Graph written 2026-09-03
Supply Chain
Supplier–customer relationships between named firms, and the layers with only one supplier.
Supplier–customer relationships between named firms, and the layers with only one supplier.
The ordering composite, decomposed. One bar per tracked chokepoint. Each bar’s total length is the composite this page orders by, and the two spans are the two operations that built it: the solid span is the layer’s Herfindahl after the substitute deduction, and the outline extends it by the demand-growth multiplier. The caret marks the raw Herfindahl before either operation, so where it falls inside the outline, demand growth more than replaced what substitutes deducted; where it falls past the bar’s end, the substitute deduction won and the composite is smaller than the raw Herfindahl. On the single row with no substitute at any price, the solid span is the full Herfindahl and the outline is pure demand growth. The top bar reproduces from its own operands: EUV Lithography Systems at a raw Herfindahl of , 0% substitute availability and 15% annual demand growth, composing to 11,500. The composite is an editorial ordering device and its own provenance record says so: every row is flagged estimated, the estimation method is recorded as editorial, and no part of it is a measured severity. It is drawn this way so a reader can reproduce the published number from the published operands.
The same 14 chokepoints ranked four ways: by Herfindahl alone, by each single editorial term applied to it, and by the published composite. 4 of 14 keep the same position under all four constructions; those rankings are robust to the editorial layer. The largest mover is EML Laser Chips (800G+ Optical Interconnect), which shifts 3 places across the constructions; its position is the editorial judgment.
| Chokepoint | HHI alone | HHI × (1 − substitutes) | HHI × (1 + growth) | Published composite | Spread |
|---|---|---|---|---|---|
| EUV Lithography Systems | 1 | 1 | 1 | 1 | 0 |
| ABF Substrates (Ajinomoto Build-up Film) | 3 | 3 | 3 | 2 | 1 |
| Gallium & Germanium (Compound Semiconductors) | 2 | 2 | 2 | 3 | 1 |
| EML Laser Chips (800G+ Optical Interconnect) | 6 | 7 | 4 | 4 | 3 |
| High Bandwidth Memory (HBM) | 5 | 4 | 6 | 5 | 2 |
| Advanced 2.5D Packaging (CoWoS-class) | 4 | 5 | 5 | 6 | 2 |
| Chiplet Interconnects (UCIe) | 9 | 8 | 7 | 7 | 2 |
| Wafer Inspection & Metrology Equipment | 7 | 6 | 8 | 8 | 2 |
| Semiconductor-Grade Neon Gas | 8 | 9 | 9 | 9 | 1 |
| US Data Center Grid Interconnection | 10 | 11 | 10 | 10 | 1 |
| EDA Tools (Advanced Node) | 11 | 10 | 11 | 11 | 1 |
| EUV Photomasks & Pellicles | 12 | 12 | 12 | 12 | 0 |
| Photoresist (EUV-grade) | 13 | 13 | 13 | 13 | 0 |
| Silicon Wafers (300mm) | 14 | 14 | 14 | 14 | 0 |
ReadingRank 1 is the highest score under that construction; tied scores share a rank. A spread of 0 means the chokepoint’s position owes nothing to the hand-assigned substitute and growth terms; a large spread means the position is the editorial judgment. The published composite column reproduces the ordering of the plate above by construction.
14 points where AI chip supply structurally narrows. 6 are concentrating, 6 stable and 2 diversifying. Trajectory reports Herfindahl movement without a verdict: a point diversifying away from a near-monopoly narrows export-control leverage for whoever held that lever and widens resilience for buyers downstream. Each Herfindahl here is a single-market figure. The stress index below reports a portfolio-weighted Herfindahl across 93 CSET steps, and Chokepoints reports per-layer figures; the three measure different markets and are not comparable.
| Technology | Category | Top Supplier | HHI | Substitutes | Demand Growth | Trajectory | Priority (editorial) | Gov. Leverage |
|---|---|---|---|---|---|---|---|---|
| EUV Lithography Systems | Lithography | ASML (Netherlands) | 0% | 15% | Stable | 100 NL/EUUSJP | ||
| ABF Substrates (Ajinomoto Build-up Film) | Substrates | Ajinomoto (Japan) | 5% | 20% | Deconcentrating | 54 JP | ||
| Gallium & Germanium (Compound Semiconductors) | Materials | China (aggregated producers) (China) | 10% | 15% | Concentrating | 0 | ||
| EML Laser Chips (800G+ Optical Interconnect) | Optics | Lumentum (United States) | 15% | 160% | Concentrating | 49 US | ||
| High Bandwidth Memory (HBM) | Memory | SK Hynix (South Korea) | 5% | 55% | Stable | 82 US | ||
| Advanced 2.5D Packaging (CoWoS-class) | Packaging | TSMC (Taiwan) | 15% | 40% | Deconcentrating | 57 US | ||
| Chiplet Interconnects (UCIe) | Packaging | TSMC (CoWoS/InFO) (Taiwan) | 10% | 50% | Concentrating | 49 US | ||
| Wafer Inspection & Metrology Equipment | Equipment | KLA Corporation (United States) | 10% | 15% | Concentrating | 92 USNL/EUJP | ||
| Semiconductor-Grade Neon Gas | Materials | Ukraine (Ingas, Cryoin) (Ukraine) | 15% | 10% | Stable | 6 | ||
| US Data Center Grid Interconnection | Infrastructure | PJM Interconnection (United States) | 20% | 35% | Concentrating | 57 US | ||
| EDA Tools (Advanced Node) | EDA | Synopsys (United States) | 5% | 12% | Stable | 96 USNL/EUJP | ||
| EUV Photomasks & Pellicles | Materials | Toppan (Japan) | 5% | 20% | Concentrating | 58 JPUS | ||
| Photoresist (EUV-grade) | Materials | JSR Corporation (Japan) | 10% | 18% | Stable | 55 JP | ||
| Silicon Wafers (300mm) | Materials | Shin-Etsu Chemical (Japan) | 20% | 8% | Stable | 67 JPUSNL/EU |
No competitor has demonstrated EUV capability; ASML order backlog stands at €38.8B (year-end 2025) with €7.4B EUV bookings; SK Hynix committing $8B for ~30 EUV systems through December 2027 (Samsung 20 systems for Pyeongtaek P5, ~$4B). High-NA EUV (0.55 NA) shipping to Intel and TSMC.
The Netherlands holds the sole node for sub-5nm fabrication globally; Dutch export-licensing decisions (operating through the Wassenaar Arrangement) determine which countries can manufacture frontier chips.
Ajinomoto investing ¥25B for 50% capacity expansion by 2030; in early 2026 Ajinomoto announced a ¥1.2B land purchase in Gifu Prefecture for a new factory (construction 2028, operations 2032). Intel launched the first commercial glass core product in early 2026 — a meaningful resolution-path advance vs the prior 2028-2030 estimate, though high-volume glass-substrate qualification for frontier AI accelerators still trails ABF by years. Morgan Stanley projects 42% ABF supply-demand gap by 2028; Ajinomoto AI-driven ABF margins exceed 50%.
5% of this market is a residual bucket, squared here as one supplier. HHI is an upper bound by however much that bucket is itself fragmented.
A single Japanese company supplies the dielectric film for every advanced AI chip package; no US or European producer exists at any scale. Glass substrates are the eventual escape path, but the timeline is 2028 at the earliest.
China export suspension expires November 27, 2026 (a known, calendared deadline). USGS Mineral Commodity Summaries 2026 (the latest periodical) puts China at 99% of primary low-purity gallium production in 2024 and 2025 and confirms China as the leading global producer/exporter of germanium metal in 2025. The military-end-user export ban remains in effect; only the broader prohibition was suspended. USGS modeled GDP impact: complete restriction of China gallium net exports could cut US GDP by $3.1B (range $1.7B-$8.2B); germanium $0.4B (range $0.01B-$1.1B); combined $3.4B (range $1.7B-$9.0B). Non-renewal of the suspension removes the primary global source of gallium (GaN power electronics, GaAs optical components) and germanium (fiber optics, IR optics) from the market.
1% of this market is a residual bucket, squared here as one supplier. HHI is an upper bound by however much that bucket is itself fragmented.
The November 2026 deadline is a binary event sitting inside the broader US-China trade-framework renegotiation, and it falls in the same window as the rare-earth deadline — one negotiation, several levers, all on the same calendar. Risk profiles diverge: gallium (99% China) is a near-monopoly; germanium (~68% China) has more non-Chinese supply but no rapid scale-up path.
800G+ transceiver demand growing 2.6x YoY (24M units 2025 to ~63M 2026). McKinsey June 2025 projects 800G shortfalls 40-60% through 2027 and 1.6T shortfalls 30-40% through 2029. Lumentum holds 50-60% of global EML chip production; demand exceeds Lumentum supply by 25-30%. March 2026 — Lumentum debuted 1.6T DR4 OSFP pluggable transceivers using its 200G/lane EML technology (the binding-constraint chip generation), confirming module-side roadmap but not loosening the chip-fab bottleneck. NVIDIA pre-allocation of supplier capacity persists. Silicon photonics is a substitute on the horizon but needs a different packaging stack. Shares here are chip-level (the binding constraint), not module-level.
15% of this market is a residual bucket, squared here as one supplier. HHI is an upper bound by however much that bucket is itself fragmented.
Optical interconnect capacity is a binding constraint on AI cluster scale-out that GPU procurement timelines do not surface. A facility with GPUs and insufficient 800G+ transceivers runs at degraded throughput, no matter what the rack diagrams say. Lumentum holds ~55% of EML laser-chip production (the transceiver-internal component that cannot be substituted at the module level); US companies hold ~85% of EML chip production between them, but 2.6x YoY demand growth outruns every announced capacity expansion.
Samsung recovered from ~15% (Q2 2025) to ~22% (Q3 2025) after HBM3E qualification with NVIDIA. SK Hynix’s lead narrowed slightly from its Q2 baseline to ~57% but retains clear market leadership. Cross-qualification achieved (Samsung/Micron both NVIDIA-qualified for HBM3E), but ~79% production remains in South Korea.
South Korea holds ~79% of HBM production (SK Hynix + Samsung). A peninsula-wide disruption — grid failure, seismic event, geopolitical break — would erase the majority of AI-accelerator memory supply. Micron (US, ~21%) is the only non-Korean producer, and no allied coordination mechanism exists for crisis-time HBM allocation.
TSMC scaling CoWoS from 75-80K to 120-130K wafers/month by end-2026 across Taiwan facilities (Chiayi AP7 advanced packaging hub phases coming online through 2027); CoWoS yield reported >98% (May 2026); OSAT outsource adding ~240-270K wafers/year (Amkor ~180-190K, SPIL ~60-80K). The binding constraint is allocation concentration (NVIDIA >50% of 2026 capacity; Broadcom >240K wafers; AMD next), not gross capacity.
Taiwan concentration in advanced packaging mirrors the fabrication chokepoint. Even with logic-fab geographic diversification under way (TSMC Arizona, Samsung Taylor), packaging stays Taiwan-centric through at least 2028.
UCIe 1.0 is finalized; chiplet adoption is outrunning interconnect capacity. As monolithic die scaling stalls past 3nm, chiplet architectures stop being optional, and high-bandwidth die-to-die interconnect demand rises in lockstep. Q1 2026 reads show the same concentration shape — TSMC ~55% of advanced 2.5D/3D packaging, Intel ~20% (EMIB/Foveros), Samsung ~10% (I-Cube), OSAT/startups ~15%.
15% of this market is a residual bucket, squared here as one supplier. HHI is an upper bound by however much that bucket is itself fragmented.
Once monolithic die scaling stalls past 3nm, chiplet assembly becomes the binding constraint on next-generation accelerator design. UCIe is open; the packaging capacity that implements it at scale is not — TSMC controls ~55% of advanced 2.5D/3D packaging. The chokepoint is the packaging line, not the standard.
KLA share has continued rising — 73.8% of metrology + inspection in 2025 (up from 72.6% in 2024) per Dr. Robert Castellano process-control coverage; KLA holds an estimated 55-60% of the overall inspection/metrology market and roughly 75-80% in patterned wafer inspection. AI/HPC and HBM-driven defect-budget intensity, EUV/High-NA rollouts, and advanced-packaging adoption are widening the share gap further. Onto Innovation continues gaining in overlay metrology but not in the e-beam inspection segment where KLA dominance is strongest. No Chinese competitor for sub-3nm inspection tools. KLA advanced-packaging revenue estimated $1.3B+ for FY2026 (40%+ growth).
7% of this market is a residual bucket, squared here as one supplier. HHI is an upper bound by however much that bucket is itself fragmented.
US companies hold ~83% of wafer-inspection equipment, KLA alone 60%. Unlike EDA (where the May 2025 episode proved the chokepoint was politically actionable in principle), inspection equipment is already controlled under Wassenaar and BIS rules; it is the enforcement layer that keeps advanced fabs from running rather than the lever waiting to be pulled.
Post-2022 diversification meaningfully cut Ukrainian dependence: by 2026, China has emerged as the largest active producer and commercial supplier of neon (chipmakers across East Asia shifted sourcing to Chinese suppliers during the Ukrainian interruption window). Air Liquide commissioned a new ultra-high-purity neon purification unit in Baton Rouge, Louisiana (+200,000 m³/yr, February 2026); Linde secured a multi-year exclusive ultra-high-purity supply agreement with a major US semiconductor manufacturer (January 2026). Ukraine still supplies a meaningful share, so a fresh infrastructure disruption would tighten DUV laser-gas supply, but the system is more resilient than at the 2022 baseline.
8% of this market is a residual bucket, squared here as one supplier. HHI is an upper bound by however much that bucket is itself fragmented.
Neon feeds the DUV excimer lasers (248nm KrF, 193nm ArF) used for the majority of chip layers even at leading nodes. The 2022 Russia-Ukraine shock priced the vulnerability; diversification since has dampened the concentration without removing it.
PJM interconnection queue has 6,093 entries with ~30% completion rate; PJM TC1 process under FERC Order 2023 compliance ran ~544 days against a 540-day target, yet queue-entry-to-cluster-study-end remains ~1.75 years (exceeds the one-year efficient-process target). FERC issued a Dec 18, 2025 final order directing PJM to file new co-location tariff rules: PJM compliance filings due Jan 20, 2026 (provisional interconnection access, sub-nameplate service, acceleration) and Feb 16, 2026 (service options + procedures for co-located loads). PJM expects to finish reviewing an additional 63,000 MW through 2026. The reforms are progressing but the queue remains binding.
20% of this market is a residual bucket, squared here as one supplier. HHI is an upper bound by however much that bucket is itself fragmented.
Grid interconnection is a binding constraint on US AI-compute expansion that exists independently of GPU supply. Hyperscaler announcements of 1+ GW campuses without interconnection approvals on file are press releases, not capacity plans.
May-July 2025 BIS restriction episode was imposed and rescinded inside six weeks after Chinese rare-earth retaliation. No controls are in force as of mid-2026; the episode left the chokepoint demonstrably actionable on the technical side and demonstrably fragile on the political side. The collective Synopsys + Cadence + Siemens share remains north of 70% per 2026 market reports — Cadence has closed the gap with Synopsys but the US-headquartered duopoly + Siemens (allied) structure is unchanged.
24% of this market is a residual bucket, squared here as one supplier. HHI is an upper bound by however much that bucket is itself fragmented.
US-headquartered vendors hold ~62% of advanced EDA tools; adding Siemens (Germany) brings allied share to ~76%. No burden-sharing framework exists for splitting the cost of Chinese retaliation if EDA restrictions return, which is the load-bearing reason the May-July 2025 episode was politically expensive.
At advanced nodes (<7nm), EUV photomask defect tolerance tightens to sub-nanometer scale, so mask fabrication becomes the yield-binding step. EUV pellicles (the ultrathin membranes protecting the mask during exposure) are still supply-constrained; ASML remains the only producer with qualified EUV-pellicle production. Japan controls over half the world's EUV-grade blanks through Toppan, DNP, and HOYA; DNP is pushing toward EUV photomask mass production in FY2027 (FY2026 onward = production-technology establishment) and signed a joint development agreement with imec for next-generation 2nm photomasks. AGC and HOYA are the two named commercial-delivery-capable EUV-mask-blank suppliers per current industry reports. The EUV mask blanks market is projected to scale from $0.3B (2026) to $1.3B (2035) at 16.5% CAGR — concentration unchanged.
10% of this market is a residual bucket, squared here as one supplier. HHI is an upper bound by however much that bucket is itself fragmented.
Photomask and pellicle concentration is a chokepoint comparable to EDA — invisible in most supply-chain analyses, binding at advanced nodes. Japan controls the mask blanks; ASML controls the pellicles. A disruption to either halts EUV lithography regardless of how many ASML systems are on order. EUV systems sit under Wassenaar; the photomask materials feeding them do not, so the leverage is structural rather than regulatory.
Japan controls ~88% of EUV photoresist production (95% of high-end EUV resists per multiple 2025-2026 trackers). JSR was acquired by Japan Industrial Partners (JIP) in 2024, keeping production domestic; JSR is building a MOR (metal-oxide resist) production facility in South Korea, operational end-2026. TOK is expanding its Koriyama plant (EUV/ArF/KrF), operational H2 2026. Both expansions are allied-jurisdiction — no Chinese-domestic EUV-grade qualification path.
4% of this market is a residual bucket, squared here as one supplier. HHI is an upper bound by however much that bucket is itself fragmented.
Japan controls the chemical inputs for EUV lithography as completely as ASML controls the hardware; Japan has pulled the lever before — photoresist export controls on South Korea in 2019, later relaxed — so this is not a theoretical instrument.
Market structure has been stable for a decade. Japan holds ~54% share (Shin-Etsu + SUMCO, Omdia 2024); 2026 sources put Shin-Etsu at ~28% and SUMCO at ~23% of 300mm output. SUMCO announced termination of 200mm production at its Miyazaki plant by late 2026 to shift capacity into 300mm AI-grade — net effect tightens Japanese concentration in the high-end node, not loosens it. The 2011 Tohoku earthquake remains the calibration precedent: ~25% global supply reduction, 6-month recovery to full capacity.
7% of this market is a residual bucket, squared here as one supplier. HHI is an upper bound by however much that bucket is itself fragmented.
Every semiconductor fab on Earth depends on Japanese silicon-wafer production. EUV is a single-company chokepoint; wafers are not — production spreads across multiple Japanese sites — but Japan's country-level concentration stays at ~54%. Taiwan (GlobalWafers, 17%) is the runner-up producer.
All assessments are Tier 3 (analyst judgment + public data). HHI (Herfindahl-Hirschman Index) is computed from estimated market shares: sum of squared market share percentages on a 0–10,000 scale. Editorial priority composite = HHI × (1 − substitute availability) × (1 + demand growth rate). Substitute availability is an editorial assessment on a 0–1 scale (0 = no substitute exists at any price; 1 = drop-in replacement available at scale). Demand growth rates are annual estimates from industry sources (TrendForce, 650 Group, Omdia). Trajectory assessments are updated quarterly. Because two of its three inputs are hand-assigned, the composite is an editorial ordinal that prioritizes analyst attention, and no measured severity enters it; the formula is deterministic and fully reproducible from the inputs shown in each card. The composite scale is unbounded above 10,000 when demand growth exceeds 0% (the HHI component is capped at 10,000 but the growth multiplier is not). Color thresholds: critical (>7,000), elevated (>4,000), moderate (<4,000).Where multiple competitors are bundled into a residual group, that group is squared as though it were a single supplier, so the HHI is an upper bound relative to the actual fragmentation among them. That is the convention in force here: 11 of 14 rows close their share list with such a bucket, ranging 1 to 24 points, and each row’s bucket is shown with its share in the outlook card below so the overstatement can be bounded directly. The opposite convention is implemented and currently dormant: were a row’s known shares to sum below 100%, its HHI would be a lower bound and marked with an asterisk. No row is in that state today; all shares sum to 100%.
Sources: ASML, TSMC, SK Hynix annual reports and earnings calls; TrendForce, Omdia, 650 Group / LightCounting, Counterpoint Research market trackers; USGS Mineral Commodity Summaries 2025; MOFCOM Announcement No. 72/2025; PJM Interconnection Queue data; SemiAnalysis; Ajinomoto Co. Annual Report 2025.
The Governance Leverage Score (0–100) reads as the share of structural policy surface allied export-control regimes can, in principle, reach over a given chokepoint; a score of 87 means 87% of the leverage points over that node sit inside an existing jurisdictional or regulatory mechanism. It reads as capacity, and recommends nothing about using that capacity. Allied coordination is structurally feasible at the top of the range (EUV lithography, HBM memory, wafer-inspection equipment). At the bottom (gallium, neon) the leverage points fall outside allied reach, where the available levers are alternative-supply development and diplomacy.
Composite of three weighted components. (1) Allied Production Share (50% weight): sum of market-share percentages from suppliers headquartered in countries with ‘full’ or ‘allied’ compute-access tier; suppliers coded “Various” are excluded (conservative lower bound). (2) Regime Jurisdiction Count (20% weight): how many of the three tracked regimes (US BIS, Netherlands/EU, Japan METI) have jurisdiction over at least one supplier with >10% market share, determined by supplier HQ location; South Korea and Taiwan map to the US regime via FDPR extraterritorial reach over items produced using US-origin technology. (3) Active Export Controls × Penetrability (30% weight): whether the chokepoint’s technology category has active restrictions in the regime-overlap matrix, weighted by an editorial penetrability assessment: impenetrable (1.0), very difficult (0.85), difficult (0.65), porous (0.35), unrestricted (0.0).
Caveats. Enforcement effectiveness, existing stockpiles, and the timeline for alternative-supply build-out are not in the score. Penetrability is editorial (Tier 3). A residual bucket coded to no single jurisdiction is excluded from allied share, which understates that share wherever the bundled companies are individually identifiable as allied: 9 nodes have one, the largest being EML Laser Chips (800G+ Optical Interconnect) at 15 points and Chiplet Interconnects (UCIe) at 15 points. Where the residual is instead assigned to a jurisdiction it is counted in full, which is the opposite exposure and the one a reader cannot see from the supplier list: 24 of EDA Tools (Advanced Node)’s allied points are a bucket coded to United States for its dominant remaining supplier, and 20 of US Data Center Grid Interconnection’s allied points are a bucket coded to United States for its dominant remaining supplier. FDPR extraterritorial jurisdiction is treated as equivalent to domestic jurisdiction, which overstates enforceability in the other direction.
How scores shift under alternative weighting of the three components (allied share / regime jurisdiction / active controls). Scores move by up to 26 points. The ordering is not invariant: production-heavy weighting changes 3 pairs, and controls-heavy weighting changes one pair. Equal weighting reorders nothing.
| Technology | Default (50/20/30) | Equal (33/33/34) | Production-heavy (60/10/30) | Controls-heavy (30/20/50) |
|---|---|---|---|---|
| EUV Lithography Systems | 100 | 100 | 100 | 100 |
| ABF Substrates (Ajinomoto Build-up Film) | 54 | 42 | 60 | 35 |
| Gallium & Germanium (Compound Semiconductors) | 0 | 0 | 0 | 0 |
| EML Laser Chips (800G+ Optical Interconnect) | 49 | 39 | 54 | 32 |
| High Bandwidth Memory (HBM) | 82 | 73 | 89 | 79 |
| Advanced 2.5D Packaging (CoWoS-class) | 57 | 44 | 63 | 37 |
| Chiplet Interconnects (UCIe) | 49 | 39 | 54 | 32 |
| Wafer Inspection & Metrology Equipment | 92 | 93 | 91 | 90 |
| Semiconductor-Grade Neon Gas | 6 | 4 | 7 | 4 |
| US Data Center Grid Interconnection | 57 | 44 | 63 | 37 |
| EDA Tools (Advanced Node) | 96 | 95 | 96 | 93 |
| EUV Photomasks & Pellicles | 58 | 52 | 61 | 40 |
| Photoresist (EUV-grade) | 55 | 43 | 61 | 35 |
| Silicon Wafers (300mm) | 67 | 64 | 66 | 48 |
A single 0–100 composite across four sub-indices: demand (WSTS monthly billings, three-month moving average, year on year), concentration (CSET weighted Herfindahl, the same construct the Chokepoints ladder distinguishes from its per-layer figures), capacity (FRED semiconductor utilization), and power (the PJM data-center-region queue-to-operational ratio). Weights default to equal and are adjustable; the monthly series begins at the earliest date all four sub-index sources are available, and its vintage is 2026-01. Two of the four sub-indices update monthly and two are held constant: demand and capacity move on WSTS billings and FRED utilization, while concentration and power read single-vintage snapshots fixed at that date. The two flat lines in the series below are those two, and no part of the composite is a forecast.
How concentrated the AI-accelerator designer market is, quarter by quarter: a Herfindahl over compute-weighted cumulative shipment shares across Nvidia, Google, AMD, Huawei, Amazon and Cambricon. This is the one quarterly concentration series on the page, and a distinct construct from the per-market figures above and the stress index’s portfolio-weighted Herfindahl. The shaded band is a Scrutica derivation over Epoch’s inputs. It spans the least and most concentrated Herfindahl reachable when each designer’s cumulative shipments move independently across the p5 to p95 range Epoch gives for that designer, so its width tracks how many designers are in the cross-section and how uneven their shares are. It does not measure the looseness of any one estimate. The series falls from a near-monopoly toward a broader field as entrants ship; late quarters flagged incomplete have fewer confirmed designers, so a late tick upward is an artifact of reporting lag.
Source: Epoch AI · AI Chip Sales (CC-BY 4.0). Herfindahl over H100-equivalent-weighted cumulative shipment shares by designer; band derived by Scrutica as the Herfindahls reachable across Epoch’s per-designer p5 to p95 shipment ranges. The per-market Herfindahls in the plate and table above are point-in-time editorial estimates. No supply-chain-layer or facility-capacity concentration series is derivable, because the append-only capacity snapshots track individual-facility build-out and have no market cross-section; none is manufactured.
Three detectors run over 29 ownership-change deals across the 7 layers the ownership feed classifies (CoWoS packaging, chiplets, EDA, advanced substrates, specialty photoresists, datacenter liquid cooling, wide-bandgap power electronics). That set overlaps the tracked chokepoints above without coinciding with them: liquid cooling and power electronics are not tracked as chokepoints, and most tracked chokepoints have no layer here. The detectors are per-layer quarterly anomalies (z > 2 vs the trailing eight-quarter mean), consolidation clusters (one acquirer, three or more deals in a rolling twelve months), and cross-layer actors (one acquirer moving in two or more layers). No anomaly has fired in the recent window: M&A activity in these layers has stayed within its historical baseline. The archive preserves the historical patterns the detectors surfaced retroactively.
The feed aggregates ownership-change events (Buyout, M&A, Corporate Divestiture, Asset Sale, Public-to-Private, Secondary Buyout, Spin-Off, Reverse Merger, Joint Venture) from a licensed corporate-ownership database, filtered to the AI-compute chokepoint universe via per-organization layer classifications, and aggregated by quarter. Three detectors run over that series:
Anomalies persist archivally with stable identifiers. The date badge distinguishes recent activity (≤ 18 months) from historical patterns the detectors surface retroactively; an “open” anomaly can reference a quarter from decades ago, so the recent counts above are the subset whose reference window lands in the trailing eighteen months. Disclosure of deal dollar amounts runs about 46%; the event counts are complete. All classification and anomaly assessment is Tier 2–3 (a licensed research database plus rule-defined detection over it).