Quantifying Systemic Vulnerability
in the Foundation Model Industry

• Taiwan: 92% of advanced semiconductor manufacturing capacity
• NVIDIA: 86% of AI accelerator market
• Top 5 corporate labs: 72% of top-tier AI conference publications

Central Question

How vulnerable is the foundation model industry to supply chain disruptions across its critical inputs?

No systematic framework exists for assessing aggregate vulnerability when production requires complementary inputs.

1. Theoretical

Apply O-Ring production theory (Kremer 1993) to industrial vulnerability measurement, recognizing complementarity rather than substitutability among critical inputs.

2. Empirical

Construct Artificial Intelligence Industrial Vulnerability Index (AIIVI) using validated human-in-the-loop methodology for data-scarce, rapidly-evolving industries.

3. Policy

Identify energy infrastructure as emerging binding constraint, challenging current policy focus solely on semiconductor capacity.

Foundation model production exhibits extreme complementarity among inputs:

Key insight from Kremer (1993): When inputs are strict complements, the weakest component determines production capacity.

• No substitutability: Abundant energy cannot compensate for semiconductor shortages
• If any single input fails, aggregate output collapses
• Multiplicative vulnerability structure appropriate

Criticality weights based on impact analysis: effect of 50% reduction in each input on production capacity. Weights sum to 1.0.

Each vulnerability sub-index aggregates multiple indicators using weighted arithmetic means:

• SC_comp: √HHI of semiconductor fabrication market
• GEO_comp: Share in most concentrated region
• PR_comp: √HHI of GPU/accelerator design market

Weight Rationale

Weights reflect substitutability: high weight where alternatives are unavailable. Example: GEO_comp receives highest weight (0.50) as no short-term geographic substitutes exist for Taiwan's capacity.

Challenge

• All state-of-the-art models released September-November 2025 (3-month window)
• Private companies: minimal disclosure, no regulatory filings
• Information dispersed across grey literature (66 sources, 15 parameters)
• Traditional data collection infeasible at required scale and speed

Solution: Human-in-the-Loop AI-Assisted Protocol

Efficiency gain: 5× faster than manual (40 vs 200+ hours) while maintaining verification rigor.

Six State-of-the-Art Foundation Model Developers

Data Sources

• Academic: Nature, Stanford HAI, arXiv (Epoch AI)
• Agencies: IEA, BCG, SIA
• Industry analysts: TrendForce, Jon Peddie Research, KPMG, Deloitte
• News outlets, company announcements

Interpretation scale: 0.0-0.2 Low | 0.2-0.4 Moderate | 0.4-0.6 High | 0.6-0.8 Very High | 0.8-1.0 Extreme

Key observation: Two inputs (Compute, Energy) drive overall extreme vulnerability. Under O-Ring production, high criticality weights amplify their impact on aggregate AIIVI.

Normalization rule:

When ratio > 1 and AI growth positive, value capped at 1.0 to indicate maximum vulnerability. Grid capacity cannot scale at AI industry rates.

Key Insights

• Monte Carlo (10,000 simulations, ±20%): 95% CI entirely within Very High to Extreme range
• Equal-weighted scenario still indicates High vulnerability (0.55)
• Excluding low-confidence parameters yields AIIVI = 1.0, as energy constraint alone creates maximum vulnerability under O-Ring structure

High-confidence specification demonstrates energy bottleneck is sufficient condition for extreme vulnerability.

Current Policy Approaches

Both strategies incomplete: Treating inputs as substitutes when production function requires complementarity.

Optimal Intervention Under O-Ring Production

With α_C=0.40 and α_E=0.30, marginal returns to reducing energy vulnerability currently exceed returns to reducing compute vulnerability for western economies.

Market shows awareness of compute constraint (Anthropic Oct 2025: diversifying to TPUs, Trainium), but energy constraint lacks market adjustment mechanism.

Addressing Data Scarcity in Fast-Evolving Industries

Generalizability

Methodology applicable to other industries sharing: (1) opacity (minimal standardized disclosure), and (2) rapid evolution (capabilities change faster than traditional data cycles). Examples: cryptocurrency, quantum computing, synthetic biology.

Current Limitations

• Elite talent concentration relies on indirect indicators; direct researcher counts by affiliation needed
• Financial parameters limited by private company opacity; disclosure will improve as industry matures
• Weights preliminary; formal expert panel validation recommended
• Snapshot from November 2025; periodic updating required for tracking

Extensions

• Geographic disaggregation: Construct region-specific AIIVI (US/EU vs China)
• Downstream effects: Extend to full value chain including application vulnerability
• Temporal dynamics: Panel structure once time-series accumulates
• Causal analysis: What determines vulnerability patterns? (econometric approach when data available)

Main Findings

Contributions

• Theoretical: O-Ring production applied to vulnerability measurement formalizes why balanced intervention matters under complementarity
• Empirical: First systematic assessment identifying energy infrastructure, not just semiconductors, as emerging bottleneck
• Methodological: Validated human-in-the-loop approach enables composite index construction in data-scarce environments