Compute Visibility Index — Capacity-Weighted Ownership Transparency by Country

Gringras, David

Scrutica

Method 2026-07-22

Compute Leverage

Leverage over a frontier training run, measured as switching cost times redeployment time.

73.0% of tracked AI compute is not attributable to a controlling jurisdiction

Attribution by country. Every country hosting tracked AI compute, drawn as a track of its own hosted capacity and split by how far the ownership chain behind that capacity was actually documented. Ink style is the only encoding of certainty: solid where the record names a controller inside the host country, outlined where it names one somewhere else, dotted where the chain was walked and stopped before reaching any jurisdiction, blank where no operator, owner or hardware owner is linked at all. Every track is the same width and reads as a share, because the question is what proportion of a country’s hosted compute is attributable; the absolute capacity is printed beside each row so the weight is never inferred from the proportion. The four zones are a partition, checked against every country’s hosted total and closing exactly. The last two are held apart on purpose: an opaque chain is a fact about the public record, an unlinked facility is a fact about this platform’s coverage, and a figure that merged them would report the second as the first. 25 countries below the sixteen drawn are aggregated into the final row, so the denominator is inside the figure. What the field cannot show: 390 tracked facilities across 68 further countries state no capacity at all, so they have no width on a capacity-weighted field and appear in none of these figures; the largest by site count are Belgium, New Zealand, Turkey. Snapshot 2026-09-04.

69.6 GW of AI compute capacity is hosted in facilities Scrutica tracks across 41 countries. For 18.8 GW of it (27.0%) the documented ownership chain reaches a jurisdiction that can be named, and only 25.0% of the total is under the control of the country it physically sits in. The rest of the attributable share answers to a controller somewhere else, which is the ordinary shape of a hyperscaler region. The remainder divides in two: for 43.2 GW (62.0%) an operator is on record and the public record itself does not answer the question, and for a further 7.6 GW this platform has not yet reached an answer. Neither figure is a country ranking, and a ranking built on these records would mostly rank where corporate registries happen to be legible.

Where the chain stops

Attribution fails in two ways. In 43.2 GW (62.0% of hosted capacity) an operator is on record and the walk up the ownership chain terminates at an entity whose jurisdiction is not documented anywhere the walk can reach. In 7.6 GW (11.0%) there is no operator, owner or hardware-owner record linked to the facility to start from. The first is the limit of the public record. The second is the limit of this platform’s coverage, and it is reported separately because a page that folded it into the first would be citing its own coverage gap as evidence about corporate secrecy.

A second instrument reads the same chains by how far they trace, not by where they land: the ownership-transparency tiers, from a beneficial owner traceable through public filings down to a linked organisation with no ancestors documented. It runs on a different population, 49.9 GW across 4,257 facilities, against the 69.6 GW across 4,050 above. The two sets of percentages are not comparable and are never combined here. On its own denominator it reads:

52.6%: 26.2 GW · Tier 1 — public-filing beneficial owner
32.1%: 16.0 GW · Tier 4 — linked org named, no ancestors documented
15.3%: 7.6 GW · Data Gap — no linked org, not yet investigated

Empty by capacityTier 2 (30 facilities) and Tier 3 (60 facilities) have facilities but no stated capacity, so a capacity-weighted tier reading is 3-valued in practice. They are named here rather than printed as zero rows, which would read as a measurement.

Whether the gap is closing

Across 4 snapshots between 2026-04-25 and 2026-04-29, the share of tracked capacity documented to a public filing has risen, from 39.6% to 43.8%. That is the question this index exists to answer, and a single dated snapshot cannot answer it.

How to read a move in these lines. A move is not a measurement of the world. Two things move them. The record can change (a plant built, an operator disclosed, a chain walked to a filing), and this platform’s own coverage can change, because a batch of newly-ingested facilities enters at the data gap and moves every share without anything happening anywhere. The two are separable, and the separator is the facility count in the same series: the corpus went from 4,529 facilities to 3,669 over the span, so a share that moved while it grew is at least partly a fact about this project rather than about the field. The data gap itself reads 12.4% at the first snapshot and 13.1% at the last, and it is the line to watch for exactly that reason.

4 snapshots from 2026-04-25 to 2026-04-29.

Capacity-weighted tier shares per snapshot, global: every country including the unrecognized bucket, on the same 4,257-facility population the tier register above states, not the attribution population the field at the top of this page draws. Bands are the five documentation tiers in order; the two that have no capacity are drawn at zero width and named in the note above rather than implied to be absent.

The register

Every one of the 41 countries hosting tracked capacity, in capacity order, with the four zones of the field above as megawatts. Percentages in the last column are the attributable share: the first two zones over the hosted total.
Host country	Hosted	Its own	Elsewhere	Chain opaque	No operator	Attributable
United States · 1,490 sites	51.0 GW	15.8 GW	340 MW	28.3 GW	6.5 GW	31.7%
United Arab Emirates · 23 sites	5.2 GW	0 MW	0 MW	5.2 GW	0 MW	0.0%
France · 309 sites	2.4 GW	2 MW	0 MW	1.4 GW	1.0 GW	0.1%
Saudi Arabia · 20 sites	2.0 GW	14 MW	0 MW	2.0 GW	0 MW	0.7%
India · 136 sites	1.6 GW	1.0 GW	181 MW	379 MW	0 MW	75.7%
Malaysia · 28 sites	1.5 GW	0 MW	180 MW	1.3 GW	0 MW	12.1%
Japan · 121 sites	763 MW	18 MW	254 MW	491 MW	0 MW	35.6%
Brazil · 113 sites	761 MW	0 MW	0 MW	761 MW	0 MW	0.0%
China · 57 sites	632 MW	411 MW	0 MW	221 MW	0 MW	65.0%
Italy · 79 sites	606 MW	0 MW	0 MW	606 MW	0 MW	0.0%
Germany · 226 sites	460 MW	0 MW	55 MW	404 MW	0 MW	12.1%
United Kingdom · 203 sites	382 MW	0 MW	13 MW	279 MW	90 MW	3.3%
Netherlands · 141 sites	350 MW	0 MW	0 MW	350 MW	0 MW	0.0%
Hong Kong · 35 sites	300 MW	0 MW	0 MW	300 MW	0 MW	0.0%
South Korea · 40 sites	290 MW	109 MW	64 MW	116 MW	0 MW	59.8%
Finland · 32 sites	234 MW	0 MW	86 MW	140 MW	7 MW	36.9%
Chile · 31 sites	156 MW	0 MW	0 MW	156 MW	0 MW	0.0%
Australia · 123 sites	155 MW	0 MW	55 MW	100 MW	0 MW	35.6%
Singapore · 61 sites	111 MW	0 MW	20 MW	91 MW	0 MW	18.1%
Indonesia · 87 sites	96 MW	0 MW	0 MW	96 MW	0 MW	0.0%
Spain · 46 sites	93 MW	0 MW	8 MW	85 MW	0 MW	8.6%
Switzerland · 47 sites	80 MW	0 MW	0 MW	80 MW	0 MW	0.0%
Ireland · 35 sites	74 MW	0 MW	0 MW	74 MW	0 MW	0.0%
Canada · 111 sites	60 MW	0 MW	52 MW	5 MW	3 MW	87.4%
Denmark · 17 sites	55 MW	0 MW	0 MW	55 MW	0 MW	0.0%
Sweden · 41 sites	46 MW	0 MW	0 MW	40 MW	6 MW	0.0%
Taiwan · 13 sites	41 MW	37 MW	1 MW	3 MW	0 MW	92.0%
Portugal · 13 sites	33 MW	0 MW	0 MW	33 MW	0 MW	0.0%
Israel · 9 sites	33 MW	0 MW	33 MW	0 MW	0 MW	100.0%
Norway · 20 sites	29 MW	0 MW	0 MW	29 MW	0 MW	0.0%
South Africa · 49 sites	20 MW	0 MW	0 MW	20 MW	0 MW	0.0%
Argentina · 47 sites	10 MW	0 MW	10 MW	0 MW	0 MW	100.0%
Thailand · 24 sites	9 MW	0 MW	0 MW	9 MW	0 MW	0.0%
Iceland · 25 sites	7 MW	0 MW	7 MW	0 MW	0 MW	95.2%
Russia · 61 sites	6 MW	0 MW	0 MW	6 MW	0 MW	0.0%
Mexico · 40 sites	5 MW	0 MW	0 MW	5 MW	0 MW	0.0%
Vietnam · 9 sites	4 MW	0 MW	0 MW	4 MW	0 MW	0.0%
Poland · 48 sites	2 MW	0 MW	0 MW	2 MW	0 MW	0.0%
Luxembourg · 12 sites	1 MW	0 MW	0 MW	1 MW	0 MW	0.0%
Czech Republic · 24 sites	1 MW	0 MW	0 MW	1 MW	0 MW	0.0%
Slovenia · 4 sites	0 MW	0 MW	0 MW	0 MW	0 MW	0.0%

Coverage, not deploymentThese are shares of the capacity Scrutica tracks. They are not shares of a country’s deployed compute. The facility corpus is biased toward jurisdictions with public-disclosure obligations, so a country with few tracked sites reads as a distribution across those few sites and nothing more. The bulk of deployed Chinese compute, in particular, does not appear in any source corpus this site ingests.

Related instruments

1 · The same chains, read for allied control

This page asks which jurisdiction a facility answers to. The Coordination view asks the narrower question the same chains can answer: whether compute sitting inside an allied export-control regime traces to a parent beyond it. It walks the identical ownership edges, so the two views cannot disagree about what the record says. They differ only in what they ask of it.

2 · Where the facilities are

The facility corpus behind every number here is drawn on the map, which has its own honesty apparatus about what each mark attests. Per-country facility detail, tier by tier, keeps serving at its own address under the visibility index. The full derivation.