Organizations whose measurement this platform reads, and the flagship work of each.
Seven organizations, and what this platform reads from each
Scrutica is a join layer over other people’s measurement. This page names the organizations that measurement comes from, the flagship publication each is read through, what Scrutica builds on top, and — the section that matters most for citing it correctly — the questions Scrutica does not answer and whose work to read instead.
The peer network
Seven organizations whose work Scrutica cites and builds on, covering frontier-compute measurement, export controls, firm-network systemic risk, AI capability and investment indices, frontier-capability evaluation, power-demand research and governance scholarship. Each entry names the flagship dataset or paper Scrutica reads from most often, and what it is read for.
The organizations whose work Scrutica builds on, across compute measurement, export controls, supply-chain risk, the AI Index, capability evaluation, grid impact and governance research. Each entry names the flagship publication Scrutica cites.
Epoch AI
Frontier compute measurement; the closest peer for AI compute infrastructure.
Epoch's Frontier Data Centers methodology anchors facility estimates on satellite imagery (SkyFi, Sentinel-2, Google Earth), permitting documents, and corporate disclosures, with per-record citations. Epoch's AI Chip Owners explorer, ML Hardware database, and Notable AI Models corpus feed several Scrutica computations directly (chip-vendor cross-validation, FLOP/s + TDP constants, the model side of the threshold-atlas).
Epoch AI publishes the most rigorous facility-level dataset in the field: 74 frontier data centers (as of the 2026-07-16 catalog pull), each verified against satellite imagery and permits with per-record citations. Scrutica ingests and attributes their work, and carries sovereign programs, supply-chain edges, and export-control designations alongside it.
Frontier Data Centers HubCC-BY 4.0; 74 sites with per-record citations as of 2026-07-16 (launched 2025-11-04 with 13 US sites).
CSET (Georgetown)
Export-control + chip supply-chain analysis.
The Center for Security and Emerging Technology's Emerging Technology Observatory ships the Advanced Semiconductor Supply Chain Dataset (May 2025), drawing on TechInsights CMRS, WSTS, and SIA, with per-input market shares and headquarters of ultimate parents. CSET ETO is one of the corpora Scrutica's cascade and concentration explorers join over (alongside licensed supply-chain databases, a licensed financial-fundamentals database, and chip-deployment chains — 18,999 edges total).
CSET's policy briefs and Chip Explorer dataset frame the export-control analysis Scrutica's BIS cross-reference and coordination-gap features build on. Scrutica's contribution is the cascade engine that propagates a designation through the supply graph and the gap analyzer that maps which allies can enforce coordinated controls.
RAND working paper WR-A2625-1 (Welburn et al., 2023) is the closest published methodology peer for Scrutica's cascade simulator: the same licensed bilateral-relationship database underpins both. RAND's semiconductor corpus (PE-A1394-1, RR-A2354-1, RR-4185, WR-A2625-1) frames supply-chain risk as firm-level input-output linkage estimation; Scrutica's cascade runs weighted BFS propagation over the shared edge substrate, with 3-month price-correlation weighting and sole-source criticality flooring.
RAND's Taiwan-strait scenario analyses and microelectronics policy work are the institutional reference for stress-testing the supply chain. Scrutica's Cascade Simulation operationalizes this as an interactive engine: a user selects a chokepoint and the simulation walks the failure forward through the supply graph, surfacing which downstream nodes survive on inventory buffers and which substitute under what timeline.
AI Index: the institutional reference for AI capability + investment metrics.
The Stanford Institute for Human-Centered AI's AI Index aggregates training-compute, capex, and capability-benchmark trends across the field, with explicit acknowledgment of Epoch AI as a primary data provider for the compute and capability sub-indices. The AI Index's per-country investment series is the reference series Scrutica's Sovereign AI dashboard cites for the capital side of program execution.
Stanford HAI's AI Index is the citation standard policymakers reach for when they need a stocktake on training compute, AI capabilities, and AI investment. Scrutica's Sovereign AI Reality Dashboard cites HAI's per-country investment numbers and reconciles them against announced-vs-deployed program execution.
Frontier-capability evaluation; the empirical bound on what compute thresholds mean.
METR's HCAST methodology ("Measuring AI Ability to Complete Long Tasks", Kwa et al., 19 March 2025) and the Time Horizons 1.1 dataset (released 29 January 2026) are the primary surface Scrutica's Autonomy Capability Threshold Monitor reads from: the monitor refreshes against METR's live time-horizons feed at metr.org/time-horizons/ (v1.1 entries through 2026-04-07; last refreshed 2026-05-09). The FLOP enrichment on Scrutica's autonomy-monitor surface joins METR's per-model time-horizons to Epoch AI's training-compute database, asking what training compute gates which 50%-time-horizon. The relationship is upstream-citation: Scrutica reads METR's published time-horizons at face value (authority tier 2) and adds the per-facility complement (which facilities have the training capacity to enter which time-horizon band).
METR is the independently funded nonprofit (no AI-company funding; supported by The Audacious Project, the Sijbrandij Foundation, and individual donors) whose autonomous-task evaluations bound what compute thresholds mean policy-wise. A 10²⁵ FLOP line (EU AI Act) or 10²⁶ FLOP line (rescinded US EO 14110; NY RAISE Act 10²⁶ + $500M revenue, effective 2027-01-01) is only policy-meaningful when paired with empirical capability bounds; the Threshold Atlas cites METR's time-horizon work (the frontier 50%-horizon doubling roughly every 7 months from 2019 onward) as that capability complement. METR partners with the UK AI Security Institute, the US AISI Consortium, OpenAI, and Anthropic for pre-release model evaluations.
Power-demand and grid-impact research for data centers.
The Electric Power Research Institute publishes facility-level power-demand methodology used by US utility planners and grid operators, including PUE conventions and load-curve work that informs Scrutica's Cost Index decomposition (chip TGP × interconnect × MFU × hours-utilized). EPRI's 2024 data-center-power study is the canonical reference Scrutica's PUE anchor table cross-checks against (default 1.15; range 1.05–1.45).
EPRI is the bridge between AI compute infrastructure and the electric grid. Their data-center load research informs Scrutica's Compute Cost Index and the energy-side framing on facility detail pages: what the grid has to deliver, where queues bind, where heat-recovery and behind-the-meter generation reshape the regional picture.
Scrutica's threshold-atlas and coordination-gap surfaces are the operational complement to GovAI's research on compute governance (training-compute thresholds, on-chip mechanisms, verification regimes, frontier-AI scaling curves): GovAI's methodological work frames the empirical question — which facilities can train above a 10²⁵ FLOP cumulative threshold, under whose jurisdiction, with what enforcement reach — and Scrutica operationalizes it at facility grain.
GovAI is the academic center studying the governance of advanced AI. Their work on compute thresholds (the EU AI Act 10²⁵ line, the China CAC line, on-chip verification) shapes the policy register Scrutica's Threshold Atlas and Coordination Gap analyzer translate into facility-level findings.
The substrate Scrutica builds and the engines that run on it. Each entry links to the live instrument; Methodology carries the full derivation chain behind every number in them.
What Scrutica builds and ships. Each entry links to the instrument itself; Methodology carries the derivation behind each.
18,999 weighted directed edges with relationship-rank-derived criticality (rank 1–3 → bucket 8; 4–10 → 6; 11–25 → 4; 26+ → 3) and 3-month price-correlation fields populated for 78% of licensed-database rows. Drives /cascade propagation, /supply-chain/chokepoints, /coordination-gaps reach analysis, and the regulatory-impact pre-scorer.
The supplier-customer relationships among the firms that build, operate, and sell AI compute. Drives the cascade simulation and the concentration analysis: what propagates when a chokepoint breaks, and where the supply chain narrows to single suppliers.
36 programs with announced-vs-committed-vs-disbursed columns, per-country derivation chain (facility / chip allocation / utility), inflation decomposition (government-only vs. private + FDI), authority-tier flags on every cited figure, and a public-records traceability tab walking licensed-database LP→fund→portfolio chains where available.
36 government AI programs, with the announced number, the committed number, and the disbursed number tracked separately. The page decomposes the gap between announcement and execution: France's €109B Action Summit pledge is 97% private capital; Saudi Arabia's program runs through PIF financing, with limited public disbursement disclosure; South Korea's ₩100T (~$74B) headline sits against about $1.5B in government-only AI spending (₩556B for the five-consortia model program + ₩1.6341T for NAICC compute).
3,435 Federal-Register-anchored designations + a closure walker that propagates structural exposure to compute-universe organizations via name-similarity matching and an affiliate-closure walk over the ownership graph. States the BIS Affiliates Rule suspension (effective Nov 10, 2025 through Nov 9, 2026): cascade signal is heightened-due-diligence trigger, not implied designation.
Every Entity List, MEU List, and SDN designation tracked, with cross-references against the 4,573-facility / 24,874-organization compute-relevant substrate (127,366 canonical-resolved organizations in the universe). When BIS designates an entity, Scrutica's cascade walks the affiliate chain and surfaces which compute-relevant organizations are structurally exposed.
Daily Vercel cron refreshes AWS/Azure/Oracle on-demand prices (vendor-canonical APIs); GCP/CoreWeave/Lambda refreshed via Playwright-verified manual ingest. Prices normalized to $/petaFLOP-day under the dense BF16 convention (no 2:4 structured sparsity; H100 SXM5 = 989.5 TFLOP/s) anchored to Epoch's ML Hardware database.
What it costs to train or serve at unit scale across the major cloud providers, broken down by hardware generation. The denominator is petaFLOP-days (one petaFLOP sustained for one day), so the comparison is normalized across H100, H200, B200, MI300X, GB200.
BFS forward + reverse + bidirectional propagation over the 18,999-edge graph. Per-scenario propagation decay (0.65–0.95) layered onto per-edge `flowShare × criticality/10` multiplier; six prebuilt scenarios (TSMC disruption, ASML EUV halt, US-China decouple, CoWoS bottleneck, HBM disruption, allied-only EUV) plus freeform mode. Back-tested against three historical disruptions: two reconciled with the observed impact ordering across the top five affected nodes (the 2020 ASML EUV restriction, and the 2023 CoWoS packaging bottleneck — the latter reported with the caveat that its packaging parameters were calibrated after the event, making it in-sample fit rather than out-of-sample prediction). The third did not test as posed: the disruption it named was not the one that occurred, and it is reported as not comparable rather than scored.
Pick a chokepoint (TSMC, ASML, SK Hynix, Nvidia) and watch the failure propagate through the supply graph. The simulation weights edges by licensed-database relationship rank and per-scenario decay; downstream nodes survive on inventory buffers (ASML EUV cliff at zero weeks; advanced wafers linear at 10 weeks; HBM cliff at 3 weeks) and recover via parameterized substitution paths.
Facility-level FLOP capacity vs. each regulatory threshold (EU AI Act 10²⁵ cumulative; China CAC ~10²⁴; rescinded EO 14110 10²⁶) with provenance click-through on every numeric cell. Path-A hardware estimate when GPU inventory is known; Path-B power-derived estimate (facility MW / PUE × GPU-fraction-of-IT / chip TGP) when only nameplate is known.
Which AI data centers around the world have enough compute to put a model across each regulatory line, under whose jurisdiction, with what enforcement reach. Pairs the model-side question Epoch's Notable Models data answers (cumulative training compute per model) with the facility-side question: which sites hold the physical capacity to cross each line.
Country-level deployment with hosted-vs-controlled attribution (capacity sitting on soil vs. capacity controlled by domestic-incorporated owners) and ownership-transparency tiering across the global footprint. Cross-validates against Epoch's company-level 9-entity chip-owners series (Google, Microsoft, Meta, AWS, China-aggregate, Oracle, CoreWeave, xAI, Other).
How much AI compute each country actually controls, distinct from how much sits on its territory: an estimate of 'compute under control' by jurisdiction, rolling Epoch's company-level series up into country-level totals where ownership chains resolve cleanly, with explicit ownership-opacity flags where they don't.
Per-allied-country reach over each chokepoint tier (lithography → wafer → packaging → memory → accelerator → cloud), with explicit flagging of where the allied legal authority for coordinated export controls breaks down (Korean MSIT semi-coordination via WA Multilateral but not fully aligned on Entity List; Singapore's narrower extraterritorial framework; Taiwan-specific gaps).
Which US allies have the legal authority to coordinate export controls on which chokepoints. Where the gaps fall (Singapore, Korea, Israel for some technologies), the page surfaces the structural reason and the policy lever that would close it.
Every numeric cell rendered through the ProvenanceValue infrastructure carries the full envelope on click: data_source, source_url, vintage, authority tier (T1 primary measurement → T4 editorial assessment), method, confidence, and any documented conflicts. Sliders on the FLOP engine and cascade page propagate parameter changes through the entire derivation chain.
Click any number on Scrutica and read where it came from. Every value carries its source URL, when the data was originally produced, the authority tier (primary measurement, research database, press, estimate), and the conflicts where sources disagree, surfaced inline rather than buried in a methodology footnote.
Outside Scrutica's scope
What Scrutica explicitly does not cover, and whose work to read instead. Knowing where the line falls is what keeps the data from being cited for a claim it cannot support. None of these are gaps queued for later coverage; they are scope decisions that hold for the foreseeable horizon.
What Scrutica explicitly does not do, and where to go instead. None of these are gaps queued for later coverage; the scope decisions hold for the foreseeable horizon.
Scrutica does not run independent satellite-imagery analysis or permitting-record investigation on individual facilities. Where we ingest Epoch's frontier data-center rows, each is attributed to them on the record and read at face value rather than re-investigated. The verification benchmark for the frontier facilities Epoch covers (74 sites at the 2026-07-16 catalog pull) is theirs.
AI Models corpus (cumulative training compute per model)
Scrutica does not catalog ML models. The model side of the threshold question (which trained system has crossed which compute line) is Epoch's Notable Models database; Scrutica's Threshold Atlas asks the facility-side complement and links into the Notable Models corpus where appropriate.
AI Companies revenue / funding / staffing tracking
Scrutica does not track AI-company revenue, headcount, or rolling funding totals. The AI Companies dataset Epoch publishes (revenue, funding, staff, usage signals with 'Confident' / 'Likely' tier per data point) is the citation source for company-level commercial trajectory; Scrutica's company pages surface compute-relevant signals (deployments, ownership chains, supply-chain edges) only.
AI capability benchmarking (METR HCAST, ARC-AGI, FrontierMath, etc.)
Scrutica does not run capability benchmarks or curate evaluation suites. METR's autonomous-task length curve, the FrontierMath benchmark, the ARC-AGI corpus: capability work belongs to the labs and benchmark organizations doing it. Scrutica's Capability Scaling explorer cites their findings rather than reproducing them.
Scrutica does not publish an annual AI Index. Stanford HAI's AI Index is the field's institutional reference for the broad question 'where is AI overall'; Scrutica covers the narrower compute-infrastructure layer.
Scrutica relies on public-records substrate (USAspending, EU TED, simpler.grants.gov, Federal Register, SEC EDGAR, vendor disclosures) plus licensed corporate-ownership and supply-chain databases held under subscription. Classified contracts (intelligence community, advanced military procurement) sit outside the public record by design and outside Scrutica's coverage by construction. Where the public record carries gaps that distort the picture, the methodology pages flag them explicitly.
Real-time grid-load measurement at the data-center bus
Scrutica tracks announced power-capacity nameplate (in MW) and queued grid-interconnection requests where filed publicly (ERCOT, MISO, PJM queues; published utility data). We do not run live grid-load telemetry against individual facilities; that work is utility-internal and EPRI-adjacent.
Scrutica. “Related Work: how Scrutica relates to the compute-governance research network.” 2026-08-14. https://scrutica.com/related-work
The date fixes the PROSE — the peer set and the section text — to a specific revision. The substrate counts quoted inside the entries render from the live snapshot and move independently of it, so include the date when citing and a reader can map the wording, not the numbers, to the version they read. The numbers themselves are citable from the citation guide.
