Semiconductor Supply Chain

In the AI Story

Semiconductor manufacturing separates into design (architecture and logic), fabrication (physical production), and packaging/testing (assembly into usable form). NVIDIA dominates AI chip design with 80-90% market share; AMD competes distantly; hyperscalers (Google, Amazon, Microsoft, Meta) develop custom chips but remain smaller players. Design concentration matters because architectural decisions—which operations to optimize, which precision to support, which memory bandwidth to provide—shape what AI models can efficiently compute. CUDA, NVIDIA's software ecosystem, creates lock-in: models optimized for NVIDIA chips cannot easily migrate to alternatives without performance penalties. The design layer's concentration means one company's roadmap decisions disproportionately influence the AI revolution's computational substrate.

Fabrication concentration exceeds design concentration. TSMC manufactures all NVIDIA chips, most Apple silicon, and significant fractions of AMD, Qualcomm, and other advanced logic. Samsung operates as secondary manufacturer but with lower yields—the percentage of functional chips per wafer—meaning higher costs and longer timelines for equivalent output. Intel's foundry ambitions, announced in 2021 and backed by CHIPS Act subsidies, represent a multi-year, multi-billion-dollar effort to re-enter leading-edge manufacturing after falling behind TSMC technologically. Even with government support, Intel's Arizona fabs face four-year construction timelines and workforce development challenges—experienced semiconductor engineers are scarce, and training takes years. Geographic concentration in Taiwan creates the most discussed geopolitical risk: the island produces roughly 90% of the world's most advanced chips while situated 100 miles from mainland China.

The EUV lithography bottleneck is the supply chain's deepest constraint. ASML's monopoly reflects twenty years of development and tens of billions in investment to produce light at 13.5-nanometer wavelength—generated by vaporizing tin droplets with lasers—and focus it with mirrors polished to atomic-scale precision. The machines weigh 180 tons, require multiple 747 flights to ship, cost $380 million each, and contain 100,000+ components from hundreds of suppliers. No competitor is within a decade of replicating the capability. ASML's single manufacturing facility in Veldhoven is a unique industrial asset whose disruption would halt frontier chip production globally until repairs were completed. The company produced roughly forty to fifty EUV systems annually as of 2024-2025—a production rate that limits how fast global chip fabrication capacity can expand regardless of capital availability.

Upstream supply chain dependencies extend to materials, chemicals, and gases whose sourcing concentrates geographically and whose substitution requires years. Rare earth processing: ~90% in China. Neon gas: historically Ukraine/Russia, partially diversified post-2022 but not replaced. Gallium and germanium: subject to Chinese export controls imposed 2023. Ultra-pure water: requires treatment to 18-megohm resistivity, roughly 1,000x purer than drinking water, demanding specialized facilities. Photoresists and etchants: chemically complex compounds from specialized suppliers. Each material represents a potential constraint where supply disruption propagates through the chain, delaying or halting production at facilities whose capital value exceeds $20 billion and whose construction timelines span years. The complexity is not accidental—leading-edge semiconductor manufacturing is among the most technically demanding industrial processes humans have developed, and the demands concentrate capability in a small number of firms and geographies.

Origin

Semiconductor supply chain analysis emerged as a strategic concern during the U.S.-Japan trade conflicts of the 1980s, when American firms lost DRAM market share to Japanese competitors. The framework gained urgency with the 2000s offshoring of fabrication to Taiwan and Korea, the 2010s rise of mobile computing (demanding advanced chips at consumer-device volumes), and the 2020-2022 chip shortage that halted automotive production worldwide. AI's emergence as the dominant computational workload in the mid-2020s elevated supply chain concentration from industrial concern to national security priority, triggering the CHIPS Act (U.S.), European Chips Act, and Japanese semiconductor investment programs—each committing tens of billions to domestic fab construction.

Smil's treatment synthesizes his work on supply chain fragility (How the World Really Works documents food, energy, materials, globalization), construction timelines (Growth analyzes infrastructure inertia), and innovation realism (Invention and Innovation warns against hype). His specific warning about AI chip supply chains appears in interviews and the Bankinter presentation: the concentration is a vulnerability that diversification efforts address too slowly relative to demand growth. The implication—that AI capability could be supply-constrained by hardware scarcity even as software capability continues advancing—inverts the usual assumption that computation is the abundant input and human attention the scarce one.

Key Ideas

TSMC fabrication monopoly. One company manufactures essentially all frontier AI chips; its facilities concentrate in Taiwan (geopolitical risk) and Arizona (under construction, delayed); no alternative achieves comparable yields or capacity at leading nodes.

ASML lithography bottleneck. Fewer than 200 extreme ultraviolet lithography machines exist globally, all manufactured by one Dutch company; every frontier chip passes through one of these machines, creating an irreplaceable chokepoint.

Four-year fab timeline. New semiconductor fabrication facilities require four years minimum from groundbreaking to production, $20-40 billion capital investment, and specialized workforces that cannot be hired or trained quickly—limiting how fast capacity can scale.

Rare earth concentration. Critical materials for semiconductor manufacturing—neodymium, dysprosium, lanthanum, gallium, germanium—concentrate in Chinese mining and processing; export controls can constrain supply independent of fabrication capacity.

Yield as hidden variable. The percentage of functional chips per wafer determines effective capacity; Samsung's lower yields mean it produces fewer working chips from equivalent silicon, energy, and time—making high yield a strategic advantage not easily replicated.