Cascade Simulation — Disruption Propagation Through the AI Compute Supply Chain

Gringras, David

Scrutica

Graph written 2026-09-06

Cascade

A single supplier failure, walked outward through the firms that depend on it.

The cascade engine propagates a failure over 14,937 supplier–customer edges across 5,080 organizations

The cascade engine

Scenario definitions

The reach curve. Each organization in the graph you can inspect, ranked left to right by how many others a failure at it reaches within 3 hops downstream. The curve is a lossless step series, so every rank in the graph is on it.

The walk is unweighted, counting everyone structurally downstream without regard to how much impact survives the trip, so every value is an upper bound on the cascade engine’s weighted run. It runs over the inspectable subset, a strict subgraph of the 14,937-edge graph the engine propagates over, and removing edges can only reduce reach, so every value is at the same time a lower bound on structural reach in the engine’s own graph.

The graph also includes historical relationships, and a few counterparties named at the head have since been acquired or dissolved.

Open the cascade engine

A disruption travels by reaching someone who depends on it, then someone who depends on them. Averaged across the inspectable graph, a failure touches 1.8 organizations at one hop, 10 within two, and 32 within three; the fan-out is why one chokepoint is a systemic question. The mean conceals the distribution, though: the curve has a short head, then a long decline that ends at a cliff. 117 organizations reach a hundred or more; 305 reach nobody. A seed’s place on that curve decides whether a run is a systemic event or a local one, and the cascade engine above runs from any of them. How propagation is derived

Propagated edges, inspectable edges, stored rows

Three numbers on this page all sound like “how many links”, and they answer different questions.

14,937: edges the cascade engine propagates over. The full-density graph, spanning 5,080 organizations. This is the graph every simulation on this page actually runs against, and the figure the simulation API reports back with each result.
2,816: edges you can inspect. The public-provenance subset served to the browser, spanning 1,498 organizations: the graph drawn above, downloadable edge by edge, and the graph the reach curve is walked over. The rest of the cascade engine’s graph is held under licence: it propagates, but its edges are not redistributable.
18,980: stored relationship rows. Every row in the corpus, including 4,043 that either duplicate another source’s record of the same relationship or record an unnamed counterparty aggregate.

The pre-built scenarios 20 seeds

Severity is the impact magnitude at the seed; decay attenuates it per supply-chain hop, and every decay value is editorial (Tier 4). Severities are sourced from SEC 10-K filings, USGS Mineral Commodity Summaries and TrendForce, IDC and SemiAnalysis reporting, with editorial assessments where no primary source exists.
Scenario	Category	Seed nodes	Severity	Tier	Decay
TSMC Taiwan Disruption	Natural disaster	TSMC	1.00	T4	0.90
ASML Export Halt	Export control	ASML	1.00	T1	0.95
US–China Full Decouple: Revenue Impact	Geopolitical	NVIDIA, AMD, Intel, Qualcomm Inc	0.15	T1	0.70
US–China Full Decouple: Access Denial	Geopolitical	SMIC, Huawei Technologies, Baidu, ByteDance, Alibaba, Tencent Holdings Ltd.	0.55	T3	0.80
CoWoS Advanced Packaging Bottleneck	Supply bottleneck	TSMC	0.40	T2	0.70
SK Hynix HBM Disruption	Supply bottleneck	SK Hynix	0.60	T2	0.80
EDA Export Restriction	Export control	Synopsys, Cadence	0.90	T3	0.80
Chiplet Assembly Disruption	Supply bottleneck	TSMC, Synopsys, Cadence	0.80	T4	0.75
Helium Supply Crisis (Hormuz)	Supply bottleneck	Samsung, SK Hynix, TSMC	0.65	T1	0.75
InP/EML Laser Shortage	Supply bottleneck	Coherent, Lumentum Holdings	0.85	T2	0.80
ABF Substrate Monopoly (Ajinomoto)	Supply bottleneck	Ajinomoto Co., Inc.	0.90	T1	0.95
Gallium/Germanium Restriction (Nov 2026)	Geopolitical	Coherent, Skyworks Solutions, Qorvo, Wolfspeed, Infineon Technologies	0.70	T1	0.65
CoWoS 2026 Allocation Constraint	Supply bottleneck	TSMC	0.85	T1	0.85
Japan Semiconductor-Input Seismic Shock (Korea coupling)	Natural disaster	Shin-Etsu, SUMCO, Tokyo Electron, Advantest, Disco, Lasertec, Tokyo Ohka Kogyo, Resonac Holdings, Kioxia	0.70	T4	0.82
Allied Coordination: TSMC Advanced Node Restriction	Policy intervention	TSMC	1.00	T4	0.90
EDA License Restriction: China Entity List Expansion	Policy intervention	Synopsys, Cadence	0.95	T4	0.80
Netherlands DUV Export Restriction to China	Policy intervention	ASML	0.70	T4	0.85
China Critical Minerals Retaliatory Restriction	Policy intervention	Coherent, Skyworks Solutions, Qorvo, Wolfspeed, Infineon Technologies	0.70	T4	0.65
Multilateral Cloud Compute Access Restriction	Policy intervention	Amazon.com, Inc., Microsoft Corporation, Alphabet, Oracle Corporation	0.80	T4	0.75
Netherlands EUV Restriction: Allied Countries Only	Policy intervention	ASML	1.00	T4	0.95

Related research

Epoch’s analysis of the CoWoS packaging and HBM memory chokepoint documents the concentration at both ends of the supply relationship: NVIDIA, Google, AMD and Amazon on the demand side; TSMC, SK Hynix, Samsung and Micron on the supply side. Relationships of that kind are the edges the simulator above propagates over.