Every advanced AI chip requires components from over thirty countries working in coordination. The narrative framing semiconductors as a US-China conflict obscures a more complex reality: a globally interdependent system where no single power controls the complete value chain.
The eight critical dependencies
- EDA software. 90% of the market sits with three firms: Synopsys and Cadence (~30% each) and Siemens EDA (~13%). Every chip flows through one of these, and a redesign costs $50-100M on advanced nodes.
- Silicon wafers at 99.999999% purity depend on Japan's Shin-Etsu and Sumco, controlling roughly half the critical 300mm segment.
- Photoresist chemicals for EUV are a Japanese-Korean oligopoly; JSR holds ~22% and the top five make ~50%.
- Process gases. Before 2022, Ukraine produced ~50% of the world's neon and supplied up to 90% consumed by the US industry; the dependency shifted after Russia's invasion.
- ASML is the sole maker of EUV lithography machines (100% share). Each costs ~$180M (low-NA) or ~$380M (High-NA), with 100,000+ components from 5,000+ suppliers. Two are critical: Carl Zeiss SMT (sole supplier of EUV mirrors) and Cymer in California (sole maker of the laser source). Cymer's US-origin tech gives Washington veto power over sales to China.
- Etch, deposition and inspection concentrate in US firms. Lam Research and Applied Materials hold ~60% of dry etch; KLA leads inspection; Tokyo Electron controls 100% of coaters/developers accompanying every EUV scanner.
- Photomasks and pellicles. Lasertec alone controls 100% of the actinic mask inspection market at 5nm or below: a 1,460-employee firm with technical veto power over advanced production.
- Foundries concentrate at TSMC: 70.2% of global foundry share in Q2 2025; Samsung 7.3% with ~50% yields on 3nm vs TSMC's 90%.
The hidden half: packaging and beyond
The real bottleneck is advanced packaging. TSMC's CoWoS processes every H100, H200 and B200 GPU, saturated with 52-78 week lead times; Nvidia reserved over 50% of all 2026 CoWoS capacity. High Bandwidth Memory (HBM) doubled from $17B (2024) to $34B (2025), heading to ~$100B by 2028. In Q3 2025 SK Hynix held 57% (overtaking Samsung in DRAM for the first time since 1983).
Software is the third bottleneck: Nvidia holds 90%+ of data-center AI GPUs, but the moat is CUDA, hardened over nineteen years into libraries, framework integration and accumulated knowledge. AMD's ROCm closed the hardware gap in 2025, but the software maturity gap persists.
Geology is the fourth: China controls 85%+ of rare-earth processing, ~98% of dysprosium and ~80% of tungsten (no substitute in chip interconnects). Physical resources are the fifth: the IEA projects data centers growing from 485 TWh (2025) to 950 TWh (2030), with hyperscalers buying dedicated nuclear power.
Europe's unrecognized position
European technical monopolies are largely unacknowledged. ASML controls 100% of EUV lithography. Carl Zeiss SMT (revenue €1.2B in 2016 to €4.1B in 2024) supplies 100% of EUV mirrors: "Without Zeiss, there are no AI chips." SOITEC controls 70-80% of SOI wafers; Infineon is #1 in automotive semiconductors; ARM has shipped 300B+ cumulative chips. IMEC in Leuven is arguably the world's most important semiconductor R&D center precisely because of its neutrality: TSMC, Intel, Samsung and others collaborate there pre-competitively. In March 2026 IMEC secured exclusive access to ASML's High-NA technology.
In September 2025, ASML led a €1.7B round in Mistral, investing €1.3B for 11%: sovereign European compute plus IP integration, occurring quietly while public discourse stayed fixated on US-China dynamics.
The decoupling paradox
Both the US and China pursue technological independence that experts deem physically impossible. The US rescinded the AI Diffusion Rule in May 2025 because it classified NATO allies as Tier 2; China partially suspended rare-earth controls in November 2025 after a bilateral agreement. These are successive dependency management, not coherent industrial policy. IMEC persists because no one can afford to break it: doing so would fracture Moore's Law for all.
Strategic imperatives for Europe
- Abandon "self-sufficiency". Europe need not replicate TSMC or Nvidia, but convert existing monopolies into non-negotiable levers through "indispensability".
- Accelerate Chips Act 2.0, classifying semiconductors as defense-equivalent with Commission investment authority, executed at startup velocity.
- Improve self-narration. ASML's Mistral investment deserved front-page coverage; High-NA qualification at IMEC deserves systematic analysis. Continued self-description as observer perpetuates observer behavior.
Every global AI chip requires mirrors polished in Germany, resin mixed in Japan and packaging built in Taiwan, each depending on rare earths refined in China, software written in California, gases captured in Ukraine and increasingly nuclear electricity from reopened US plants. This is a fabric, not a war. And Europe occupies the place the chain already granted it.