Method 2026-09-05

Compute Leverage

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

Where allied control binds

Per facility, whether compute inside an allied export-control regime is owned by a parent outside it.

2 corrections · latest Aug 6, 2026

STRONG: 104; chain reaches a non-aligned or restricted jurisdiction · 838,575 H100-eq GPUs
OPAQUE: 3,095; chain terminates before a public-filing or government endpoint · 11,509,585 H100-eq GPUs
MILD: 214; chain reaches allied-partial only · 2,547,619 H100-eq GPUs
NONE: 844; every documented jurisdiction allied-coordinated · 10,894,103 H100-eq GPUs

The 30 largest of 3,199 facilities classified STRONG (the chain reaches a non-aligned or restricted jurisdiction) or OPAQUE (it stops before any public-filing or government endpoint), ranked by H100-equivalent GPUs, STRONG first. Population: 4,257 public-source facilities have a resolved ownership-chain classification. Facility rows sourced from licensed databases are excluded from this analysis entirely. Regime snapshot v2026-04-25. The interactive analyzer (requires JavaScript) loads the complete set with severity/jurisdiction filters, search, and per-facility chain detail; the toolbar above exports the full set as CSV or JSON.
Facility	Located	Operator	Severity	Hops	H100-eq
DayOne Nusajaya Tech Park	MY	DayOne Data Centers Limited	STRONG	1	~281,455
Alibaba Zhangbei	CN	Alibaba	STRONG	1	~132,895
Huawei Horinger	CN	Huawei Technologies	STRONG	1	~123,799
VNET Bayin Ulanqab	CN	VNET Group ADR	STRONG	1	~122,789
G42 Microsoft 100 MW UAE Cluster	AE	G42	STRONG	2	~55,000
YTL AI Johor	MY	YTL Power	STRONG	2	~38,979
G42 Microsoft 30 MW UAE Cluster A	AE	G42	STRONG	2	~15,000
G42 Microsoft 30 MW UAE Cluster B	AE	G42	STRONG	2	~15,000
Foxconn Big Innovation Cloud AI factory	TW	Foxconn	STRONG	1	~13,050
Saudi Data & AI Authority Sovereign AI factory	SA	Saudi Data And Artificial Intelligence Authority	STRONG	1	~12,633
Foxconn Hon Hai Kaohsiung Supercomputer	TW	Foxconn	STRONG	1	~11,642
Aramco Groq Inference Cluster	SA	Saudi Aramco	STRONG	1	~7,475
KAUST Shaheen-III	SA	King Abdullah University of Science and Technology	STRONG	1	~2,816
Core42 SuperPOD	AE	G42	STRONG	2	~1,272
Yandex Chervonenkis	RU	Yandex	STRONG	1	~502
Saudi Aramco Dammam-7	SA	Saudi Aramco	STRONG	1	~500
Condor Galaxy 3 (CG-3)	US	G42	STRONG	2	~485
Saudi Aramco Tuwaiq-1	SA	Saudi Aramco	STRONG	1	~416
Core42 AI-03	AE	G42	STRONG	2	~395
Yandex Lyapunov	RU	Yandex	STRONG	1	~346
Yandex Galushkin	RU	Yandex	STRONG	1	~343
Condor Galaxy 1 (CG-1)	US	G42	STRONG	2	~243
Condor Galaxy 4 (CG-4)	—	G42	STRONG	2	~243
Condor Galaxy 5 (CG-5)	—	G42	STRONG	2	~243
Condor Galaxy 6 (CG-6)	—	G42	STRONG	2	~243
Condor Galaxy 7 (CG-7)	—	G42	STRONG	2	~243
Condor Galaxy 8 (CG-8)	—	G42	STRONG	2	~243
Condor Galaxy 9 (CG-9)	—	G42	STRONG	2	~243
G42 Artemis	AE	G42	STRONG	2	~82
Bridge Data Centres MY06 (Sedenak Tech Park)	MY	Bridge Data Centres (BDC)	STRONG	2	—

How the analyzer classifies each facility

Four deterministic steps, each reproducible from the facility records and the versioned regime snapshot linked below.

1. Chain walk

Per facility, the classifier starts at COALESCE(operator_org_id, owner_org_id, hardware_owner_org_id) and walks ownership_chain_edges upward, picking the highest-authority documented parent at each hop. Termination conditions: a publicly-listed entity (CIK present), a government entity, a public-filing-grade ancestor edge, a cycle, or 10 hops, whichever lands first. The walker is the same one the Facility Intelligence Card uses, so the card view and the analyzer never disagree about what the documented chain says.

2. Posture lookup

Each hop's country_hq normalizes to ISO 3166-1 alpha-2 and resolves through jurisdictional_regimes.json (v2026-04-25, built from the Wassenaar participant list, EU 2021/821, EU Delegated Regulation 2025/2189, national export-control authorities, and CSIS Allen & Goldston, March 2025). Hops whose country does not resolve surface as 'unrecognized'.

3. Severity rule

STRONG when the chain reaches a non-aligned or restricted jurisdiction. MILD when the chain reaches at least one allied-partial jurisdiction without reaching a non-aligned/restricted one. OPAQUE when the chain terminates short of a public-filing or government endpoint AND no non-aligned/restricted jurisdiction has been documented. NONE when every documented jurisdiction is allied-coordinated AND the chain terminates at a documented endpoint.

4. Conservative defaults

If a chain terminates opaquely AND a non-aligned hop is already documented, the row is STRONG (the documented signal carries). If a chain has no documented jurisdictions at all, the row is OPAQUE. Operators with no organization linked at all surface as a separate data-quality bucket inside OPAQUE so they can be filtered out without contaminating the gap analysis.

Severity reference

STRONGThe chain reaches a non-aligned or restricted jurisdiction, so compute sitting in an allied posture answers upward to a jurisdiction outside the coordinated framework.

OPAQUEThe chain terminates short of a public-filing or government endpoint.

MILDThe chain reaches at least one allied-partial jurisdiction: a Wassenaar member, or a state applying EU dual-use rules, with no AI-compute-specific instruments of its own.

NONEEvery documented jurisdiction in the operator-to-ultimate-parent chain runs AI-compute export controls aligned with the US-led trilateral framework.

Regime classifications are versioned; snapshots live under data/processed/jurisdictional_regimes_versions/ so a researcher citing a Scrutica snapshot can re-run the classification against the regime data as it stood, after the live file has moved on. Current snapshot: v2026-04-25 JSON snapshot.

The Center for a New American Security publishes a program-level Sovereign AI Index (Chavez, Chilukuri, Scanlon, April 2026), which readers working on jurisdictional control over AI compute will want alongside this page: cnas.org.

Regime coverage, scenarios, and the export-control timeline

Reference material for the analyzer above: which regimes cover which technologies, how the multilateral regimes are structured, named coordination scenarios, and the export-control timeline. The jurisdictions listed are supply-chain participants outside coordination, in no ranked order.