Facility estimates, hardware specifications and model-training observations come from Epoch AI.
CSET supplies semiconductor supply-chain records, used alongside the licensed corporate databases.
RAND’s firm-network research is a methodological reference for the cascade analysis. Other entries provide context on AI capability, its economic effects, electricity demand and compute governance.
Data-center capacity, AI hardware and model training compute.
Epoch estimates data-center capacity from satellite imagery, permits and company disclosures. Scrutica uses its facility estimates and hardware specifications in capacity calculations; Epoch's model-training budgets supply the model observations on Capacity Thresholds.
Semiconductor supply chains and export controls.
CSET's Emerging Technology Observatory maps semiconductor production inputs and the firms that supply them. Its mid-2025 dataset draws on TechInsights CMRS, WSTS and SIA; country market shares are assigned by the headquarters of each firm's ultimate parent. Scrutica incorporates these supply relationships into its company network.
Firm-network systemic-risk research for semiconductor supply chains.
Welburn and colleagues use customer–supplier relationships to estimate firm-level input–output linkages and study how a shock at one company affects aggregate production. Their 2023 working paper addresses missing data through inference and examines how network structure concentrates systemic risk. It is a methodological reference for Scrutica's cascade analysis.
Annual statistics on AI research, performance and investment.
Stanford HAI's AI Index collects annual evidence on AI development and its economic and social effects. The 2026 report includes research infrastructure, technical performance, private-sector investment and national AI policy, with separate chapters on science and medicine.
Evaluations of AI agents on extended tasks.
METR measures the length of tasks an AI agent can complete at a specified success rate, using the time those tasks take human professionals. Its March 2025 study fits success rates against human task duration on software and reasoning tasks. The resulting time horizon depends on the task suite and human baselines.
Power-demand and grid-impact research for data centers.
EPRI's Powering Intelligence research examines data-center electricity demand and its implications for the grid. Its 2024 report defines power usage effectiveness as total data-center energy use divided by IT-equipment energy use and discusses how cooling requirements affect that ratio.
Compute governance and AI regulation.
GovAI studies how compute infrastructure can support AI oversight and enforcement. Its research on computing power examines the practical advantages of concentrated hardware supply chains alongside the risks of intrusive monitoring and greater concentration of power.
Methodology gives the assumptions and source definitions used in these analyses.
18,980 supply-chain records from licensed corporate databases and CSET ETO. Relationships identify suppliers and customers, with product or service descriptions where available. The cascade model uses the recorded criticality and supply share; missing values default to 5 out of 10 and 30%, respectively. Sole-source relationships have a minimum criticality of 9. These assumptions affect the simulated disruption.
36 programs, with separate amounts for announced funding, commitments and disbursements. Country records identify the disclosures behind these amounts and distinguish government funding from private investment included in a headline pledge. Facility and chip-allocation records provide evidence of implementation where available.
3,435 BIS Entity List designations, linked to their published sources. Company matching combines names with recorded ownership relationships. An ownership connection can identify a company for further review; the applicable restrictions depend on the rule and its effective date.
GPU rental prices divided by dense BF16 throughput, expressed in dollars per petaFLOP-day. Daily refreshes cover AWS, Azure and Oracle pricing; GCP, CoreWeave and Lambda prices are checked manually against published price pages. Each observation carries its retrieval date. Hardware throughput comes from Epoch AI's ML Hardware database; prices on this basis exclude the extra throughput available from structured sparsity.
Simulates disruption through supplier and customer relationships, with adjustable severity and propagation assumptions. Downstream effects depend on supply share and criticality; upstream effects use a customer's share of the supplier's weighted relationships. In historical checks, the ASML blanket-halt scenario did not reproduce the selective 2020 restrictions, and the CoWoS comparison used parameters calibrated after the event. A comparison labelled ‘Russia helium 2022’ used a Qatar–Hormuz helium scenario unrelated to the 2022 disruption of Ukraine’s neon supply.
Compares model training budgets and estimated facility throughput with published compute thresholds. Facility calculations use recorded hardware where available and convert power capacity to throughput otherwise. Training duration and utilization are assumptions; the threshold descriptions state the associated legal conditions and whether the instrument remains in force.
Attributes data-center capacity to its host country and to the jurisdiction identified through corporate ownership. A separate ownership-transparency assessment classifies facilities by the available sources for their corporate ownership. The two assessments use their own stated facility populations and capacity totals.
Examines facilities whose host country and corporate ownership connect different export-control jurisdictions. Ownership chains are assessed against a versioned classification of national regimes. A documented connection to a jurisdiction outside the coordinated group remains visible when a later part of the ownership chain is unknown.
Documents the formulas and source definitions behind the analysis, including default values and uncertainty. The capacity and cascade calculators allow researchers to change assumptions and inspect the resulting estimates.
Scrutica uses Epoch AI's estimates for frontier data centers. Epoch's methodology explains how satellite imagery, permits and company disclosures support those estimates; Scrutica does not conduct an independent satellite survey of each site.
Capacity Thresholds uses Epoch AI's model-training observations. Epoch maintains the underlying model catalog and its estimates of cumulative training compute.
Epoch's AI Companies dataset tracks revenue, funding, staffing, usage and compute spending. It labels observations Confident or Likely according to their source and specificity. Scrutica's company records focus on infrastructure and corporate relationships; the financial disclosures they include are not a comprehensive financial history.
Scrutica does not conduct capability evaluations. METR measures AI agents on tasks timed against human professionals; its published methods explain the task selection and the limits of the resulting time-horizon estimates.
Stanford HAI's annual AI Index covers AI development and its economic and social effects across the field, including research infrastructure, performance, investment and policy.
Scrutica uses public procurement disclosures and licensed corporate databases. Classified contracts are outside its coverage; public award records may also omit information needed to identify the facilities or equipment involved.
Facility records describe reported or planned power capacity and public grid-connection requests. These figures do not measure a site's electricity consumption in real time.
Scrutica. “Related Work.” 2026-09-08. https://scrutica.com/related-work
