About this report

A supply chain research report on AI infrastructure, tracing the physical path from raw quartz to running GPU clusters. Covers semiconductor materials, lithography, chip design, foundries, memory, packaging, networking, photonics, power, cooling, construction, and deployment.

How bottlenecks are scored

Bottlenecks are scored using FMEA (Failure Mode and Effects Analysis), a risk methodology from aerospace and automotive engineering. Each failure mode gets three scores from 1-10:

• Severity: impact on AI infrastructure if it fails
• Occurrence: likelihood of failure
• Detection: how hard it is to see coming

Multiplied together, these give a Risk Priority Number (RPN) from 1 to 1,000:

• Critical (RPN > 140)
• Significant (RPN 90-140)
• Moderate (RPN 50-89)
• Low (RPN < 50)

How companies are classified

• Gate: no alternative at any price. Removal halts the supply chain. Examples: ASML (EUV lithography), Carl Zeiss SMT (EUV optics), Ajinomoto (ABF substrate film).
• Bottleneck: alternatives exist but take 12+ months to qualify. Examples: TSMC (advanced nodes), SK Hynix (HBM).
• Dominant: market leader with real competitors. Switching is feasible within 6-12 months. Examples: Applied Materials (etch), Broadcom (switch silicon).

Sources

SEC filings, earnings calls, government reports (DOE, IEA, CSIS), and industry analysis (SEMI, TrendForce, Dell'Oro, Yole). Academic literature lags this domain by 12-24 months, so the report favors current practitioner sources. Where estimates conflict, the report presents the range.

Scope

Covers companies whose revenue is materially tied to AI data center construction, globally, including private companies in critical positions. Does not cover AI software, enterprise applications, or financial instruments.