Evidence for compute governance
AI infrastructure records for research on ownership, supply dependencies and compute policy.
Linking what is published separately
A facility’s power capacity says little about the training it can support until its hardware and utilization are known, and whether a regulation reaches it depends on the companies behind it and the jurisdictions they operate in. These facts are published separately, in facility catalogs, company filings, the Federal Register and government program announcements. Scrutica links them record by record, so a facility can be followed to its owners, their suppliers and customers, and any Entity List designation that reaches them.
Coverage
Sources include Epoch AI’s facility and GPU-cluster datasets, the Federal Register, SEC EDGAR, CSET ETO and licensed corporate databases. The methodology page describes their coverage and the gaps that affect the analysis.
Records and tools
- FacilitiesData centers and fabrication plants across 110 countries, with location, operator, power capacity and GPU deployment where available.
- Supply-chain relationships11,858 distinct supplier–customer relationships, counted separately by product or service. Of these, 1,924 come from a licensed layer limited to relationships between organizations already in Scrutica's records. Corporate ownership is recorded separately.
- Sovereign AI programs36 government compute programs, with announced funding, commitments and disbursements recorded separately where disclosed.
- Export controls3,435 BIS Entity List designations, 41 active designation-to-company matches. A designation may match more than one company record.
- Compute capacity estimatesDaily compute estimated from a chip count, a power capacity or a construction budget, with adjustable assumptions and uncertainty bounds.
The Query Builder answers questions over these layers in natural language, using 55 tools for records, calculations and page content. The MCP server exposes 10 read-only tools for use in an agent client.
Citing a figure
Individual pages and Query Builder answers can use a different date or population from these totals. Cite the figure you used, with its date and scope.
| Quantity | Value | Coverage | Vintage |
|---|---|---|---|
| Facility records | 4,583 | all recorded facilities, across classes and statuses; includes records from licensed sources that are not publicly readable | 2026-10-22 |
| Listed facility records | 4,267 | facility records a reader can open individually; excludes the records held under licence | 2026-10-22 |
| Countries | 110 | distinct country codes in facility records | 2026-10-22 |
| Supply-chain records | 20,798 | all supply-chain records, including licensed sources and repeated reports of the same relationship | 2026-10-22 |
| Distinct supplier–customer relationships | 11,858 | distinct supplier–customer relationships as the cascade engine counts them: company identities reconciled, keyed on the published product or service, so records that differ only by a withheld label count once | 2026-10-22 |
| Export-control designations | 3,435 | recorded export-control designations, with Federal Register sources | 2026-10-22 |
| Designation-to-company matches | 41 | active matches between designations and company records; a designation can have several matches | 2026-10-22 |
| Sovereign programs | 36 | national or regional government compute programs in the curated list | 2026-10-22 |
| Public query tools | 55 | tools available to the public Query Builder; excludes administrative tools | current tool list |
Limitations
Coverage depends on disclosure. Public-company filings can identify subsidiaries and ownership relationships that remain unknown for private or state-owned operators; missing reports also limit facility-capacity estimates.
Power and capital cost support capacity estimates only through assumptions about the hardware they serve. Hardware inventories reduce that uncertainty, but utilization and training duration still affect estimates of how much compute a facility can supply.
A reported supply relationship rarely establishes how much production would be lost if it failed. The cascade model uses default supply shares where values are missing and analyst-selected propagation rates; estimated severity depends on those assumptions.
Built by
David Gringras, physician (Edinburgh) and law graduate (University of Law), Frank Knox Fellow at Harvard (MPH Health Policy, 2026; cross-registered at MIT, Harvard Law School, and the Kennedy School). Expert panelist and co-author on the Evals-Consensus.AI Delphi study of AI evaluation practice; supported by a BlueDot Impact career transition grant. Previously Evaluations and Collaborations Lead on the FATF-to-AI governance translation project at Arcadia Impact / The Future Society, and project supervisor at Orion AI Governance on evaluation-independent governance mechanisms.
Send corrections or methodological criticism to the address above. Published corrections are listed at Corrections.