Scrutica
| Layer | Veto · 18mo — Can a holder halt a run within 18 months? YES when switching-cost × redeployment-time exceeds the horizon. NO when the lab can route around them. UNCERTAIN when the routing time straddles the horizon. | Veto · 5yr — Can a holder halt structural independence over 5 years? Same rule, longer horizon — a closing window (HBM) reads YES→NO, a permanent lock (EUV) reads YES→YES. | Checked by — Accountability is the strongest governance hook over the layer under a fixed priority order, assigned per layer against the rule: EXPORT_CONTROL (a BIS/Dutch/multilateral export-control regime governs the layer — directional: it governs diversion, not a domestic lab’s purchase), ANTITRUST (a named authority with a live or available proceeding), PUBLIC (the holder is itself a public body), COMPETITION (no legal or public hook, but an alternative supplier exists in principle), or NONE (no hook of any kind — the alarm set). Whether a competitor is reachable within a horizon lives in the veto, not this tag. | Stated rule | Redeployment time — Time to route around the holder. Combined with switching cost, this is the leverage metric (not market share). Tier-4 values (no source found) are flagged "needs verification"; sourced-but-estimated values are flagged "(estimated)". |
|---|---|---|---|---|---|
| EUV lithography · ASML | YES | YES | EXPORT_CONTROL | A 100% monopoly, but governed by Dutch + US multilateral export-control law — a governance hook exists over the layer, so it is checked-in-principle rather than fully uncontested. | No commercial alternative at any horizon for a Western lab; state-built EUV at scale assessed 5-10+ years out (estimated) |
| Advanced logic foundry (≤5nm/3nm) · TSMC; Samsung Foundry; Intel Foundry | YES | UNCERTAIN | EXPORT_CONTROL | Leading-edge logic fabrication is governed by US export-control law: the Foreign Direct Product Rule reaches any chip made with US tooling, and the Jan-2025 Foundry Due-Diligence Rule mandates quarterly OSAT/foundry KYC compliance reporting to BIS — a public lever exists over the layer. | No published figure for a full TSMC→Samsung inter-foundry port of a leading-edge AI accelerator (estimated — needs verification) |
| HBM memory (HBM3E / HBM4) · SK Hynix; Samsung; Micron | YES | NO | EXPORT_CONTROL | HBM is governed by US export-control law — the Dec-2024 rule controls HBM at/above a bandwidth-density threshold (>2 GB/s/mm²), on top of the Oct-2022/Oct-2023 advanced-computing controls — a public lever exists over the layer. | 12-24 months to qualify a secondary supplier, anchored to the Samsung→NVIDIA case (~18-month qualification for a supplier with a finished product) (estimated) |
| Advanced packaging / CoWoS · TSMC; OSATs (ASE, Amkor, SPIL) | YES | UNCERTAIN | NONE | No export-control regime over the packaging service, no live antitrust proceeding, not a public body — no governance hook of any kind. The sole layer with no public lever. | New OSAT production line ~6-9 months; customer-level requalification for a given accelerator unverified (estimated — needs verification) |
| AI accelerator · NVIDIA; Google TPU; AWS Trainium; AMD | YES | UNCERTAIN | EXPORT_CONTROL | Frontier AI accelerators are the primary object of US export-control law (ECCN 3A090/4A090; the Oct-2023 Total-Processing-Performance metric closed the A800/H800 work-arounds) — the most-governed layer in the stack. | Committed labs (CUDA-bespoke): >18 months. Diversified/vertically-integrated labs: already routed around NVIDIA at production scale (not applicable). (estimated) |
| Scale-up interconnect (NVLink / NVSwitch) · NVIDIA | YES | UNCERTAIN | EXPORT_CONTROL | Reached by the same US advanced-computing export-control architecture as the accelerator it binds: ECCN 3A090 carries an I/O-bandwidth control parameter (≥150 GB/s aggregate bidirectional) and 4A090 covers electronic assemblies/components incorporating controlled accelerators (e.g. GB200 NVL72 racks) — the scale-up fabric is governed by the regime, not merely by an emerging market alternative. | No production UALink alternative until ~late 2026 at earliest (committed labs) (estimated) |
| Scale-out interconnect (InfiniBand vs Ethernet) · NVIDIA/Mellanox InfiniBand; Broadcom/Arista Ethernet | NO | NO | COMPETITION | Ethernet (Broadcom, Arista, Ultra Ethernet Consortium) is a production-scale alternative; demonstrated at frontier scale. | Routable to Ethernet now for a lab with fabric-engineering capacity (estimated) |
| Cloud / datacenter capacity · AWS; Microsoft Azure; Google Cloud; CoreWeave; Oracle; neoclouds | UNCERTAIN | NO | COMPETITION | Capacity is fragmenting (97GW of new colocation projected 2025-30); multi-vendor structure means no single cloud provider is a unilateral veto for a top lab. | Binding for single-vendor-locked startups; not for multi-vendor pre-positioned labs (estimated) |
| Power / grid interconnection · ISO/RTO (MISO, PJM, ERCOT, SPP); local utility; state PUC; FERC; EPA | YES | UNCERTAIN | PUBLIC | The veto-holders are public bodies (ISO/RTO, utilities, state PUCs, FERC, EPA) subject to democratic/regulatory/FOIA accountability — a DIFFERENT KIND of power than a private monopoly. | New site: ~4.5yr average to commercial operation (transformer ~160wk OR interconnection study alone exceeds 18 months). Already-energized clusters: veto already cleared. (estimated) |
| Capital ($1B+/yr runs) · hyperscaler balance sheets; sovereign funds (QIA, PIF, Mubadala); mega-VC | UNCERTAIN | NO | COMPETITION | Capital access is structurally expanding (sovereigns + infrastructure investors entering); diversified + compute-entangled funding means no single provider is a unilateral veto for a top lab. | Single-funder veto plausible only for <5 single-funder-dependent startups; implausible for hyperscalers/multi-investor labs (estimated) |
1 leverage point can halt a committed lab’s frontier run within 18 months and answer to no one: Advanced packaging / CoWoS.
The same 10 layers, now on two axes: can-halt / uncertain / cannot-halt down the side, checked / unchecked across the top. The whole argument lives in one cell, the top-right (can halt a run, and nothing checks the holder), and the only question is whether anything occupies it at the run and horizon you have selected in the ledger above. Change either and the chips re-settle into new cells as a lab routes around a holder. Share is kept off the grid on purpose: a 100% share a lab can re-source within a quarter is no veto at all, while a smaller share no one can qualify around in time would be one. The criterion here is leverage, not size.
The reading runs up and to the right. The top-right cell — can halt a run, and unchecked — is the power-concentration finding; everything to its left can still halt a run but answers to something (a competitor a lab could reach, an export regime, or a public body), and everything below it cannot halt a run inside the horizon at all.
Concentrated-and-unaccountable matrix at 18 months for a committed lab. 1 leverage point both can halt a run and are unchecked: Advanced packaging / CoWoS, held by TSMC and OSATs (ASE, Amkor, SPIL). Can halt: 7. Uncertain: 2. Cannot halt: 1.
The naive story holds that NVIDIA’s share isthe chokepoint, so every frontier run carries an NVIDIA veto. Put that way it is falsifiable, and it fails. Run the three real cases (xAI on NVIDIA, Google on pure TPU, Anthropic on a Trainium-and-TPU mix) and NVIDIA’s accelerator veto binds the lab that committed to CUDA while dropping out of the two that did not. Were the veto count flat across runs, this instrument’s leverage metric would have collapsed back into market share; it moves by 2 at 18 months, which is the test passing.
NVIDIA, the most-named chokepoint of the lot, vetoes only xAI — the lab that committed to it — and drops out of Google and Anthropic, which run a non-NVIDIA training path (TPU, Trainium). The total veto count moves by 2 across runs at 18 months. A pure market-share read sees none of this, because the divergence lives in the 3 leverage points that re-resolve per run, not in any share table.
From the accelerator leverage point: DIRECTIONAL: the regime governs export/diversion, not a domestic lab’s GPU access. FALSIFICATION RESULT (the orthogonal finding, carried by the per-run veto below, NOT by this tag): NVIDIA — the most-named chokepoint — vetoes only the labs that chose it; Google (TPU) and Anthropic (Trainium/TPU) train with no NVIDIA in the path, so the default 18mo veto applies to NVIDIA-committed labs (xAI, OpenAI) only.
A struck-through cell is a veto that binds the committed baseline yet drops out for that run, meaning the holder cannot halt it. Pivot rows (tinted) are the layers that re-resolve per run; the rest of the stack reads identically across all three. One pivot has nothing to do with chip choice. The power / grid veto clears for every run here because each sits on already-energised capacity; it would re-bind a new most-locked build, which is why the cross-run baseline still counts it.
NVIDIA falsification test at 18 months. Committed-baseline veto count: 7. xAI Colossus (Memphis): 6 vetoes, with Power / grid interconnection dropping out. Google Gemini: 4 vetoes, with AI accelerator and Scale-up interconnect (NVLink / NVSwitch) and Power / grid interconnection dropping out. Anthropic Claude: 5 vetoes, with AI accelerator and Power / grid interconnection dropping out. The veto-count spread across runs is 2, so the prediction held: the most-cited chokepoint does not bind every run.
Each veto is read at two horizons — whether a holder can stop a planned run inside 18 months, and whether it can block structural independence over five years — and the divergence between the two is what this view exists to show. A flat line held high is a permanent lock that no second supplier dissolves within five years (EUV lithography, ASML alone); a line falling toward green is a window open for only the next 18 months (HBM, once a third qualified supplier matures). What moves a line between its endpoints is the redeployment time on each row, carried through with its verification status intact.
This band holds the unchecked veto — Advanced packaging / CoWoS: the only leverage point that can halt a run today and answers to no one. An eroding trajectory does not make it accountable — only reachable, eventually.
Horizon divergence across 10 leverage points. Permanent lock: EUV lithography (YES at 18 months, YES at 5 years). Eroding — contested by 5yr: Advanced logic foundry (≤5nm/3nm) (YES at 18 months, UNCERTAIN at 5 years); Advanced packaging / CoWoS (YES at 18 months, UNCERTAIN at 5 years); AI accelerator (YES at 18 months, UNCERTAIN at 5 years); Scale-up interconnect (NVLink / NVSwitch) (YES at 18 months, UNCERTAIN at 5 years); Power / grid interconnection (YES at 18 months, UNCERTAIN at 5 years). Closing window: HBM memory (HBM3E / HBM4) (YES at 18 months, NO at 5 years). Not binding — or loosening: Scale-out interconnect (InfiniBand vs Ethernet) (NO at 18 months, NO at 5 years); Cloud / datacenter capacity (UNCERTAIN at 18 months, NO at 5 years); Capital ($1B+/yr runs) (UNCERTAIN at 18 months, NO at 5 years).
The view an analyst opens before citing the surface. Every verdict reduces to two physical inputs per layer, switching cost and redeployment time, plus one categorical accountability rule; there is no blended score anywhere, by design. All of it is here, sourced and tiered, down to the two redeployment numbers no primary source would commit to, which carry a marker instead of a clean figure. The leverage definition sits just below; everything else (every coefficient, the taxonomy, the falsifier behind each verdict, the construct ceiling) is one click away in its section. The construct ceiling closes the panel rather than hiding in it, because it is the caveat most worth reading: what the instrument measures is control of inputs, and the step from there to “power” is an argument, contestable at the rule level.
Leverage = switching-cost × redeployment-time, not market share.
A layer holds a veto over a run, at a given horizon, when two conditions hold together: a dominant or sole actor exists, and the lab cannot route around that actor inside the horizon, where the cost of routing around is switching cost times redeployment time. Share is shown as context and never enters the verdict, because a share-based criterion assumes the thing it is meant to test. A 100% share a lab can re-source in a quarter is no veto at all; a smaller share no one can qualify around in time wouldbe. The criterion is leverage, not size. Every veto-holder on this stack also happens to be a near-monopoly, so the distinction earns its keep not in some low-share veto (there is none here) but in the per-run falsification, where NVIDIA’s commanding accelerator share is not a veto for the labs that route around it. What the instrument measures is the asymmetry, the ability to make a lab wait, and the two horizons exist because that ability is time-scoped: 18 months asks whether a holder can stop a planned or in-progress run, 5 years whether it can block a lab from building independence.
One veto in the set is held not by a firm but by a stack of institutions acting as a single gate — power and grid interconnection, where an ISO or RTO, the local utility, a state public-utility commission, FERC, and EPA must each clear a new frontier-scale site. The instrument counts that stack as one gate (a run is halted whether it is one veto or five that stops it) but flags it as an institutional stack rather than a private monopoly, because the two are different kinds of concentration and should never be silently summed into the same number.
Every coefficient is below with its source, authority tier, and vintage; click any value to see them. Two of the redeployment numbers (the time to port a leading-edge accelerator from TSMC to another foundry, and the time to requalify a customer at an alternative packager) carry a needs verification marker rather than a clean figure, because no primary or secondary source would commit to one; the sources acknowledge the barrier and decline the number. They are shown, not hidden, and they matter less than their prominence suggests, because the 18-month vetoes they sit under rest on observable capacity allocation (leading-edge foundry booked through 2026, NVIDIA holding more than half of CoWoS capacity through 2027) rather than on the missing figure. The number that would move a verdict if it were sourced is the five-year one, which is why those two layers read ?· Uncertain at five years instead of YES or NO.
Two numbers on this surface have no primary or secondary source, and rather than estimate them we mark them. The first is the time to port a leading-edge AI accelerator from TSMC to another foundry: the closest published proxy is a same-foundry tape-out-to-volume figure of roughly 12-to-18 months (Electrek, on Tesla's AI5), itself a partial lower bound, since an inter-foundry port adds full PDK re-implementation and re-qualification on top. An earlier “18–36 months” figure that circulated was unsourced inference, and it has been retired rather than dressed up. The second is the customer-level requalification time at an alternative packager versus TSMC's CoWoS, which the sources acknowledge as a barrier and decline to quantify. Both carry a needs verification marker, and a dedicated verification pass owns them before any publish gate. Neither changes the near-horizon reading: both are five-year-relevant, because the 18-month vetoes they sit under rest on capacity allocation, which is observable.
Where a number is genuinely estimated-but-sourced — a modeled market share, an analytical synthesis — it carries a quieter est. marker instead, so “we synthesised this from a real source” never reads as “this is unverified,” and neither reads as the concentration alarm. The flag tracks Scrutica-side derivation, not source authority: a figure a source reports directly (a site’s megawattage, say) is not estimated even at Tier 3, because we did not derive it; a figure we modeled or synthesised is, even from a Tier-1 or Tier-2 source.
What checks a veto-holder is a single categorical tag with a stated, contestable trigger, never a number, because a blended accountability score is the failure mode this instrument was built to avoid. The tag records the strongestgovernance hook over the layer under a fixed priority order (export control › antitrust › public body › competition › none); each one is assigned per layer against that rule and audited, never computed. The definitions are below, in alarm-first order, and each layer's specific application is one expand away, or in the ledger's “Checked by” column.
One distinction carries more weight than any other on this surface, so it gets said out loud: in-horizon reachability lives in the veto, not in the tag. Accountability answers a question the veto does not, “if this holder exploited its position, what could check it, in principle, at any timescale?”, and it stays silent on “can a lab route around it in time?” A competitor that exists but cannot be reached inside 18 months is therefore a veto of YES carrying a competition hook, which is a different thing from an absence of accountability. Collapse the two and a reachability fact starts masquerading as a governance fact, which is the one substitution this instrument cannot afford.
A worked consequence: leading-edge foundry, HBM, and the AI accelerator are all tagged export control, not because a domestic lab's access is governed (it is not) but because a public lever exists over the layer — the Foreign Direct Product Rule and the January-2025 foundry due-diligence rule over fabrication, the December-2024 bandwidth-density threshold over HBM, the 3A090/4A090 controls and the 2023 Total-Processing-Performance metric over accelerators. Advanced packaging is the layer where that pattern breaks: the December-2024 rule added packaging equipment to the controlled list but not the packaging service, so no export regime reaches TSMC's allocation of CoWoS capacity among Western customers, which is why it, alone, is tagged none.
A 100% monopoly, but governed by Dutch + US multilateral export-control law — a governance hook exists over the layer, so it is checked-in-principle rather than fully uncontested.
Leverage acts at the national-policy timescale (a lab keeps its fabbed chips unless a government acts, with notice/process), not the individual-lab commercial timescale.
Leading-edge logic fabrication is governed by US export-control law: the Foreign Direct Product Rule reaches any chip made with US tooling, and the Jan-2025 Foundry Due-Diligence Rule mandates quarterly OSAT/foundry KYC compliance reporting to BIS — a public lever exists over the layer.
DIRECTIONAL: the hook governs diversion (whom TSMC may fabricate for, where advanced capacity may expand — FDPR, CHIPS-Act guardrails), NOT a domestic lab’s access or TSMC’s allocation among Western customers. Samsung/Intel exist as competitors but are not switchable within 18 months — that unreachability is the VETO=YES, not an absence of accountability.
HBM is governed by US export-control law — the Dec-2024 rule controls HBM at/above a bandwidth-density threshold (>2 GB/s/mm²), on top of the Oct-2022/Oct-2023 advanced-computing controls — a public lever exists over the layer.
DIRECTIONAL: governs diversion, not a domestic lab’s supply. Three functional suppliers (SK Hynix, Samsung, Micron) also exist — a market check that matures by 5yr; that maturation is the VETO trajectory (18mo YES → 5yr NO), not the accountability.
No export-control regime over the packaging service, no live antitrust proceeding, not a public body — no governance hook of any kind. The sole layer with no public lever.
The honest distinction from foundry/HBM/accelerator: the Dec-2024 BIS rule added advanced-packaging EQUIPMENT to controlled items, but not the packaging SERVICE — so no export regime reaches TSMC’s allocation of CoWoS capacity among Western customers. Whether a rival packager could be reached in time is a VETO question, deliberately excluded here. The acute lab-level commercial veto — more operationally binding than ASML, and the sole member of the power-concentration alarm set.
Frontier AI accelerators are the primary object of US export-control law (ECCN 3A090/4A090; the Oct-2023 Total-Processing-Performance metric closed the A800/H800 work-arounds) — the most-governed layer in the stack.
DIRECTIONAL: the regime governs export/diversion, not a domestic lab’s GPU access. FALSIFICATION RESULT (the orthogonal finding, carried by the per-run veto below, NOT by this tag): NVIDIA — the most-named chokepoint — vetoes only the labs that chose it; Google (TPU) and Anthropic (Trainium/TPU) train with no NVIDIA in the path, so the default 18mo veto applies to NVIDIA-committed labs (xAI, OpenAI) only.
Reached by the same US advanced-computing export-control architecture as the accelerator it binds: ECCN 3A090 carries an I/O-bandwidth control parameter (≥150 GB/s aggregate bidirectional) and 4A090 covers electronic assemblies/components incorporating controlled accelerators (e.g. GB200 NVL72 racks) — the scale-up fabric is governed by the regime, not merely by an emerging market alternative.
DIRECTIONAL and reached INDIRECTLY — NVLink/NVSwitch is not a standalone ECCN line item; it is controlled via the 3A090 I/O-bandwidth parameter and as part of 4A090 assemblies — so the hook governs diversion, not a domestic lab’s purchase. The in-horizon market check (UALink first switch hardware ~late 2026; Google’s TPU pods use proprietary ICI and need no NVLink) is the reachability story, which lives in the VETO, not this tag; the priority order tags it export control over competition.
Ethernet (Broadcom, Arista, Ultra Ethernet Consortium) is a production-scale alternative; demonstrated at frontier scale.
Capacity is fragmenting (97GW of new colocation projected 2025-30); multi-vendor structure means no single cloud provider is a unilateral veto for a top lab.
The veto-holders are public bodies (ISO/RTO, utilities, state PUCs, FERC, EPA) subject to democratic/regulatory/FOIA accountability — a DIFFERENT KIND of power than a private monopoly.
Applies to NEW capacity. The strongest 18mo veto for a new site — but in-principle accountable, so it is a contrast case to the private-unchecked cluster, not the same risk.
Capital access is structurally expanding (sovereigns + infrastructure investors entering); diversified + compute-entangled funding means no single provider is a unilateral veto for a top lab.
A verdict is robust when its leverage inputs sit far from the horizon; it is sensitive when the redeployment time straddles the 18-month or 5-year line. Those cases read ?· Uncertainat one horizon here — an honest “could go either way” rather than a forced call — and a stress test would move them first. Where the driver is also unsourced (marked below), the uncertainty is doubled: the verdict rests on observable capacity allocation, not on the missing figure.
needs verificationRedeployment time unsourced — the 18-month veto rests on capacity allocation, not this number.
needs verificationRedeployment time unsourced — the 18-month veto rests on capacity allocation, not this number.
Each binding veto below is paired with the single observation that would overturn it; a verdict whose flip-condition is far out of reach is robust, one whose flip-condition sits close to the horizon is brittle.
YES at both horizons — the permanent lock.
Falsified by a second commercial EUV supplier reaching frontier scale, or by DUV multipatterning at sub-5nm becoming yield-competitive. Neither is on a five-year horizon for a Western lab; state-built EUV at scale is assessed five-to-ten-plus years out. This is the verdict policy is the only lever for.
YES at 18 months, UNCERTAIN at 5 years.
The 18-month veto rests on capacity — leading-edge nodes booked through 2026 — not on a redeployment figure, so it would be falsified by leading-edge capacity opening up, not by a faster port. The five-year UNCERTAIN is where the unsourced inter-foundry port time bites: source it low and this reads NO at five years, source it high and it reads YES.
YES at 18 months, NO at 5 years — the closing window.
The 18-month veto would be falsified by a lab qualifying a secondary HBM supplier faster than the ~18-month Samsung-to-NVIDIA qualification case suggests; the five-year NO would be falsified by the three-supplier market failing to mature (consolidation, a sustained sell-out). What the verdict asserts is the trajectory between those two endpoints, not either one read on its own.
YES at 18 months, UNCERTAIN at 5 years — the sole unchecked veto.
The 18-month veto rests on allocation (NVIDIA holding more than half of CoWoS capacity through 2027, OSATs fully booked), so it would be falsified by free qualified capacity appearing at an alternative packager, not by the unsourced requalification time, which only governs the five-year reading. The accountability tag of NONE falsifies on a separate axis entirely: an export-control hook reaching the packaging service, or a live antitrust proceeding, or the holder becoming a public body. An eroding veto does not make this layer accountable; it makes it reachable, eventually, which is a different fact.
YES in the default, NO for the labs that routed around NVIDIA.
The default 18-month veto is falsified, per lab, by that lab running a non-NVIDIA training path at production scale — which Google (TPU) and Anthropic (Trainium and TPU) already do, and which is why their cells read NO. The verdict tracks each lab's committed training path, not NVIDIA's share of the market, and the surface resolves it per run for exactly that reason.
YES in the default at 18 months.
Falsified by a production-ready UALink alternative, targeted around late 2026, or — already, for Google — by a proprietary pod fabric (ICI) that needs no NVLink at all.
YES for a new site at 18 months; cleared for the three named runs.
This is the verdict most sensitive to framing, so the framing is explicit: the 18-month veto applies to a new most-locked build — interconnection alone exceeding 18 months, transformer lead times around 160 weeks, against a ~4.5-year average to commercial operation for 2024 projects of all types. It does not halt the three real runs, which sit on already-energised capacity (the per-run override reads NO for all three), and it re-binds only when a lab needs new power. Falsified by interconnection reform that compresses the queue, or by behind-the-meter generation that bypasses it — itself now gated by a January-2026 EPA rule.
The three layers that never bind — scale-out interconnect (Ethernet is routable now), cloud capacity (multi-vendor by design, fragmenting), and capital (diversified and compute-entangled) — are in the instrument deliberately, so the stack is not read as uniformly concentrated. A veto count that found a veto everywhere would be measuring its own assumptions.
What the veto count measures is control of AI's inputs: which actors can make a frontier lab wait, and for how long. It does not measure power, and the gap between the two is where most misreadings of this surface will start. Input-control is necessary for leverage without being sufficient for it: a holder who can stall a training run is not thereby able to dictate what the lab trains, while a layer holding no veto at all can still shape outcomes through price, through queue priority, or through the standing relationship that decides who gets allocated first when capacity runs short. The veto count is a floor on the concentration of leverage, not a reading of power; treat it as the former and the surface is rigorous, treat it as the latter and you have over-read it.
The bridge from “who can halt a run” to “who holds power” is therefore an argument, and it is left contestable by design. The contestable joint is the accountability rule: every tag carries its stated trigger, so the disagreement is never with a verdict you cannot inspect but with a rule you can — and if you reject a rule, the verdict moves with it. That is the intended mode of use. Argue the rule, not the number.
Training a frontier model runs through a long supply chain: the chips, the handful of machines that can make them, the packaging that fuses them, and the power to run it all. At seven points along that chain, a single player can stall the work for a year or more. Six of the seven sit under some check, whether an export-control rule or a public body. The last answers to no one.
That one is advanced packaging: the step that fuses bare silicon into a working accelerator, controlled by TSMC. No export rule reaches it and no antitrust case touches it; it answers only to the customers in its queue.
Most of these loosen within five years as rivals qualify. One does not: chip lithography, where ASML remains the only supplier in sight.
This is the brief. Switch to Analyst view for the full instrument: every holder, the leverage behind each verdict, the sources, and the methodology.