Chokepoint Anatomy — Hourglass Concentration in the AI Compute Supply Chain

Gringras, David

Scrutica

Graph written 2026-09-03

Supply Chain

Supplier–customer relationships between named firms, and the layers with only one supplier.

5 of 9 layers have no measured concentration

4

of 9 layers have a Herfindahl: 2 measured outright, 2 as a lower bound whose unattributed remainder can only push the figure higher.

5

have none. 4 have a supply market that no tier-1 or tier-2 source breaks into shares; 1 is a consumer layer, where supply concentration is not a question that applies.

10,000

the highest measured figure on the ladder, Equipment, EUV lithography sub-layer. Ten thousand is the arithmetic ceiling: one supplier, the whole market.

The measurement gap. One row per layer of the compute supply chain, on a Herfindahl–Hirschman axis running from zero to its arithmetic ceiling of 10,000: one supplier holding the entire market. A solid bar is a figure recomputed in code from published market shares; a bar with an open edge is a lower bound, drawn short of a true value that lies somewhere to its right because part of the market is unattributed; a dotted span is a real supply market for which no tier-1 or tier-2 source publishes shares, so no honest Herfindahl exists; a bracketed span is a consumer layer that buys compute and supplies none, where supply concentration is not a question that applies. The assessed criticality beside each layer name is an editorial judgement of supply-chain importance and deliberately never touches the axis: drawn on the same scale it would read as a Herfindahl band. Each figure is a single-layer market-share Herfindahl for one market. The stress index reports a different construct, a portfolio-weighted supply-chain Herfindahl across 93 CSET steps, and the two are not comparable.

The hourglass below shows who occupies each step of AI chip production, from the machines that pattern the silicon to the packaging that assembles the finished part. The ladder above answers the prior question. Of 9 layers, 2 have a concentration figure computed from published shares and 2 have a lower bound whose unattributed remainder can only push it higher. The other 5 have nothing: 4 are real supply markets that no tier-1 or tier-2 source breaks into shares, and 1 is a consumer layer where the question does not apply. Scrutica publishes the gap, because a plausible number in those rows would be indistinguishable from a measured one.

The hourglass

43 companies across 9 layers, sized by supply-chain importance. Hover any node to trace what depends on it downstream. Relationships draw from the full dependency graph. Block widths order and size the layers for legibility; on the two layers whose complete charted set has published shares they are those shares, and everywhere else they are editorial weights that are never squared into any figure.

ASML is the sole manufacturer of EUV lithography machines; no alternative supplier exists at any price. No other node in the semiconductor supply chain comes this close to a single point of failure.

The right-hand column shows two separate judgments per layer: an editorial chokepoint-criticality level (assessed supply-chain importance, tier 4; it sorts and colors the rows and is computed from no data) and the measured concentration state. Measured figures, each recomputed in code from published shares: HBM HHI 4,174 (Counterpoint Q3 2025 revenue shares 57/22/21, tier 2); Fabrication HHI 5,007 (TrendForce FY2025 named top-6, a named-share lower bound, tier 2); EUV sub-layer HHI 10,000 (ASML sole supplier, tier 1); Cloud HHI ≥ 1,421 (Synergy Q4 2025 top-3 lower bound, tier 2; ~37% of the market is unattributed, so the true value is higher). Layers with no tier-1/2 share publication render “sourced-unavailable” (Advanced Packaging; Materials; EDA Tools and Chip Design each have tier-3 share figures only), and AI Labs have no HHI, because they buy compute and supply none. Measured HHIs above 2,500 exceed the “highly concentrated” threshold of the 2010 DOJ/FTC Horizontal Merger Guidelines (the 2023 Guidelines lower that line to 1,800); the comparison applies only to measured figures, never to editorial weights.Each layer shows two separate judgments. The colored level is our editorial assessment of how critical that supply-chain step is; no data computes it. The HHI number, where one appears, is a measured market-concentration figure built from published market shares, and each one names its source, year, and authority tier inline. Where no reliable share data is published, the display says “sourced-unavailable” in place of an estimate; AI labs get no number because they buy compute and sell none. A measured HHI above 2,500 counts as “highly concentrated” under the 2010 US merger guidelines (the 2023 guidelines use 1,800). Three measured layers clear that bar: HBM memory, chip fabrication, and EUV lithography equipment, where one company, ASML, is the only supplier on Earth.

Frontier-facility callout. Anthropic’s compute access also includes the entire Colossus 1 facility (SpaceX-owned post-merger; Memphis, >300 MW, >220K NVIDIA GPUs) under a 100% lease announced 2026-05-06. Colossus 1 is a single-tenant facility, so it is excluded from the Cloud & Infrastructure layer’s concentration count, which measures cloud providers. Sources: xAI announcement (Tier 1), DataCenterDynamics 2026-05-06 (Tier 2).Anthropic also has exclusive access to the entire Colossus 1 facility in Memphis (over 220,000 NVIDIA GPUs, >300 MW), owned by SpaceX since its 2026 merger with xAI and leased 100% to Anthropic from 2026-05-06. We track this separately from the cloud-provider concentration layer, which counts shared cloud services; this is a single-tenant lease. Sources: xAI and DataCenterDynamics.

Methodology

Two ingredients, kept strictly separate. (1) Measured concentration. Where a tier-1/2 source publishes market shares, the HHI is recomputed in code from those shares: HBM 4,174 (Counterpoint Q3 2025: 57² + 22² + 21²); Fabrication 5,007 as a named-share lower bound (TrendForce FY2025: 69.9² + 7.2² + 5.32² + 4.35² + 3.87² + 2.6²); EUV sub-layer 10,000 (ASML 2025 Annual Report + Q4-2025 6-K, tier 1; ASML is the sole EUV supplier: 48 systems recognized 2025, sales +39% to €11.6B); Cloud ≥ 1,421 as a top-3 lower bound (Synergy Q4 2025: 28² + 21² + 14²; ~37% of the market is unattributed, so the true value is strictly higher). Where no such source exists the display says sourced-unavailable: Advanced Packaging (the TrendForce capacity trajectory is not a share; the circulating “40% revenue share” is paywalled and unverifiable), Materials (the Omdia per-supplier wafer breakdown has no public locus), and the broader Equipment market beyond EUV.

Chip Design publishes a share band only: NVIDIA ~80–85% of merchant AI accelerators (tier-3 analyst band). Squaring a mid-band share (~82%) would alone contribute ≈ 6,700 HHI points, far above any concentration threshold, but that arithmetic inherits the band’s tier-3 status and its denominator sensitivity (counting captive ASICs like TPU and Trainium moves NVIDIA toward ~75%; excluding ASICs entirely moves it above 90%), so no HHI is published for this layer. The OECD’s “over 80%” GPU line (Competition in AI Infrastructure, Nov 2025, p.14 §2.3.1) is itself a tier-3 analyst estimate (Farooque, Yahoo Finance, ~85%) reported by the OECD, which makes it a citation chain and not an OECD measurement. AI Labs have no HHI: labs buy compute and supply no market (context: Menlo Nov 2025 enterprise API-spend survey, n=495, tier 3; Epoch AI compute-share estimates, tier 2). Advanced node: Taiwan ~92% of leading-edge pure-play foundry manufacturing (CSET “No Permits, No Fabs”, VerWey 2021) is a country-level share including UMC and PSMC, so it is not a TSMC company share; no tier-1/2 company-level advanced-node share exists.

(2) Editorial importance weights. Entity block sizes use the measured shares where the layer’s full charted set is published (HBM, Fabrication) and assessed supply-chain-importance weights elsewhere (tier 4, informed by CSET TechInsights 2024, a licensed supply-chain database, and SEC EDGAR). Weights order the visual; they are never squared into any displayed or exported number. For empirically derived supply-chain relationships with relationship values and price correlations, see the Supply Chain Explorer.

Where reliable published market-share data exists, the concentration numbers here are computed directly from it, and every number names its source, year, and authority tier: memory chips (Counterpoint Research, Q3 2025), chip fabrication (TrendForce, full-year 2025), cloud computing (Synergy Research, Q4 2025, shown as a minimum since only the top three providers are attributed), and EUV lithography machines (ASML’s own annual report; ASML is the sole supplier on Earth, so that market scores the maximum 10,000). Where no reliable share data is published we say “sourced-unavailable” instead of estimating: advanced packaging, wafer materials, and the broader equipment market. For AI chip design we report a range (NVIDIA roughly 80–85% of openly-sold AI accelerators) because the answer depends on whether you count chips, like Google’s TPUs, that are never sold to anyone. AI labs get no concentration score: they buy compute and sell none. Company block sizes reflect published shares where they exist and our labeled editorial judgment of importance where they don’t; those judgments never feed any number. Relationships are drawn from licensed bilateral supply-chain data and CSET market classification. Sole-source exposures are flagged wherever no known alternative supplier exists. For the underlying empirical data, see the Supply Chain Explorer.

Show all 9 supply chain layers as a table (alternative to the interactive hourglass)

All 9 layers of the compute supply chain, from raw materials to AI labs. Each row has the editorial chokepoint-criticality level (assessed supply-chain importance, tier 4; it is not a Herfindahl band), the measured concentration state (recomputed in code from published market shares: Counterpoint Q3 2025 for HBM; TrendForce FY2025 named top-6 for foundry, a lower bound; Synergy Q4 2025 top-3 for cloud, a lower bound; ASML 2025 Annual Report for EUV, tier 1; sourced-unavailable is stated where no tier-1/2 share publication exists, and AI labs have no figure because they buy compute and supply none), the citable basis with metric, source, vintage and authority tier, and the top-3 nodes in editorial importance order.
Layer	Criticality (editorial) — Editorial chokepoint-criticality assessment (moderate / high / extreme): assessed supply-chain importance, authority tier 4. It is not a Herfindahl band; the measured concentration state is in the next column.	Measured concentration — Herfindahl–Hirschman Index recomputed in code from published market shares (range 0–10,000). "≥" marks lower bounds; "sourced-unavailable" means no tier-1/2 share publication exists; AI labs have no figure because they buy compute and supply none. Never computed from editorial weights.	Basis (metric · source · vintage · tier)	Top nodes (editorial importance order) — Ordered by block-sizing weight: the measured share where the layer’s full charted set is published (HBM, Fabrication), otherwise editorial supply-chain-importance weight. Ordering only; weights are never rendered as market-share percentages.
AI Labs · Frontier compute consumers	moderate	NO HHI — consumer layer	No HHI: AI labs are frontier compute consumers, not suppliers in a defined product market. Context stats (hard scope labels — NOT market shares in the HHI sense): US-enterprise LLM API spend Anthropic 40 / OpenAI 27 / Google 21 (Menlo Ventures survey n=495, Nov 2025, tier 3); OpenAI ~10–15% of world operational AI compute and top-five labs combined <50% at end-2025 (Epoch AI estimate, tier 2).	OpenAI · US; Google DeepMind · GB; Anthropic · US
Cloud & Infrastructure · Compute deployment	moderate	HHI ≥ 1,421 — top-3 lower bound	LOWER BOUND (top-3 only): 28² + 21² + 14² = 1,421 from Synergy Research Group Q4 2025 worldwide cloud infrastructure revenue shares (tier 2). The remaining ~37% (Oracle, CoreWeave, Alibaba, long tail) is unattributed — the true HHI is strictly higher.	Amazon Web Services · US; Microsoft Azure · US; Google Cloud · US
Chip Design · GPU & accelerator designers	high	HHI — sourced-unavailable (tier-1/2)	No HHI at tier-1/2. Share band: NVIDIA ~80–85% of merchant AI accelerators (tier-3 analyst band, 2025). Market-definition note: including captive ASICs (TPU, Trainium) moves NVIDIA toward ~75%; excluding ASICs entirely moves it above 90%. The band-squaring arithmetic appears in methodology text only, with caveats; the OECD "over 80%" line is a disclosed tier-3 citation chain (Farooque via OECD), not an OECD measurement.	NVIDIA Corporation · US; Advanced Micro Devices · US; Broadcom Inc · US
EDA Tools · Electronic design automation	high	HHI — sourced-unavailable (tier-1/2)	No HHI at tier-1/2. Shares on record (canonical store, all is_estimated, vintage 2024): Synopsys ~32% (FY2024 10-K revenue $6.13B against tier-3 total-market estimates), Cadence ~30% (FY2024 10-K, 71% EDA segment), Siemens EDA ~14% (industry estimate, no named source). Three named firms sum to ~76%; squaring estimated shares is deliberately withheld — a derived HHI would assert precision the tier-3 inputs cannot carry.	Synopsys · US; Cadence Design Systems · US; Siemens EDA · DE
Advanced Packaging · CoWoS, InFO, chiplet integration	high	HHI — sourced-unavailable	No tier-1/2 share publication for advanced packaging (CoWoS) — HHI sourced-unavailable. Capacity context (TrendForce, tier 2): ~75K wafers/month end-2025 → ~130K projected end-2026; TSMC dominant with ASE (SPIL) and Amkor as overflow. Capacity is not a market share.	TSMC (CoWoS) · TW; ASE Technology · TW; Amkor Technology · US
HBM Memory · AI accelerator memory	high	HHI 4,174 — measured	Measured: 57² + 22² + 21² = 4,174 from Counterpoint Research Q3 2025 HBM revenue shares (SK Hynix / Samsung / Micron — all suppliers named; tier 2, via Semiecosystem). Samsung recovered 15% → 22% Q2→Q3 2025 on HBM3E qualification; refresh on Q4 publication.	SK Hynix · KR; Samsung · KR; Micron Technology · US
Fabrication · Semiconductor foundries	extreme	HHI 5,007 — named-share lower bound	Named-share lower bound: 69.9² + 7.2² + 5.32² + 4.35² + 3.87² + 2.6² ≈ 5,007, recomputed in code from TrendForce FY2025 foundry revenue shares (tier 2, via Taipei Times 2026-03-14). Advanced node: Taiwan ~92% of leading-edge pure-play foundry manufacturing (CSET/VerWey 2021, COUNTRY-level incl. UMC + PSMC); no company-level advanced-node HHI exists at tier-1/2.	TSMC · TW; Samsung Foundry · KR; SMIC · CN
Equipment · Lithography, deposition, etch	extreme	HHI 10,000 — EUV sub-layer only	EUV sub-layer (tier 1): ASML 100% of EUV lithography systems → HHI 10,000 — the strongest figure in this dataset. Source: ASML 2025 Annual Report, financial-performance section + Q4-2025 investor presentation (Form 6-K, Jan 28, 2026) (48 EUV systems recognized 2025, 4 EXE + 44 NXE; EUV system sales +39% to €11.6B). Broader wafer-fab equipment market: sourced-unavailable — no HHI published. The prior blended equipment figure is retired.	ASML Holding · NL; Applied Materials · US; Lam Research · US
Materials & Chemicals · Wafers, photoresists, specialty gases	high	HHI — sourced-unavailable	Sourced-unavailable, split layer: silicon-wafer per-supplier shares have no public locus (Omdia subscription-gated; ledger 2026-06-09) — a wafer HHI does not publish; photoresist/specialty-chemical shares have no tier-1/2 publication. Country-level context: Japan ~56% of global silicon wafers (CSET TechInsights 2024, tier 2).	Shin-Etsu Chemical · JP; SUMCO Corporation · JP; JSR Corporation · JP

What a Herfindahl here does and does not say

A Herfindahl–Hirschman index is the sum of every supplier’s squared percentage share of one market. Squaring is what makes it useful: it weights a seventy-per-cent supplier far more heavily than seven ten-per-cent ones, so the index separates a market with one dominant firm from a market of several mid-sized ones. Scrutica recomputes each figure in code from the published shares, so the arithmetic behind every number on the ladder is auditable and the sources are named per layer.

Where only part of a market is attributed, squaring the named shares alone yields a floor: the unattributed remainder can only add to the sum. Those rows are labelled lower bounds and drawn with an open edge. Where nothing publishes, nothing is computed. The row stays empty on purpose, with no share band squared into a stand-in and no proxy substituted. The editorial criticality levels shown beside the layer names are assessments of supply-chain importance at authority tier 4; they order the hourglass and are never squared into any figure, on this page or in any export.