Compute Visibility

The classification starts at the organisation a source establishes as the facility’s operator, owner or AI-hardware owner and follows the recorded parent chain, stopping before any link reviewed as not establishing control. It reads the headquarters country recorded for that starting organisation and for its parent, and assigns host-country control when both match the facility’s country. Control here is corporate domicile: where the company and its parent are headquartered.

A name without source support for that role counts as no owner or operator established; an owner named only by an unreviewed licensed source is classified separately.

The ownership-transparency tiers classify the documentation attached to each company and its recorded parents (definitions below). They run on a different population: every public facility record, without the licence-held rows the attribution above includes as aggregates, 16.6 GW operating on 322 of 4,267 records, 86 MW of it on records whose country label does not resolve, which the attribution above (~16.6 GW on 306 facilities) cannot place; so the two sets of percentages do not add. On its own denominator:

83.7%

13.9 GW · Tier 1 — company has an SEC identifier or a parent linked through filing, registry or supply-chain data

9.9%

1.6 GW · Tier 4 — company identified; parent unrecorded

6.5%

1.1 GW · Data Gap — no company linked

Empty by capacity Tier 2 (30 facilities) and Tier 3 (60 facilities) have facilities but no stated capacity, so they contribute no MW to these shares.

Across 26 snapshots between 2026-09-24 and 2026-10-19, the operating-capacity share classified as Tier 1 has risen, from 73.4% to 83.7%. The 2026-09-23 snapshot is left out of the comparison: it counts Epoch’s records at IT power, the compared snapshots at facility power (the power-basis note below explains the step).

The snapshots contain 4,266 and 4,267 facilities respectively, so the share moves with the set of records held as well as with reported capacity and ownership. The data-gap share was 6.5% at both.

How a tier is decided. The classifier is the materialized view v_facility_visibility_tier. It starts at the facility’s linked organisation, COALESCE(operator_org_id, owner_org_id, hardware_owner_org_id), and reads mv_ownership_chain_closure, which holds every organisation reachable upward through the recorded parent links within six hops, along every path. It selects no parent and applies no confidence or source-priority threshold. It asks four questions in order: is no organisation linked (Data Gap); does the organisation hold a CIK, or does any link within those six hops come from a Tier 1 source (Tier 1); does any come from the licensed private-market database (Tier 2); is any link recorded at all (Tier 3). A facility that answers no to all four is Tier 4. Every recorded link counts, including links reviewed as not establishing control, because a tier measures documentation; control is the attribution register’s question, set out under Where the chain stops.

The chain a reader can open. A country’s page lists its facilities with a View chain control, which draws one ownership path from a separate routine, getOwnershipChain() in src/lib/data/ownership-chain.ts. At each hop it follows the parent whose link ranks highest by source, then by confidence, then by the most recent as-of date (pickBestParent()); a source outside its ranking goes last. It stops at the first parent that holds a CIK, is a government body or is reached through a filing-grade link; at a parent already on the path; where two different parents tie on source and confidence and their as-of dates do not separate them (detectContestedParents()), except that two claims from the same source and the same as-of date read as joint ownership recorded by one document, and the walk continues through pickBestParent(), while undated same-source ties stay contested; or at ten organisations, the starting one included. Each hop shows its link’s source, confidence and as-of date. The drawn path shows the records behind a tier without being the classifier’s evidence: it can follow a different parent from the link that decided the tier, and its list of filing-grade sources is longer than the Tier 1 list, so it can mark a link as a filing on a Tier 3 facility.

The two are kept apart because they describe different things: Tier 4 describes a linked organisation’s ownership records, Data Gap describes Scrutica’s coverage of a facility. The order of the classifier’s questions separates them. Take three facilities.

• The first links no operator, owner or AI-hardware owner. The first question settles it: Data Gap, with nothing further read.
• The second links an operator that holds a CIK and has no recorded parent. The second question settles it: Tier 1, because the operator’s own filings document it and no parent is needed.
• The third links an operator with no CIK and no recorded link above it. It answers no to every question and falls to Tier 4.

A missing parent therefore never makes a facility Data Gap, and never makes a filer Tier 4. The tier shares above list Data Gap separately from Tier 4 for the same reason: two countries’ Tier 4 shares compared without their Data Gap shares would treat their records and Scrutica’s coverage of them as one quantity.

A tier records whether a chain could be walked. It says nothing about why it could not, and treating the residual as opacity conflates states that are different objects: an organisation that told a public registry no entity consolidates it has answered the question; one that declined to name its parent has refused it; one with no LEI recorded in Scrutica’s registry data has no declaration here to read either way. Disclosure posture is a second axis over the same facilities, read from the reporting exceptions filed to the Global LEI system, where a registrant that reports no consolidating parent must declare a reason. It decomposes the residual and ranks nothing: no single ordinal could rank “no LEI recorded” against “declines to say”, because one is a fact about the registry data Scrutica holds and the other is a registrant’s stated choice. The Companies register’s Disclosure view reads each declaration against the facilities its organisation holds.

Four states, and the boundary that matters most runs between the second and the third. Documented: the chain is on the record, Tiers 1 to 3. Declared: no chain, and the registrant stated why; the method behind the absence is the registrant’s own filing, so it is citable. Absent: no chain and no statement, where nothing distinguishes “no parent exists” from “parent undisclosed”. No linked organisation: the Data Gap, a gap in Scrutica’s coverage, where the other three are findings about the records.

The declared classes preserve the registrant’s own reason code, and each says something different: NO_KNOWN_PERSON declares that no entity consolidates it, which ends the chain by statement where the other states end it by running out of records; NATURAL_PERSONS declares control by individuals, so no parent entity exists to record; NON_CONSOLIDATING declares a parent that does not consolidate it under the applicable accounting standard; NO_LEI declares that a parent exists and holds no identifier, which asserts the parent’s existence without naming it; and NON_PUBLIC, with the finer consent and legal-obstacle codes GLEIF records alongside it, declares that an obstacle prevents publishing a parent the registrant has. Only the last is a refusal, and it is the one class never merged with an absent record, because it alone asserts both that a parent exists and that it is being withheld.

Each non-documented class also names whose limitation it is, because the failure this axis exists to prevent is reading a registry’s reach as an organisation’s secrecy. Where an LEI is recorded for the organisation and no parent report or exception is linked to it in Scrutica’s registry data, the class is registered with no parent report linked, and the silence is attributed to the organisation as far as those records go. Where a parent report is linked but Scrutica cannot yet join the parent to a tracked entity, the gap is this platform’s entity resolution and is attributed here. Where no LEI is recorded for the organisation, the class is no LEI recorded: Scrutica’s records do not show whether the organisation is registered elsewhere or what reporting it owes, so the absence is attributed to the registry data as held. That class is the largest of the three, and reading it as concealment would be the central error of the instrument.

Two rules govern the edges. Where a registrant answered the direct-parent and ultimate-parent questions differently, a declared refusal takes precedence over every other declaration, because it gives the most information: a parent exists and is unnamed. A declaration that no parent exists ranks below every other named reason, because it is an answer and states no obstacle. And a declaration never moves a facility between tiers. An SEC filer can decline to name a parent at the LEI registry; its chain is still documented through the filing regime, so the tier holds, and the company’s record shows the declaration beside its filings.

• A declaration is self-reported. The registry records what the entity declared and does not attest that the declared reason is true. A declared apex is the registrant’s claim to be the top of its own chain, and this axis reports the claim with its permalink without endorsing it.
• Registry reach is uneven by geography, in a direction that matters here. Identifier adoption is driven substantially by EU and US financial regulation, so entities outside those regimes are systematically more likely to fall in the no-LEI-recorded class regardless of how they behave. That class is therefore not comparable across countries as a disclosure measure, and it compounds the Western skew noted under Limitations below.
• The exception’s reference field does not name the parent. Where a registrant declares an unregistered or unpublishable parent, the registry holds no recoverable identity for it: measured on 8 August 2026 across the full golden copy, the field is populated on well under one per cent of exceptions, and on none of those Scrutica holds. No parent-recovery path is built on it.
• Organisations are matched through Scrutica’s canonical identity map, never by name alone. Legal-form variants of one entity resolve together through that curated map; corporate form is never stripped as noise, because it is often what distinguishes a parent from its subsidiary.

• The population is the public-source facility set. The tiers are computed over facilities whose own records come from public or freely redistributable sources. Facility records sourced from licensed databases are outside the tier population entirely, which is why its totals differ from the attribution register at the top of this page, where licence-held records enter as aggregates. Licensed sources still contribute parent links inside ownership chains, per the tier definitions above.
• Licensed links are dated extracts. Parent links from the licensed databases are held under subscription and not redistributed. They come from extracts taken on stated dates, so a restructuring after an extract is absent from its links until another source records it. Each hop’s as-of date is shown in the chain a reader can open, so freshness can be judged link by link.
• Tier 1 source coverage skews Western. SEC filings, the LEI registry and the licensed databases between them cover most US-, EU- and Japan-listed parents and underweight private and state-controlled entities domiciled elsewhere. A country with a high Tier 4 share is not thereby less well governed: the same skew makes a high Tier 4 share equally consistent with a compute base that is more privately held and outside the reach of those sources, and the tiers cannot distinguish the two explanations.
• A facility with no operating power figure adds nothing to a share. Its capacity counts as zero megawatts; the facility counts on each country’s page do not depend on capacity disclosure.
• Country attribution uses host country. A facility physically in country X whose ultimate parent is domiciled in country Y counts under X in these tiers. Attribution to a controlling jurisdiction, which assigns host-country control only where the starting organisation and its deepest documented parent are both headquartered in the facility’s country, is the register at the top of this page, defined under Where the chain stops.
• Unresolved country labels are counted apart. A facility whose country label does not resolve to a country cannot enter any country’s share; where such facilities hold operating capacity, the tier paragraph above states it.

Scrutica’s identity adjudication of 14 August 2026 separates three legal entities that share the Inspur name, each read from its own primary documents. Inspur Group Co., Ltd., which is unlisted, was added to the Entity List with effect from 2 March 2023 (88 FR 13673). Inspur Electronic Information Industry Co., Ltd., listed in Shenzhen (000977) and also known as Inspur Information and as IEIT Systems, was added in its own right with effect from 25 March 2025, as a subsidiary of Inspur Group (90 FR 14046). Inspur Digital Enterprise Technology Limited, listed in Hong Kong (596) and formerly Inspur International Limited, was a Cayman Islands company as of the 26 May 2026 proxy form (HKEXnews), which also proposes a change of domicile whose completion Scrutica has not verified, and is named in neither notice; its 2025 annual report records Inspur Group, as Inspur Group Limited, holding 50.91% of it at 31 December 2025, directly and through a wholly-owned subsidiary.

The holder categories reported on the Shenzhen and Hong Kong issuers’ registers therefore describe the shareholders of two different companies. An earlier version of this case read them as two listings of one group; that reading was withdrawn on 14 August 2026.

For the tiers, the case shows what the classifier does not read. Its inputs are whether an organisation is linked, that organisation’s CIK and the sources of the parent links recorded above it; a register’s holder categories are not among them, so no categorisation can move a facility between tiers. Nor is a tier an export-control determination: designation status is read entity by entity from Federal Register notices such as the two above, and a notice that names a subsidiary, as 90 FR 14046 does, counts only as one of the sources that can document a parent link.

• A national-transparency ranking. The index measures how legible the ownership records of compute assets are. It does not score a country’s regulatory regime, its corporate-disclosure law or its export-control posture; a country with mostly Tier 1 capacity has more legible compute records, a narrower claim than being more transparent overall.
• An enforcement tool. Tier 4 capacity is no allegation of wrongdoing. Private holding structures, family-trust ownership and tax-domicile choices are common and lawful. The classifier records which ownership links are documented; what to do with that record is a governance question this page does not answer.
• A substitute for case-by-case investigation. A Tier 3 or Tier 4 facility may be entirely legitimate, and a Tier 1 facility may have governance issues its public chain does not show. The index is one structural indicator; each facility’s own page shows its findings, and the country register above is the aggregate they sit under.