Concentration climbs the AI stack: cloud and models are only modestly concentrated, but upstream inputs — from advanced chip packaging and leading-edge lithography to refined critical minerals such as gallium — are extreme chokepoints dominated by one or a few suppliers, placing the limits of competition policy beyond domestic antitrust and into industrial policy and export control.

Main Finding

AI’s input stack exhibits a clear concentration gradient: markets are relatively unconcentrated at the downstream model layer but become sharply concentrated upstream in hardware, fabrication tools, memory, packaging, electricity supply chains, and mineral refining. Many of the most concentrated nodes are chokepoints (often single-firm or single-country dominances) that antitrust cannot reach; they are best addressed by export controls and industrial policy. Much of the upstream concentration is built (processing capacity and refining), not geological, so policy can move it — and the main strategic threat is loss of supply rather than a mere upstream price rise.

Key Points

• Measured on a common scale (Herfindahl–Hirschman Index, HHI) the stack shows rising concentration upstream:
  • Downstream: frontier-model usage ≈ 1,150; cloud ≈ 1,410 (both below the U.S. “highly concentrated” threshold of 1,800).
  • Mid/Upstream examples: high-bandwidth memory (3 firms) HHI ≈ 4,174; advanced packaging ≈ 8,100; leading-edge lithography ≈ 10,000 (index ceiling).
  • Minerals/refining: gallium production ≈ 9,800; graphite processing ≈ 8,649; rare-earth processing ≈ 8,100; cobalt refining ≈ 6,241.
• Three headline findings:
• A measurable concentration gradient: low downstream, very high upstream.
• Upstream layers are strategic chokepoints and vulnerabilities; they are often state- or foreign-controlled and thus outside domestic antitrust reach.
• Concentration is frequently created at processing/refining stages (not only geology). Hence capacity and policy choices matter.
• Economic mechanism emphasized: an input with few or no substitutes that is controlled centrally gives leverage over all downstream production (foreclosure risk; propagation of shocks).
• Policy-relevant insight about pricing: because upstream inputs often represent a small share of final costs, a chokepoint’s harm is primarily through loss of supply (availability) rather than by modest upstream price increases.

Data & Methods

• Conceptual framing: production-function view of AI capability (stack of distinct input markets) and supply-chain theory (double marginalization, foreclosure, shock propagation).
• Concentration metric: Herfindahl–Hirschman Index (HHI) used to place all layers on a single comparable scale; HHI’s reciprocal gives an “equivalent number of equal firms.”
• Data provenance: all series, code, and sources are cited and reproducible; companion public dashboard and repository at https://aistackmap.org. Each measured series is assigned a confidence tier in the paper.
• Sources drawn from technical and policy databases and reports (examples: U.S. Geological Survey, International Energy Agency, industry and market-technology reports, merger-guideline thresholds). The paper converts sector-specific measures (production shares, processing shares, firm market shares, national production/refining shares) into HHI values.
• Comparative approach: production, reserves, and refining/processing are compared to test whether concentration is geological (reserves) or industrial (refining/processing capacity).
• Limitations noted in the paper (brief): market boundaries matter for interpreting concentration; data gaps and varying quality across layers; dynamic changes (policy, investment, technological substitution) can alter concentration over time.

Implications for AI Economics

• Rethink policy focus: competition debates focused solely on models and cloud miss the most binding chokepoints. Antitrust tools are insufficient where control rests with foreign states or single global suppliers.
• Tools required: export controls, economic statecraft, and active industrial policy (domestic capacity expansion, strategic stockpiles, subsidies for refining/fabrication, deployment of alternative materials, and coordinated international action).
• Prioritize supply security over price controls: because the key risk is supply denial, policy should secure physical flows (diversify suppliers, onshore/refine capacity, build buffers), and not rely only on market-price mechanisms.
• Investment targets: advanced packaging, lithography alternatives/duplication, high-bandwidth memory sources, domestic/refined mineral processing, electricity-grid upgrades to support compute demand — these are leverage points to reduce systemic vulnerability.
• Governance and measurement: embed supply-chain concentration metrics into AI policy and regulation; expand oversight beyond visible downstream firms to critical upstream inputs; maintain transparent, reproducible data monitoring (e.g., dashboards like aistackmap.org).
• Research agenda: quantify dynamic responses (how long and at what cost capacity can be built), model geopolitical incentives to weaponize chokepoints, and study substitution technologies that reduce upstream specificity.

Reference / reproducibility note: all series and code used to generate the paper’s concentration measures are publicly available at the paper’s companion dashboard: https://aistackmap.org.