

// 2024–2026 · NOW
Route Around
with What Exists
Consumer GPUs (RTX 5060 Ti, AMD RX 7900 XTX) are available today. CPU quantised inference via llama.cpp works on the R740 now. Open-weight models (Llama 3.1, Mistral, Phi-4) match closed-source on inference tasks. The supply constraint is a training problem — edge inference bypasses it entirely. Build the network now using what's available.
// 2027–2028 · TRANSITION
New Fab Supply
Changes the Market
TSMC Arizona 2nm, Intel 18A, Samsung Texas 2nm coming online means advanced node supply increases significantly. GPU prices should moderate as hyperscaler allocation pressure eases. This is when upgrading edge nodes from consumer to prosumer GPUs (L40S class, AMD MI300-equivalent) becomes economically viable at scale. Plan for hardware refresh cycles timed to this window.
// 2029–2031 · MATURITY
Distributed Supply
Matches Distributed Compute
By 2030, advanced node production will be meaningfully distributed across US, Japan, Korea, Europe, and Taiwan. The geopolitical single-point-of-failure risk reduces. Chinese domestic production (SMIC 5nm+ if EUV issues resolve) creates a parallel supply chain. The structural case for distributed edge compute becomes even stronger when the supply chain itself is distributed — redundancy compounds.