Compute Visibility Index — Capacity-Weighted Ownership Transparency by Country

Gringras, David

Compute Leverage

Corrected Aug 6, 2026Fresh22d· methodology version

Who controls the compute a frontier training run needs — read at two grains. Across the eleven layers of the stack, whether a single actor can veto a run and whether anything checks that power; across the facilities and countries that host it, how much capacity sits under whose jurisdiction and where allied export control actually binds. Leverage is switching cost times redeployment time throughout, never market share.

Compute under whose control

How much compute sits under whose control, by country: strict jurisdictional attribution, capacity-weighted ownership-transparency tiers, and where the box physically sits versus who operates it. The per-layer supply-chain reading across countries lives in the Country radar.

Strict jurisdictional control: capacity at facilities where operator country, ultimate-parent country, and host country all equal C — with bifurcated and opaque residuals reported as named line items, not smoothed into the headline number.

Global summary

Strict jurisdictional control across the hosted-compute substrate

For each facility, the classifier resolves operator country (via operator_org_id → organizations.country_hq) and ultimate parent country (via the deepest ancestor in mv_ownership_chain_closure). A facility is domestic to country C only when host country, operator country, and ultimate-parent country all equal C. Three residual classes capture the ways the strict criterion fails: jurisdictionally bifurcated (operator country ≠ ultimate parent country), foreign-concentrated (operator and ultimate parent agree on a non-host jurisdiction), and opaque (ultimate-parent country unresolved). Capacity-weighted by power_capacity_mw for the denominator and per-class numerators.

A country’s compute is “under its control” only when three things agree: the facility sits physically inside its borders, the company operating the facility is incorporated there, and the ultimate corporate parent of that operator is incorporated there too. Anything less — a Microsoft Azure region in Frankfurt, a 21Vianet-operated Azure region in Beijing, an opaque holding company — falls into one of three named residual buckets and is rendered explicitly rather than smoothed over.

Hosted compute (global)

71.3 GW

41 countries with hosted compute

Under strict control

25.3 GW

35% of global hosted

Bifurcated

180 MW

0% — op ≠ ultimate parent

Opaque

32.1 GW

45% — ultimate-parent jurisdiction unresolved

How this is derived

Strict-control criterion. Operator country, ultimate-parent country, AND host country all equal C. Anything weaker classifies explicitly: bifurcated (operator and ultimate parent disagree), foreign-concentrated (operator and ultimate parent agree on a non-host jurisdiction), opaque (ultimate parent country unresolved), or data gap (no operator_org_id linked). The residual classes are not smoothed into the under-control number — each column in the table below is its own line so a jurisdictionally-bifurcated attribution is visible to a reader instead of being absorbed.

Top countries by under-control capacity

Compute Under Control by country

Capacity at facilities where the operator AND the ultimate parent of the operator are both domiciled in this country AND the facility is physically hosted here. Drill into any row for the facility-level attribution: each facility shows operator country and ultimate-parent country side-by-side so jurisdictional ambiguity reads from the data.

#	Country	Under control	% of hosted	Bifurcated	Opaque	Foreign-conc.	Off-shore controlled
1	United States1,490 fac	24.0 GW	45%	—	21.9 GW	340 MW	—
2	India136 fac	1.0 GW	64%	—	379 MW	181 MW	—
3	China57 fac	242 MW	38%	—	390 MW	—	—
4	Taiwan13 fac	37 MW	90%	—	3 MW	<1 MW	—
5	Japan121 fac	20 MW	3%	—	489 MW	254 MW	—
6	Saudi Arabia20 fac	14 MW	<1%	—	2.0 GW	—	—
7	South Korea40 fac	5 MW	2%	—	220 MW	64 MW	—
8	France309 fac	2 MW	<1%	—	1.4 GW	—	—
9	United Arab Emirates23 fac	—	—	—	163 MW	5.0 GW	—
10	Albania2 fac	—	—	—	—	—	—
11	Armenia2 fac	—	—	—	—	—	—
12	Angola1 fac	—	—	—	—	—	—
13	Argentina47 fac	—	—	—	—	10 MW	—
14	Austria22 fac	—	—	—	—	—	—
15	Australia123 fac	—	—	—	100 MW	55 MW	—
16	Bangladesh3 fac	—	—	—	—	—	—
17	Belgium39 fac	—	—	—	—	—	—
18	Bulgaria14 fac	—	—	—	—	—	—
19	Bahrain2 fac	—	—	—	—	—	—
20	Brazil113 fac	—	—	—	761 MW	—	—
21	Bhutan1 fac	—	—	—	—	—	—
22	Belarus1 fac	—	—	—	—	—	—
23	Canada111 fac	—	—	—	5 MW	52 MW	—
24	Switzerland47 fac	—	—	—	80 MW	—	—
25	Côte d’Ivoire2 fac	—	—	—	—	—	—
26	Chile31 fac	—	—	—	156 MW	—	—
27	Colombia13 fac	—	—	—	—	—	—
28	Costa Rica5 fac	—	—	—	—	—	—
29	Cuba1 fac	—	—	—	—	—	—
30	Cape Verde1 fac	—	—	—	—	—	—

Reference

Methodology — Compute Under Control

The Compute Under Control metric answers a narrow question that ownership-transparency measurements cannot: of the compute physically hosted inside country C, what fraction is operated by an entity incorporated in C whose ultimate corporate parent is also incorporated in C? The metric is deliberately strict. Where the operator-and-parent agreement breaks — the typical hyperscaler-region pattern, the joint-venture pattern, the holding-company-in-Cayman pattern — the residual capacity is reported under named residual classes rather than treated as a continuous gradient.

Per-facility classification

Domestic. Host country = operator country = ultimate parent country. The strict in-C-control case. A 1 GW facility in the US operated by a US-incorporated firm whose ultimate parent is also US-listed classifies here.
Foreign-concentrated. Operator and ultimate parent agree, but on a different jurisdiction than the host. The textbook Microsoft-Azure-Frankfurt pattern: operator (Microsoft Deutschland) and ultimate parent (Microsoft Corporation, Redmond) both US-aligned, but the facility sits on German soil.
Bifurcated. Operator country and ultimate parent country disagree. The 21Vianet-Azure-China pattern: operator is 21Vianet (China-incorporated), ultimate parent is Microsoft (US). The facility appears in C’s drill-down twice, once as host-country jurisdictional axis and once as ultimate-parent jurisdictional axis; the “Under control” column treats it as neither’s strict in-control facility.
Opaque. Ultimate parent country cannot be resolved. Either the chain has no documented ancestors with a populated country_hq, or operator country itself is null. These are not absorbed into any country’s under-control numerator; they are reported as a named residual class.
Data gap. Facility has no operator_org_id, owner_org_id, or hardware_owner_org_id linked. Investigation pending. Reported separately from opaque so coverage gaps do not contaminate the honest-uncertainty signal.

Ultimate-parent resolver

For each operator the classifier reads from mv_ownership_chain_closure — the transitive closure of ownership_upward_edges bounded at six hops — and picks the deepest reachable ancestor as the ultimate parent. Ties at the same hop depth are resolved alphabetically (deterministic). When no documented ancestor exists, the operator itself serves as its own ultimate parent (correct: the operator is the topmost known node and further ancestors are simply not recorded in any source Scrutica ingests). The picker is conservative: it does not infer beneficial ownership from indirect signals; if the chain stops, the chain stops.

Definitional caveats

Where the definition’s choice matters. A Cayman-domiciled holding company is treated as Cayman-domiciled by this classifier, not as the jurisdiction of its top operating subsidiary. A reader who prefers the “substantive economic ownership” reading should read the drill-down and form the alternate judgement explicitly; the classifier never silently picks a winner.
Multinational holding parents. When the deepest ancestor is a multinational holding (e.g., a Luxembourg SPV) with no clear majority jurisdiction, the classifier reports it as the SPV’s country of incorporation. This is a Scrutica editorial choice and is documented explicitly so a reader who reads “Luxembourg-controlled” on a row knows the classifier is reporting the incorporation jurisdiction rather than the controlling shareholder’s.
Substrate sparsity in non-disclosure jurisdictions. The same caveat that bounds the Ownership Transparency tab applies here. Countries with limited public-disclosure obligations have structurally sparse coverage. China, Saudi Arabia, India: the under-control column reflects Scrutica’s facility coverage of those countries, never deployed reality.
Capacity weighting. Aggregates are weighted by power_capacity_mw, so a 1 GW facility carries more weight than fifty 1 MW facilities. Where capacity is null, the row contributes zero to the aggregate.
Coverage caveat. The metric is conservative by construction. A reader looking at a low under-control percentage for a country should read it as the floor (“at least this much is strict-control”), not the ceiling: bifurcated and opaque residuals may be domestically controlled under a more permissive reading, and the table separates them so a reader can apply that reading inline.

Relationship to the Ownership Transparency tab

The two metrics are complementary, not redundant. Ownership Transparency measures how legibly the chain traces to a beneficial owner through public records (a transparency-of-records dimension). Compute Under Control measures whether operator and ultimate parent agree on a single host-country jurisdiction (a jurisdictional-attribution dimension). A country can score high on one and low on the other: Ireland scores high on transparency (its hyperscaler operators are SEC filers) but low on under-control (those operators are foreign-headquartered). Saudi Arabia scores low on both for a different reason: substrate coverage of its operators is structurally sparse, and where coverage exists the chains terminate at the sovereign-investment vehicle level. Read both tabs together.

Peer attribution

Prior work on jurisdictional compute attribution

Public country-level compute attribution is sparse. Hawkins, Lehdonvirta & Wu (SSRN, June 2025, “AI Compute Sovereignty”) offer a three-tier framework across 225 cloud regions that asks a similar question through a different lens (single-jurisdiction-pick rather than explicit bifurcation). Epoch AI’s April 2026 hyperscaler-concentration post estimates that five US-headquartered firms hold 71% of cumulative AI compute as of Q4 2025 (8-point rise over 21 months), but the question is company-level rather than country-level. CNAS’s Sovereign AI Index attributes program-level capacity to sovereign jurisdictions. Compute Under Control works at a different grain: every row carries operator country and ultimate-parent country side-by-side, every residual class is named, and the drill-down is at facility granularity. The methodology is a Scrutica editorial choice; a reader who prefers a different operational definition of “control” can apply it inline from the per-facility attribution.

Threshold Atlas

Coverage limitations of threshold tracking — capacity whose ownership walk reaches no documented ancestor

Sovereign AI

Sovereign program transparency vs. private-sector opacityCompute

Countries

Country-level hosted vs. controlled attribution

Supply Chain

Corporate-graph evidence behind the ownership-transparency tier classificationConstraint

Entities

The facility-by-facility footprint behind the hosted-vs-controlled splitControl

Built for the AI governance research community