# Scrutica — Reference for AI Agents

*Coverage counts come from the snapshot dated 2026-10-03.*

Scrutica brings together records on AI computing facilities, semiconductor supply relationships, export controls and national AI programs. This guide explains the calculations and provides selected reference records. The [short index](https://scrutica.com/llms.txt) lists current pages; use the linked record pages and their sources for a particular finding.

## Access and reuse

Public pages, published dataset downloads and public read APIs are free to access without an account or API key. The [agent manual](https://scrutica.com/agents) documents MCP connections and request limits; the [OpenAPI specification](https://scrutica.com/api/openapi.json) describes public JSON endpoints.

Scrutica content uses CC BY-SA 4.0, subject to attributed sources' terms. Published dataset releases use CC BY 4.0 as stated on their download pages. Original subscription datasets are not redistributed. Follow the licence attached to the material you use.

## Coverage snapshot

These figures describe the collections counted on 2026-10-03. Public tools can return a smaller selection because their source permissions, filters and limits differ.

- 4,582 facility records across 110 countries (4,073 operating · 493 announced or under construction · 16 decommissioned or cancelled)
- 20,798 stored supply-chain records (designation matches excluded; partnership records included, though the cascade never propagates along one); 11,858 distinct supplier–customer relationships as the cascade engine counts them
- 127,034 canonical organizations; 7,420 linked to compute-related records; 5,376 appearing as suppliers or customers
- 3,435 export-control designation records
- 36 sovereign AI programs
- 3,125 investment records in the public collection, from 28,752 tracked. A licensed record gives an amount band and a year, never the amount, the date or the investor. 2,382 public and 20,000 tracked records come from two partial extracts (10,000 rows each, March 2026); they are labelled and kept out of every total and ranking
- 4,109 compute-deployment records

Facility coverage combines geographic registries, cloud-provider disclosures, research datasets and individual project sources. A facility record can describe an operating site, a proposed project or a manufacturing plant; the type and status determine whether it belongs in a capacity comparison. Registry presence alone supplies no GPU inventory.

Supply-chain totals include records with unnamed counterparties and repeated source reports. The deduplicated figure counts relationships as the cascade engine does: company identities reconciled, keyed on the published product or service, unreviewed associations left out. Of the deduplicated relationships, 1,924 come from a licensed layer limited to relationships between organizations already in Scrutica's records. Of the organizations appearing as suppliers or customers, 424 appear only through that layer. Of the organizations linked to compute-related records, 332 are linked only through that layer. Public relationship diagrams show a permitted selection. The cascade simulator also uses licensed relationships in its server calculation and returns derived outcomes.


## Citing a finding

Cite the specific Scrutica page and its access date, together with the underlying source for the reported fact. For an estimate, include the input values, their source dates, the selected hardware and the methodology version. A facility's daily compute estimate also needs its utilization and interconnect assumptions.

The [dataset citation page](https://scrutica.com/datasets/cite) supplies citations for published releases. Preserve the dataset version and CSV checksum when a calculation depends on a downloaded release; a documentation revision can occur after the data release.

A general platform citation can use:

```bibtex
@misc{scrutica2026,
  author = {Scrutica},
  title  = {AI Compute Infrastructure Intelligence Platform},
  year   = {2026},
  url    = {https://scrutica.com}
}
```

The [corrections register](https://scrutica.com/corrections) records published corrections. Check the relevant record and methodology when reproducing an older result.

---


## Pages by research question

**Facilities and ownership.** [Entities](https://scrutica.com/entities) provides facility, company and country registers. [The map](https://scrutica.com/map) locates tracked sites. Use [Capacity](https://scrutica.com/capacity) for facility evidence review and hypothetical compute scenarios, and [Training Evidence](https://scrutica.com/capacity/training-evidence) for company-level evidence about training infrastructure. [The Ownership Investigator](https://scrutica.com/entities/investigator) follows recorded corporate relationships; [Compute Visibility](https://scrutica.com/compute-leverage/visibility) summarizes corporate traceability through organizations linked to facilities.

**Dependencies and disruption.** [Supply Chain](https://scrutica.com/supply-chain) shows public supplier–customer relationships. [Chokepoints](https://scrutica.com/supply-chain/chokepoints) and [Concentration](https://scrutica.com/supply-chain/concentration) examine concentrated supply. [Cascade](https://scrutica.com/cascade) propagates a selected disruption through the graph. [Compute Leverage](https://scrutica.com/compute-leverage) evaluates suppliers and institutions that could block a proposed training run. [Scenarios](https://scrutica.com/scenarios) collects the regional analyses described below.

**Export controls and national programs.** [Export Controls](https://scrutica.com/export-controls) provides designation records and a [change log](https://scrutica.com/export-controls/changes). Its [impact](https://scrutica.com/export-controls/impact), [trade](https://scrutica.com/export-controls/trade) and [diversion](https://scrutica.com/export-controls/diversion) views address different questions about restrictions and supply relationships. [Threshold Atlas](https://scrutica.com/capacity/thresholds) estimates time to selected cumulative compute totals. [Sovereign AI](https://scrutica.com/sovereign-ai) compares program finances and hardware evidence. [Country Radar](https://scrutica.com/entities/countries/radar) summarizes national supply-chain positions; [Coordination](https://scrutica.com/compute-leverage/coordination) examines cross-country ownership relationships against policy groupings.

**Costs.** The [Compute Cost Index](https://scrutica.com/economics/cost-index) normalizes cloud prices by peak accelerator capacity and lists the [published training-cost figures](https://scrutica.com/economics/cost-index#training-cost) on record for frontier runs, each with the party that published it.

**Methods and reference.** [Methodology](https://scrutica.com/methodology) explains the calculations. The [glossary](https://scrutica.com/methodology/glossary) defines 77 terms. [About](https://scrutica.com/about) describes coverage and sources; [Related Work](https://scrutica.com/related-work) distinguishes imported datasets from methodological references and adjacent research. [Corrections](https://scrutica.com/corrections) records published amendments, while [the Record](https://scrutica.com/changelog) lists additions and updates to data collections.

**Queries and downloads.** [Query Builder](https://scrutica.com/query) accepts research questions in natural language. [Datasets](https://scrutica.com/datasets) lists published releases and their documentation. The [agent manual](https://scrutica.com/agents) covers MCP and public APIs; [Privacy](https://scrutica.com/privacy) describes submitted information and service data collection.

---


## Reading evidence and missing values

A source can establish one field while leaving another unknown. A planning application may establish a campus power envelope; a hardware-owner disclosure may identify who owns the accelerators. Neither establishes who operates every building on the site. Read the field's definition and source before combining records.

A numeric `null` means the response has no value for that field. Preserve it as missing in calculations. A reported zero is a different observation. The `is_estimated` flag identifies estimates where the response supplies it; a source classification and an estimate flag answer different questions.

Source dates, record-update dates and scheduled completion dates also differ. An announcement remains an announcement until the record has evidence of operation. Capacity totals that combine statuses describe projects at different stages.

For derived values, retain the operands and the method. Facility estimates use parameter-sensitivity ranges; the MCP throughput tool uses fixed bounds around a point estimate. Neither range has an assigned statistical confidence level. The method sections below state which inputs vary.

Company and facility pages can include conflicting source reports, accounting qualifications and corrections. Read these alongside the value. A missing parent relationship, missing identifier or incomplete source search cannot establish that no parent exists.

---


## Methodology — facility compute estimates

The [Capacity estimator](https://scrutica.com/capacity) separates named-facility evidence from hypothetical compute scenarios. A named-facility estimate requires operational compute-site status and a complete inventory whose partitions reconcile to the facility total; each partition must be present capacity, be counted in accelerators and resolve to exactly one accelerator model with a dense FP16 rate that has passed evidence review. A partition that fails any of these is held with its reason, and one held partition withholds the facility figure; every partition and its verdict remain visible. Nothing substitutes a default model, a dominant chip or a partial inventory. Named power estimates remain withheld pending separate review of the power quantity, scope, date and source. A held partition does not establish that an inventory is wrong.

The hypothetical calculators accept user-selected hardware, power or capital inputs. They multiply peak accelerator throughput by interconnect efficiency and model FLOP utilization (MFU), then by 86,400 seconds for a day of continuous operation. The following formulas describe these scenarios:

| Input | Calculation before utilization and interconnect adjustments |
| --- | --- |
| Hardware inventory | Accelerator count × per-accelerator dense FP16 FLOP/s |
| Facility power | Facility MW × 1,000 ÷ PUE × accelerator share of IT load ÷ per-accelerator TDP in kW; multiply the resulting accelerator count by its peak FLOP/s |
| Capital expenditure | Total investment × accelerator share of capital expenditure ÷ accelerator purchase price; multiply the resulting accelerator count by its peak FLOP/s |

PUE is total facility power divided by IT power. The accelerator share then allocates IT load between accelerators and other equipment. The capital-expenditure share allocates investment between accelerators and the rest of the project. Both shares must fit the scope of the reported input: a campus power connection and a completed training cluster describe different assets.

An admitted named-facility estimate retains each inventory component's count and supported dense FP16 specification, then sums the component estimates. Unknown hardware is never replaced with H100. The hypothetical calculators use the hardware selected for the scenario. H100-equivalent capacity is a normalization that needs its own source method; interpreting it as an installed H100 inventory changes the claim.

**Sensitivity bounds.** Hardware estimates vary interconnect efficiency and MFU. Power estimates also vary PUE and the accelerator share of IT load. Cost estimates also vary the accelerator share of investment and purchase price. These are parameter-sensitivity ranges with no assigned confidence level. The paths share hardware and efficiency assumptions, so agreement can reflect shared inputs. The estimator warns when the largest positive daily estimate exceeds the smallest by more than three times.

Named-facility admission is implemented by `evaluateFacilityCompute` in `src/lib/compute/facility-compute-evidence.ts`, partition by partition, with the closed dense-FP16 rate table `DENSE_FP16_EVIDENCE_RECEIPT` in the same file. The generic calculations are `estimateFromHardware`, `estimateFromPower`, `estimateFromCost` and `checkPathDivergence` in `src/lib/compute/flop-estimator.ts`. The [methodology](https://scrutica.com/methodology) identifies the versioned parameters and their sources. Installed hardware and a theoretical component sum do not establish availability or throughput for one training job.

---


## Methodology — Compute Cost Index

The [Compute Cost Index](https://scrutica.com/economics/cost-index) compares cloud rental prices per petaFLOP-day of peak dense BF16 capacity. A petaFLOP-day is 10¹⁵ FLOP/s sustained for 86,400 seconds.

For an instance with hourly price `P`, accelerator count `N` and per-accelerator dense BF16 throughput `T` in TFLOP/s:

`Instance PFLOP/s = N × T ÷ 1,000`

`Cost per petaFLOP-day = P × 24 ÷ (N × T ÷ 1,000)`

`calculateCloudScuCost` in `src/lib/compute/cost-calculator.ts` applies this formula. Missing or non-positive operands leave the normalized cost unavailable. The cloud denominator uses peak throughput. Workload MFU, memory requirements and communication overhead affect delivered training output and need a workload-specific calculation.

Compare prices within the same procurement category: on-demand, spot, one-year reserved or three-year reserved. Each observation lists its provider, region and retrieval date. An old low price may have ceased to be available; reservation commitments and interruptibility also change what the buyer receives.

The on-premises calculator in the same view uses a delivered-compute denominator that includes MFU rather than the peak-throughput denominator above. Its cost assumptions belong with that model. The [downloadable cost dataset](https://scrutica.com/datasets/compute-cost-index) documents the fields, source dates and peak-throughput normalization used in the published release.

---


## Methodology — cascade simulation

The [cascade simulator](https://scrutica.com/cascade) runs on the server using recorded supply-chain relationships, including licensed sources, after organization resolution and supplier–customer–product deduplication. The public diagram shows a permitted subset of relationships. Simulation responses contain derived impacts and aggregate graph counts; the underlying licensed relationship attributes remain on the server.

**Forward propagation.** At each relationship the model multiplies the incoming impact by supply share, criticality divided by ten, and the scenario's propagation decay. Missing supply share defaults to 0.3; missing criticality defaults to five. A sole-source relationship receives a criticality floor of nine. These defaults are model inputs, not disclosures about a particular supplier.

The optional correlation blend maps a recorded price correlation `r` to `clamp(5 × (1 + r), 1, 10)` and combines it with criticality. The default blend weight is 0.5, and the sole-source floor is applied afterward. The model uses the supplied correlation without removing common market or sector exposure; co-movement can therefore affect the weight without demonstrating a causal supply dependency.

**Direction and stopping rules.** Forward propagation follows disruption from a supplier to its customers. Reverse propagation approximates a customer's contribution using its relationship weight relative to the supplier's other customer weights, capped at 0.95. This share is not a measured revenue fraction. Bidirectional propagation combines the two results by taking the larger impact for each organization.

The traversal permits six propagation steps, limits repeated visits and admits onward propagation above a 0.05 pre-decay impact threshold. Where several routes reach an organization, the result retains the highest impact encountered; it does not add the paths. These rules affect both the affected set and the reported severity.

Preset scenarios supply initial organizations, severity and decay values. Their assumptions need to accompany an exported result. A missing relationship limits the routes the simulator can find, and an assumed weight limits interpretation of the severity. Use the result to investigate documented dependencies; it supplies no loss estimate in dollars or calibrated probability of failure.

Implementation: `buildGraph` and `simulateDisruption` in `src/lib/compute/cascade-simulator.ts`. The request runs through `/api/cascade/simulate`, `cascade-server-engine.server.ts` and `cascade-full.server.ts`. The [methodology page](https://scrutica.com/methodology) provides the scenario assumptions.

---


## Methodology — supply-chain relationships

A supplier–customer relationship needs two identified organizations and the product or service that connects them. The same pair can have several relationships. Scrutica resolves organization aliases before deduplicating on the supplier, customer and product or service; it preserves distinct products between the same firms.

The collection also contains aggregate disclosures with an unnamed counterparty. These describe a supplier's customer group or another aggregate, so a stored-record count cannot be read as a count of named bilateral relationships. The coverage snapshot above gives both the stored count and the deduplicated relationship count.

Public graph responses expose a permitted selection of the collection. The supply-chain data module applies source eligibility and attribute rules before returning relationships. An internal licensed record can contribute research context without its original record or numerical attributes becoming publicly downloadable.

Read supply share, criticality, correlation and their dates separately. Missing attributes remain missing in the source record; the cascade simulator applies its own documented defaults when it builds the model. A price correlation describes co-movement over its source's observation window and is not independent proof of physical dependence.

SEC Exhibit 21 supplies subsidiary disclosures. A listed subsidiary can be held indirectly through other subsidiaries, so the exhibit alone may establish group membership without identifying the immediate parent. Ownership evidence belongs alongside the supply relationship when corporate structure matters; it does not substitute for evidence that a product was supplied.

The [Supply Chain explorer](https://scrutica.com/supply-chain) and `scrutica_get_supply_chain` provide public relationship queries. The published [dataset directory](https://scrutica.com/datasets) lists downloadable releases; it currently publishes the sovereign-program and compute-cost datasets. The full supply-chain collection has no public packaged release.

Implementation: canonical resolution, `deduplicateEdges` and public projection rules in `src/lib/data/supply-chain.ts`; source eligibility in `src/lib/data/supply-chain-public-attributes.server.ts`.

---


## Methodology — sovereign AI programs

The [Sovereign AI register](https://scrutica.com/sovereign-ai) distinguishes announced, committed and disbursed funding. Its financial figures are in USD. Hardware deliveries and operational capacity are separate observations: disbursement alone cannot establish that a computing facility is operating.

The execution ratio is:

`disbursed_usd ÷ (announced_govt_only_usd ?? announced_usd)`

The government-only announced figure is preferred. When it is missing, the calculation uses the headline announcement, which can include private investment or other components. A missing disbursement, missing denominator or zero denominator leaves the ratio unavailable. Summary responses return the denominator field and amount with the ratio so the calculation can be reproduced.

Portfolio comparisons also need a matched population. The government-only execution rate includes programs with both a government-only announcement and a disbursement figure. Programs missing one side are reported separately; dividing totals drawn from different program sets would change the question.

**Funding components.** The field historically named `inflation_decomposition` separates components of a headline announcement, including government funds, private investment, credit and state-enterprise funds. The implementation performs an accounting breakdown, with no consumer-price deflator or conversion to constant dollars. Components can differ in funding stage and period, and one amount can include another. Notes and flags identify known overlaps, unverified amounts and missing hardware figures.

**Hardware procurement.** Sovereign programme records include reported accelerator models, counts where available, procurement status and source notes. Scrutica does not publish a numerical estimate of export-control jurisdiction over those programmes.

**Schedule comparisons.** The schedule-adjusted rate divides disbursement by the announced denominator multiplied by the fraction of the program's stated timeline elapsed, capped at one. The linear-pace projection divides remaining funding by the average annual disbursement since the start year. Both depend on a reference year and a uniform spending-pace assumption; programs with staged appropriations can depart substantially from that path. Missing required amounts or dates leave the relevant calculation unavailable.

The register is a curated set with uneven source coverage. It has no enforced $100 million inclusion threshold. For a comparison, use the same funding stage and denominator, then inspect the component sources and dates for each program.

Implementation: `computeRealityRatio`, `sovereignSummaryRow`, `computeGovernmentOnlyExecution`, `computeScheduleAdjustedRate` and `computeTrajectoryForecast` in `src/lib/data/sovereign-programs.ts`. The [dataset dictionary](https://scrutica.com/datasets/sovereign-ai-index) defines the published financial fields.

---


## Methodology — Threshold Atlas

The [Threshold Atlas](https://scrutica.com/capacity/thresholds) divides a selected cumulative compute total by an admitted facility's theoretical daily FLOP budget. A named-facility estimate requires operational compute-site status and a complete inventory whose partitions reconcile to the facility total; each partition must be present capacity, be counted in accelerators and resolve to exactly one accelerator model with a dense FP16 rate that has passed evidence review. A partition that fails any of these is held with its reason, and one held partition withholds the facility figure; every partition and its verdict remain visible. Nothing substitutes a default model, a dominant chip or a partial inventory. Named power estimates remain withheld pending separate power-evidence review; they cannot substitute for a held partition.

**Regulatory scope.** Article 51(2) of the EU AI Act sets a presumption of high-impact capabilities when a general-purpose model's cumulative training compute exceeds 10²⁵ FLOP. Providers can submit arguments against systemic-risk classification under Article 52, and the Commission can designate models on other capability grounds. These provisions classify models. Applying them requires the provider and model context as well as the training-compute total. Sources: [Article 51](https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-51), [Article 52](https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-52).

China's 2023 Interim Measures apply to generative-AI services offered to the public in China; Article 17 requires security assessment and algorithm filing for services with public-opinion attributes or social-mobilization capacity. The cited measures specify no numeric training-FLOP trigger. The Atlas therefore gives this regime no days-to-threshold calculation. Source: [CAC Interim Measures, Articles 2 and 17](https://www.cac.gov.cn/2023-07/13/c_1690898327029107.htm).

The Atlas also retains historical and research reference lines with their status stated. EO 14110's US reporting threshold is historical: [EO 14148, Section 2(ggg)](https://www.federalregister.gov/documents/2025/01/28/2025-01901/initial-rescissions-of-harmful-executive-orders-and-actions) revoked that order on January 20, 2025.

**Reading the estimates.** Thirty-, ninety- and 365-day counts are cumulative: a facility within thirty days belongs to all three counts. The estimable denominator contains only admitted, reviewed inventories; excluded and withheld records remain separate. Missing review is not zero capacity. A duration assumes continuous access to the theoretical component aggregate. It establishes neither availability to one job nor evidence that a training run occurred.

The reverse calculator computes the accelerator count needed for a target budget and duration, or the duration needed for a specified count. The shared implementation is `facilityPrimaryEstimate` and `daysToThreshold` in `src/lib/capacity/threshold-reach.ts`, with `reverseFromThreshold` in `src/lib/compute/flop-estimator.ts`.

---


## Methodology — compute attribution

The [Compute Visibility page](https://scrutica.com/compute-leverage/visibility) groups facilities by the corporate traceability of their linked organization. The classifier chooses the recorded operator, then facility owner, then hardware owner as its starting organization. These fields supply a chain anchor; the classification does not validate the facility-to-organization relationship.

The visibility categories follow these checks, in order:

- **Tier 1:** the starting organization has an SEC identifier, or a recorded ancestor belongs to a configured source group.
- **Tier 2:** an ancestor appears in the specified licensed corporate-ownership source and the Tier 1 check did not match.
- **Tier 3:** an ancestor is documented, with neither earlier condition met.
- **Tier 4:** an organization is linked, but no ancestor appears in the recorded chain.
- **Data gap:** no operator, facility owner or hardware owner is linked.

A Tier 1 result can follow from an organization's SEC identifier alone. It does not establish property ownership, facility operation or authority over the compute.

The Tier 1 source group includes filing, registry and licensed-source records. A chain can terminate at a listed company or government entity, so the category cannot certify that an ultimate beneficial owner was identified. Tier 4 includes organizations that report having no consolidating parent, alongside organizations whose parent information is missing.

Tier shares are weighted by operating capacity: the MW recorded on operational facilities, with campus build-out figures filed on operating sites and earlier phases of operating clusters held out. Figures on records under construction, expanding, announced, permitted, retired or without a status are listed beside each share by status and never added to it. Facilities with no power value can appear in facility counts without contributing to a MW total.

**Country attribution.** The country comparison uses the facility location, the headquarters country of the organization a source establishes as the facility's operator, owner or AI-hardware owner, and the headquarters country of its furthest recorded ancestor, reached through no ownership link reviewed as not establishing control. Agreement among these fields produces the domestic category; agreement between the two organizations in a foreign country produces foreign concentration; different organizational countries produce the bifurcated category. Missing country or ancestry information can leave attribution unresolved. A recorded organization that no source establishes in one of those roles starts no comparison, and an owner named only by a licensed source that has not been reviewed is counted in a category of its own.

Headquarters, legal jurisdiction, hardware title and authority to schedule a workload are distinct facts. The comparison classifies recorded locations and corporate relationships; establishing operational command over the compute requires further evidence.

The [Coordination page](https://scrutica.com/compute-leverage/coordination) examines cross-country relationships against its selected policy groupings. Read the relevant policy and relationship sources before treating a connection as legally restricted or evidence of evasion.

Implementation: `v_facility_visibility_tier` and `v_facility_control_attribution` in the committed visibility migrations, with public aggregation in `src/lib/data/compute-visibility.ts`.

---


## Methodology — Compute Leverage

[Compute Leverage](https://scrutica.com/compute-leverage) asks which suppliers or institutions could prevent a proposed frontier training run from obtaining the resources it needs. Each entry names the holder and gives an editorial veto judgment at 18 months and five years, with the evidence about switching costs and replacement time.

**Judgments and horizons.** YES means the assessment finds that the lab could not obtain an alternative within the horizon; NO means it could; UNCERTAIN records unresolved timing or substitution evidence. The 18-month count concerns the resources needed to establish a new run. Whether a holder can interrupt an operating run depends on the specific resource and the run's existing arrangements.

The implementation selects a stored judgment for the chosen run and horizon, falling back to the default when no run-specific override exists. Switching cost and replacement time are sourced explanatory fields. The code does not multiply them into a numerical score. Market-share figures provide context for the assessment.

**Accountability categories.** Each entry also records an editorial category for a mechanism that could constrain the holder. The stated priority is export controls, antitrust, public authority, competition, then NONE. A category identifies the recorded mechanism; its applicability depends on the transaction and jurisdiction. In particular, an export restriction can limit a sale to one destination without governing a domestic customer's purchase. Competition can exist even when an alternative is unavailable within the selected horizon.

The following 11 entries are the default judgments. An institutional group is counted once, with its members named.

- **EUV lithography** — ASML. Veto judgment at 18 months: YES; at 5 years: YES. Accountability category: EXPORT_CONTROL.
- **Advanced logic foundry (≤5nm/3nm)** — TSMC, Samsung Foundry, Intel Foundry. Veto judgment at 18 months: YES; at 5 years: UNCERTAIN. Accountability category: EXPORT_CONTROL.
- **EDA tools (chip design)** — Synopsys, Cadence, Siemens EDA. Veto judgment at 18 months: YES; at 5 years: YES. Accountability category: EXPORT_CONTROL.
- **HBM memory (HBM3E / HBM4)** — SK Hynix, Samsung, Micron. Veto judgment at 18 months: YES; at 5 years: NO. Accountability category: EXPORT_CONTROL.
- **Advanced packaging / CoWoS** — TSMC, OSATs (ASE, Amkor, SPIL). Veto judgment at 18 months: YES; at 5 years: UNCERTAIN. Accountability category: EXPORT_CONTROL.
- **AI accelerator** — NVIDIA, Google TPU, AWS Trainium, AMD. Veto judgment at 18 months: YES; at 5 years: UNCERTAIN. Accountability category: EXPORT_CONTROL.
- **Scale-up interconnect (NVLink / NVSwitch)** — NVIDIA. Veto judgment at 18 months: YES; at 5 years: UNCERTAIN. Accountability category: EXPORT_CONTROL.
- **Scale-out interconnect (InfiniBand vs Ethernet)** — NVIDIA/Mellanox InfiniBand, Broadcom/Arista Ethernet. Veto judgment at 18 months: NO; at 5 years: NO. Accountability category: COMPETITION.
- **Cloud / datacenter capacity** — AWS, Microsoft Azure, Google Cloud, CoreWeave, Oracle, neoclouds. Veto judgment at 18 months: UNCERTAIN; at 5 years: NO. Accountability category: COMPETITION.
- **Power / grid interconnection** — ISO/RTO (MISO, PJM, ERCOT, SPP), local utility, state PUC, FERC, EPA. Institutional group, counted once. Veto judgment at 18 months: YES; at 5 years: UNCERTAIN. Accountability category: PUBLIC.
- **Capital ($1B+/yr runs)** — hyperscaler balance sheets, sovereign funds (QIA, PIF, Mubadala), mega-VC. Veto judgment at 18 months: UNCERTAIN; at 5 years: NO. Accountability category: COMPETITION.

At 18 months, 8 entries have a YES judgment. Entries that also have NONE in the accountability field: none. This intersection follows the recorded judgments and categories; evaluating a particular holder's legal accountability requires the cited source context.

The interactive page provides the run selection and supporting evidence. Implementation: `vetoFor` and `countVetoes` in `src/lib/data/control-map.ts`.

---


## Reproducing a calculation

A reproducible result needs the source observation, its date and scope, and the model inputs used with it. The following checks distinguish calculations that otherwise share similar units.

**Facility daily budget.** Read the [facility](https://scrutica.com/entities) record's partitions: each must be present capacity, be counted in accelerators and resolve to exactly one accelerator model with a dense FP16 rate that has passed evidence review, and the partition counts must reconcile to the facility total. One held partition withholds the facility figure and names itself. For an admitted inventory, multiply each partition's count by its cited dense FP16 FLOP/s, interconnect efficiency, MFU and 86,400 seconds, then sum the partitions. Named power estimates remain withheld pending separate power-evidence review. The [Capacity estimator](https://scrutica.com/capacity) also provides hypothetical calculators with user-selected inputs; their arithmetic cannot establish a named facility's installed inventory or availability to one job.

**MCP throughput estimate.** `scrutica_estimate_flops` takes a hardware model, count, utilization and precision. Its `peak_pflops` and `sustained_pflops_point_estimate` are rates in petaFLOP/s. It applies fixed lower and upper multipliers of 0.7 and 1.15 to the sustained point estimate; neither multiplier comes from sampling a probability distribution. Both accepted sparsity settings currently use a multiplier of 1.0. This tool does not apply the facility estimator's separate interconnect factor or infer a facility's power allocation. The response returns the inputs and the estimate flag.

**Supply-chain propagation.** Preserve the scenario's initial affected organisations, severity, direction and propagation decay with its output. In the forward calculation, impact is multiplied by supply share, criticality divided by ten, and the scenario's decay at each relationship. The graph builder uses a supply share of 0.3 when the recorded share is missing and a criticality of five when that score is missing; sole-source relationships have a criticality floor of nine. Where a price correlation is available and blending is enabled, the correlation is mapped by `clamp(5 × (1 + r), 1, 10)` and blended with criticality at the configured weight. The defaults and any override affect the result. The [cascade methodology](https://scrutica.com/methodology) and [simulator](https://scrutica.com/cascade) explain the scenario assumptions.

**Sovereign expenditure.** The execution ratio is disbursed USD divided by announced government-only USD, with announced headline USD as the fallback when the government-only amount is missing. A missing disbursement, missing denominator or zero denominator produces no ratio. Read `reality_ratio_denominator_field` and `reality_ratio_denominator_usd` with an exported ratio; a fallback headline can include private investment. Committed funds are a separate stage. The [program records](https://scrutica.com/sovereign-ai) list the amounts and their source context.

**Cloud price.** Reproduce a [cost-index](https://scrutica.com/economics/cost-index) observation from hourly instance price, accelerator count and per-accelerator peak dense BF16 throughput. Divide the hourly price by instance PFLOP/s, then multiply by 24. Use the same region, pricing category and retrieval date when comparing observations. A workload-utilization discount would change the denominator used by this index.

---


## Finding the right record

Use the identifier returned by search when requesting a detail record. Company and facility names can have aliases, and the same company can appear under several source identifiers. Known aliases and recognized source identifiers can resolve to a shared company record. Constructing an identifier from a name can produce a different or nonexistent record.

| Record | Main question | Where to read it |
| --- | --- | --- |
| Facility | What physical site or project is documented, at what status and capacity? | [Facilities](https://scrutica.com/entities), with details at `/facilities/{id}` |
| Company | Which organization is named, and what ownership and supply relationships are documented? | [Companies](https://scrutica.com/entities/companies), with details at `/companies/{id}` |
| Supply relationship | What product or service connects a supplier and customer? | [Supply Chain](https://scrutica.com/supply-chain) |
| Sovereign AI program | What was announced, committed or disbursed, and what hardware evidence exists? | [Sovereign AI](https://scrutica.com/sovereign-ai) |
| Export-control designation | Which entity is named by an authority, in which notice and action? | [Export Controls](https://scrutica.com/export-controls) |
| Accelerator | Which hardware specification and control classification apply to the model? | [Chip register](https://scrutica.com/export-controls/chips) |
| Scenario | What assumptions define the disruption and which records support the case? | [Scenarios](https://scrutica.com/scenarios) |

For a facility, distinguish its operator, facility owner and hardware owner. Epoch's hardware-owner field names who owns the AI hardware; building ownership and operation are separate fields. Power can describe an entire campus, an individual building or IT load; read the source's scope. Accelerator counts can be reported inventories or estimates expressed in equivalent hardware.

For an organization, a missing parent field leaves the parent unknown unless a source supplies an explicit no-parent statement. Headquarters country, registered jurisdiction and a ticker are separate identity fields. A ticker alone gives no measure of ownership concentration.

A company match to a designation needs its own matching evidence. A name resemblance, documented affiliate relationship and explicit designation are different claims. The designation's source notice establishes its scope; a graph connection alone does not.

API response fields vary by tool and endpoint. The [agent manual](https://scrutica.com/agents) and [API specification](https://scrutica.com/api/openapi.json) describe public access. Use the returned fields rather than assuming the internal database schema is exposed.

---


## Following a source

Record pages and tool responses provide source names, links and dates where available. A link can point to a primary disclosure, a research dataset or an explanation of a licensed source's role. Inspect the linked material to establish which field it supports and whether the value was reported or derived.

Dataset pages include machine-readable JSON-LD metadata. Source links appear under `prov:wasDerivedFrom` when supplied, with Scrutica named as the processing agent. Version, publication and modification dates, and download links also depend on what the page supplies. These describe the dataset; a source link for an individual observation must establish that observation's value and scope.

For a calculation, cite both the input's source and the method. A source attached to a facility's identity cannot establish its power rating unless that source actually supplies the rating. The same rule applies to a corporate parent's name, the immediate ownership relationship and the country assigned to it.

When a record has a qualification or correction, preserve it in the resulting answer. Where the response lacks a source for a needed claim, leave that claim unresolved until another source establishes it.

---


## Source classifications

Source tiers summarize the kind of evidence assigned to a record. They do not express statistical confidence, and their exact criteria can differ by instrument.

The shared authority badge uses this scale:

| Tier | Evidence category |
| --- | --- |
| 1 | Primary disclosure, filing or measurement |
| 2 | Research database or industry-standard derived metric |
| 3 | Analyst report, press or interview |
| 4 | Inferred, estimated or Scrutica-derived value |

“Tier unavailable” means the value has no usable classification on this scale. It can reflect a missing or out-of-range tier.

A company's own announcement is primary evidence of what it announced; the same announcement can still describe a future project or an estimate.

Read the definition attached to the specific instrument. In particular, the compute-visibility tiers above classify recorded ownership evidence through their own source predicates. They are not interchangeable with the general authority badge or the estimate flag.

A derived result needs the source and assumptions for its inputs. Source quality alone cannot settle whether a power figure describes IT load, whether a subsidiary relationship is direct, or which funding amount supplied a ratio's denominator.

---


## Glossary (77 terms)

The Scrutica glossary defines the technical terms the instruments use. The full interactive glossary with cross-linked related terms is at https://scrutica.com/methodology/glossary.

### Semiconductor

**Advanced packaging**. Techniques that connect multiple dies within one package, including side-by-side integration and vertical stacking. They allow logic and memory manufactured separately to work together. *Governance relevance:* Packaging has its own process requirements and capacity constraints. A supply assessment must match the package a chip requires to the facilities able to produce it.

**CoWoS**. Chip-on-Wafer-on-Substrate: TSMC’s packaging family for connecting logic dies and high-bandwidth memory. CoWoS-S uses a silicon interposer; CoWoS-R uses redistribution layers, and CoWoS-L combines them with local silicon interconnects. *Governance relevance:* A chip designed for one CoWoS variant needs the corresponding packaging process and capacity. Aggregate advanced-packaging capacity can therefore overstate the capacity available to a specific accelerator.

**DUV lithography (DUV)**. Deep ultraviolet lithography uses 193 nm argon-fluoride or 248 nm krypton-fluoride light. Multiple patterning adds exposures and processing steps to print finer features. *Governance relevance:* DUV equipment remains in use alongside EUV in advanced chip production. Export restrictions distinguish equipment specifications and destinations; DUV covers several machine classes.

**EUV lithography (EUV)**. Extreme ultraviolet lithography uses 13.5 nm light to print circuit patterns on wafers. A tin-plasma light source and multilayer mirrors operate in a vacuum. *Governance relevance:* ASML is the sole supplier of EUV lithography systems. Access to these machines constrains which manufacturers can add processes that require EUV.

**Fabless**. A chip company that designs chips and contracts out their fabrication. NVIDIA and AMD use this model. *Governance relevance:* Producing a fabless company’s design requires access to a compatible foundry process. Restrictions or shortages at that foundry can affect chip availability even when the designer can continue its own work.

**FinFET**. Fin Field-Effect Transistor: a transistor whose gate surrounds the sides and top of a raised silicon channel, called a fin. *Governance relevance:* The transistor architecture is one part of a manufacturing process. FinFET and gate-all-around processes require different designs and production techniques, which matters when assessing replacement foundry capacity.

**Foundry**. A semiconductor manufacturer that fabricates chips designed by customers. A customer must use a design compatible with the foundry’s process. *Governance relevance:* Market concentration depends on the denominator. Revenue across all foundry services, wafer capacity at a particular process generation, and capacity qualified for one chip measure different kinds of supply.

**GAA**. Gate-All-Around transistor: a transistor whose gate surrounds the channel on all sides. Intel’s RibbonFET uses ribbon-shaped silicon channels. *Governance relevance:* GAA changes the transistor structure and the process used to make it. Access to a fab with a similar node label does not establish that it can manufacture the same design.

**HBM**. High Bandwidth Memory: stacks of DRAM dies connected vertically by through-silicon vias. Several stacks can supply memory to one accelerator package. *Governance relevance:* Accelerator production requires compatible memory as well as fabricated logic dies. Compare HBM capacity and bandwidth at the same level: an individual stack, a GPU package, or a complete system. *Example:* NVIDIA specifies 141 GB of HBM3e and 4.8 TB/s memory bandwidth for one H200 GPU.

**IDM**. Integrated Device Manufacturer: a company that both designs and fabricates chips. Intel and Texas Instruments are examples. *Governance relevance:* An IDM controls fabrication for the chips it makes internally. Its exposure still depends on the equipment, materials and outsourced manufacturing services used for the particular product.

**OSAT**. Outsourced Semiconductor Assembly and Test: a company that packages and tests fabricated chips. ASE and Amkor provide these services. *Governance relevance:* A finished accelerator needs packaging and testing capacity suited to its design. An OSAT’s conventional packaging capacity gives little information about which advanced packages it can manufacture.

**Photomask**. A patterned template, also called a reticle, from which a lithography system projects circuit features onto a wafer. *Governance relevance:* Defects in a reticle can be repeated across the chips exposed with it. Reticle inspection and pattern accuracy are therefore part of controlling manufacturing yield.

**Process node**. A manufacturer’s name for a semiconductor process generation, such as N4 or N7. The number in a modern node name does not specify a physical gate length or permit direct comparison between manufacturers. *Governance relevance:* Assess a chip’s performance and the relevant export classification separately. A node label alone leaves both its capabilities and the transaction’s licence requirements unresolved.

**Wafer**. A thin disc of semiconductor material on which chip dies are fabricated. Wafer starts per month measures how many wafers enter production. *Governance relevance:* Converting wafer capacity into chip output requires the die area and functional yield for the product being made. Capacity on a process unsuitable for that design cannot supply it.

**Yield**. The fraction of manufactured dies that meet the required specifications. *Governance relevance:* For a given number of dies fabricated, lower yield reduces the number available for packaging and sale. Wafer capacity needs a yield assumption before it can support a usable-chip estimate.

### Accelerator

**BF16**. A 16-bit floating-point format with one sign bit, eight exponent bits and seven fraction bits. Its exponent range matches FP32, with less precision for individual values. *Governance relevance:* Scrutica’s cloud cost comparison uses dense BF16 peak throughput. That convention fixes the unit of comparison; workload performance still depends on how effectively the rented hardware is used.

**FLOP**. A floating-point arithmetic operation. FLOP/s measures the rate of these operations; multiplying a sustained rate by elapsed seconds gives a cumulative operation count. *Governance relevance:* A training-compute threshold concerns the operations attributed to training a model. Applying it to hardware capacity requires assumptions about training duration and sustained throughput.

**FP16**. A 16-bit floating-point format with one sign bit, five exponent bits and ten fraction bits. Mixed-precision training can combine FP16 arithmetic with higher-precision operations. *Governance relevance:* The format affects numerical range and the throughput a processor can deliver. A chip’s quoted FLOP/s rate needs its precision format before it can be compared with another rating.

**InfiniBand**. A network technology used for high-bandwidth, low-latency communication between servers, including servers in AI training clusters. *Governance relevance:* Distributed training transfers data between accelerators on different servers. Network bandwidth and congestion can limit how much additional throughput a larger cluster supplies.

**Memory bandwidth**. The rate at which data can move between a processor and its memory, usually expressed in GB/s or TB/s. *Governance relevance:* A workload with few arithmetic operations per byte transferred can be limited by memory bandwidth. In that case, raising peak arithmetic throughput alone gives little benefit.

**NVLink**. NVIDIA’s interconnect for high-bandwidth communication between supported processors. *Governance relevance:* Training efficiency depends on the communication a workload requires and the links available between its accelerators. Compare a complete system’s topology and bandwidth before inferring performance from NVLink support.

**PCIe**. PCI Express: an interface for connecting devices within a computer. *Governance relevance:* A PCIe label does not describe all GPU-to-GPU connectivity in a system. NVIDIA’s H100 NVL, for example, supports an NVLink bridge between PCIe GPUs.

**SXM**. NVIDIA’s module form factor for GPUs installed on a server baseboard. *Governance relevance:* SXM and PCIe configurations can differ in memory, power and communication bandwidth. These specifications help assess a system’s training capacity; form factor alone is insufficient.

**TDP**. Thermal Design Power: a manufacturer’s thermal design rating, in watts, used to size cooling. *Governance relevance:* Scrutica’s power estimate divides facility power by PUE, applies the accelerator share of IT power, then divides by the selected GPU’s power rating to estimate GPU count.

**Tensor core**. NVIDIA hardware units that accelerate matrix multiply-and-accumulate operations. *Governance relevance:* AI workloads use both matrix operations and other computation. A tensor-operation peak rating describes one part of the processor’s performance; completed workload throughput requires a measurement or model of the whole job.

**TFLOP/s**. One trillion (10¹²) floating-point operations per second. Hardware ratings specify the arithmetic format and whether structured sparsity contributes to the reported rate. *Governance relevance:* Comparisons need the same precision and sparsity convention. Estimating completed training work also requires a sustained-utilization assumption and a duration.

### Training

**Chinchilla-optimal**. A compute-constrained balance of model size and training data studied by Hoffmann et al. (2022). Their results favored scaling parameters and training tokens together, with about 20 tokens per parameter in the Chinchilla training run. *Governance relevance:* The balance depends on the optimization objective. A model trained to minimize loss under a fixed training budget may differ from one chosen to minimize the combined cost of training and repeated inference.

**Fine-tuning**. Further training of an existing model, often to adapt it to a task or dataset. It may update all model weights or a smaller set of parameters. *Governance relevance:* Fine-tuning changes the model being evaluated. The relevant evidence includes the resulting capabilities and the additional training performed; the original model’s training budget alone leaves those changes unspecified.

**Inference**. Using a trained model to produce an output. Compute demand depends on the model, input and output lengths, and how requests are processed together. *Governance relevance:* Capacity planning must account for the number and composition of requests as well as compute per request. A training-compute total gives no direct estimate of the resources needed to serve a model.

**METR time horizon**. METR’s estimate of the task duration at which an AI agent reaches a specified success rate, commonly 50%. Duration refers to the time human experts need for the tasks. *Governance relevance:* The estimate depends on the evaluation’s task distribution. METR’s suite emphasizes software, machine-learning and cybersecurity tasks, so a longer horizon measures progress within that tested scope.

**MFU**. Model FLOP Utilization: model operations per second divided by the hardware’s theoretical peak rate. The model-operation count excludes activation recomputation; hardware FLOP utilization can include that extra work. *Governance relevance:* An MFU assumption converts peak capacity into a training-throughput estimate. Scrutica’s facility estimator defaults to 40% when no MFU override is supplied. *Example:* The PaLM paper reports 46.2% model FLOP utilization and 57.8% hardware FLOP utilization for PaLM 540B, with activation recomputation included in the latter.

**Parameter count**. The number of learned weights in a model. The parameters used for an individual input can be a subset of the total in a mixture-of-experts model. *Governance relevance:* Total parameter count helps estimate weight-storage requirements. Estimating computation also requires the model architecture and the parameters active for each input.

**Scaling law**. An empirical relationship between model performance and inputs such as parameter count, training data and compute. Kaplan et al. (2020) and Hoffmann et al. (2022) study such relationships for language-model loss. *Governance relevance:* A fitted scaling relationship can estimate trade-offs within its experimental setting. Applying it to a new architecture or a particular capability requires evidence that the relationship transfers.

**Training compute**. The operations used to train a model, commonly counted in FLOP. A reported total needs a defined scope, including which training stages it counts. *Governance relevance:* The EU AI Act’s Article 51(2) uses cumulative training compute above 10²⁵ FLOP for a presumption of high-impact capabilities in a general-purpose AI model. Providers may contest classification under Article 52.

### Datacenter

**GPU cluster**. A group of connected GPU servers configured to work together. *Governance relevance:* A facility’s installed GPU inventory can contain separate clusters or serve several users. Estimating time to a training-compute total requires assumptions about the resources assigned to one run.

**Hyperscaler**. An operator of large-scale computing infrastructure, often spread across many data centers. The term includes major cloud providers and companies that operate infrastructure for their own services. *Governance relevance:* An operator’s total estate can serve several workloads and customers. Its AI capacity depends on the equipment deployed and the resources available to those workloads.

**Interconnect fabric**. The links and switches that connect processors and servers. A cluster may combine local accelerator links with an InfiniBand or Ethernet network between servers. *Governance relevance:* Distributed workloads exchange activations or synchronize results across these links. Bandwidth, latency and topology affect the time spent communicating as the workload expands across more devices.

**Liquid cooling**. Cooling that carries heat away in a liquid, through cold plates attached to components or by immersing equipment in a suitable fluid. *Governance relevance:* High-density deployments require a compatible cooling design, including the server’s heat-transfer system and the facility’s heat rejection. A GPU’s power rating alone does not specify that design.

**MW**. Megawatt: one million watts, a unit of power. Energy consumption also requires a duration; one megawatt sustained for one hour is one megawatt-hour. *Governance relevance:* A data center’s announced power capacity needs a status and scope: an energized building, a future phase, or an entire campus. Those distinctions affect how much capacity is available for a compute estimate.

**Neocloud**. A cloud provider focused on supplying GPU computing capacity. CoreWeave, Lambda and Nebius offer such services. *Governance relevance:* A provider’s GPU inventory, hosting arrangements and hardware suppliers affect the capacity it can offer. Company-level sourcing and contract terms are needed to assess dependence on a particular chip maker.

**PUE**. Power Usage Effectiveness: total facility energy divided by IT equipment energy over the same period. IT equipment includes servers, storage and networking. *Governance relevance:* Scrutica uses PUE to convert facility power into IT power, then applies a separate accelerator share. PUE by itself leaves the allocation of IT power among GPUs and other equipment unspecified. *Example:* A PUE of 1.2 means facility overhead consumes energy equal to 20% of the IT equipment’s energy use.

**Utilization rate**. The fraction of available capacity used over a stated period. For GPUs, the measure may refer to occupied device time or to a particular activity counter; the denominator needs to be specified. *Governance relevance:* An occupancy measure records how much hardware time jobs use. MFU measures model throughput during a job. Combining them requires matching the hardware population and time window to avoid counting idle time twice.

### Supplychain

**Cascade**. Disruption that spreads through business dependencies. Scrutica’s simulator follows recorded relationships and weights transmitted impact by supply share, criticality and a scenario-level decay factor. *Governance relevance:* The result depends on the relationships available and the model’s propagation assumptions. Missing supply shares use a default; changing that assumption can change downstream impacts.

**Chokepoint**. A supply-chain dependency whose disruption is difficult to avoid because suitable alternatives are limited. *Governance relevance:* Assess the affected product and the time needed to qualify or expand replacement supply. Market concentration describes supplier shares; substitutability describes the options available after a disruption.

**HHI**. Herfindahl-Hirschman Index: the sum of squared market shares, expressed as percentages. A single supplier has an HHI of 10,000; splitting supply among more similarly sized firms reduces the index. *Governance relevance:* The US agencies’ 2023 Merger Guidelines classify markets above 1,800 HHI as highly concentrated. A merger that also increases HHI by more than 100 points meets their structural presumption; applying it requires a defined relevant market.

**Licensed supply-chain database**. A subscription source of business relationships. Its records can include supplier–customer links and attributes such as relationship value or price correlation. *Governance relevance:* Source coverage determines which relationships enter an analysis. Published aggregates and derived results also depend on the permitted use of the licensed records.

**Price correlation**. A Pearson correlation between two companies’ stock prices over a stated window. Scrutica’s licensed relationship records can include a three-month value. *Governance relevance:* When enabled in the cascade model, correlation maps to criticality as 5 × (1 + r), clamped to 1–10, then blends with the edge's criticality. Co-movement can reflect shared market influences as well as a business relationship.

**Sole-source dependency**. A dependency in which only one supplier can meet the requirement for a particular input or qualified process. Other suppliers may serve the wider market. *Governance relevance:* Scrutica records a relationship as sole-source only where its cited source says so for that buyer; a supplier’s market share, however high, describes the market rather than any one customer. The cascade model gives such a relationship a minimum criticality of 9 on its 1–10 scale. Supply share and the scenario’s decay factor also affect the modeled disruption.

**Substitutability**. The ability to replace an input or supplier with an alternative that meets the required specifications. It depends on the product and the time available to qualify or expand replacement supply. *Governance relevance:* Scrutica’s category-level reference scores and scenario propagation factors serve separate purposes. Assess the actual replacement options before treating either as a measure of available supply.

**Supply chain edge**. A directed connection representing a recorded business relationship. A supplier–customer pair can have separate connections for different products or services. *Governance relevance:* The relationship’s source and product determine what dependency it supports. An unreported supply share remains a missing input to any calculation of how much the customer relies on that supplier.

### Governance

**Compute governance**. Policies that use access to computing resources, or information about their production and use, to influence AI development and deployment. *Governance relevance:* Relevant questions include where compute is available, who can obtain it and which suppliers or authorities can constrain access. Hardware ownership, physical location and regulatory jurisdiction each supply different evidence.

**Dual-use**. Goods, software or technology that can serve civilian and military purposes. *Governance relevance:* A dual-use description identifies possible applications. Export obligations follow the applicable rules, including item classifications and controls on users or uses.

**EU AI Act**. Regulation (EU) 2024/1689, which sets rules for AI systems and general-purpose AI models, with obligations that depend on the applicable category. *Governance relevance:* Article 51(2) sets a rebuttable presumption of high-impact capabilities above 10²⁵ cumulative training FLOP. The Commission can also designate a model using other capability criteria.

**GPAI**. General-purpose AI models, as defined by the EU AI Act, have significant generality and can perform a wide range of distinct tasks, including when incorporated into other systems. *Governance relevance:* Article 53 sets provider obligations for documentation, information to downstream developers, copyright policy and training-content summaries, with specified exceptions. Systemic-risk classification adds further duties.

**Proliferation**. The spread of a capability to additional actors. In AI, this can involve access to models, computing resources or the expertise needed to use them. *Governance relevance:* An assessment needs to specify the capability and means of access. A hardware transfer and remote access to a hosted model create different dependencies and control options.

**Reality ratio**. Scrutica’s ratio of reported disbursed spending to announced government spending. When the government figure is absent, the calculation falls back to the headline commitment, which may include private capital. *Governance relevance:* Read the numerator and denominator together with their dates. The ratio measures reported financial disbursement; evidence of completed or operational infrastructure requires a separate assessment.

**Sovereign AI**. National initiatives to develop AI capabilities, including domestic computing infrastructure and access arrangements with commercial providers. *Governance relevance:* Funding commitments, disbursements and operational capacity describe different stages of a program. Supplier relationships help assess its dependence on equipment and services from other jurisdictions.

**Systemic risk**. Under the EU AI Act, risk associated with a general-purpose model’s high-impact capabilities that can spread at scale across the value chain, with significant effects on matters such as safety or fundamental rights. *Governance relevance:* Providers of models classified as posing systemic risk have Article 55 duties, including model evaluation and adversarial testing, risk mitigation, incident reporting and cybersecurity.

### Exportcontrol

**BIS**. Bureau of Industry and Security: the US Department of Commerce agency that administers the EAR, issues licences and maintains the Entity List. *Governance relevance:* BIS rules and individual licence decisions affect access to controlled computing equipment. The relevant rule and its effective date belong with any assessment of a transaction.

**De minimis rule**. EAR rules that assess controlled US-origin content in a foreign-made item. The applicable percentage depends on the item and destination, with some categories subject to no minimum. *Governance relevance:* Foreign-direct-product rules can separately bring a foreign-made item under the EAR. A de minimis calculation alone therefore does not settle jurisdiction.

**Deemed export**. Under the EAR, a release of controlled technology or source code to a foreign person in the United States can constitute an export to that person’s relevant country of citizenship or permanent residency. *Governance relevance:* Collaboration can involve a controlled release even when no equipment leaves the country. Whether a licence is needed depends on the technology, the recipient and applicable exclusions or authorizations.

**EAR**. Export Administration Regulations: US rules administered by the Bureau of Industry and Security that govern covered exports, reexports and transfers, along with specified activities by US persons. *Governance relevance:* Licence requirements depend on the item’s classification and the parties, destination and use involved. An advanced-computing specification is one part of that assessment.

**End-use controls**. Export restrictions tied to the intended use of an item or specified activities. They can apply in addition to controls based on the item’s technical classification. *Governance relevance:* The destination’s role in a transaction matters: an item that appears unrestricted by its specifications can still require authorization because of its intended use.

**Entity List**. A list maintained by BIS whose entries specify licence requirements and review policies for transactions involving the named entities. *Governance relevance:* Read the applicable entry’s item scope, review policy and restrictions on licence exceptions. These can differ between entities on the list.

**License exception**. An authorization under the EAR for a transaction that meets specified conditions, without an individual licence application. *Governance relevance:* Eligibility depends on the exception’s terms and applicable restrictions, including those attached to the destination, item or listed entity.

**TOPS**. Tera operations per second: 10¹² operations per second. A quoted rate needs the operation type, precision and sparsity convention. *Governance relevance:* The EAR’s advanced-computing definition of total processing performance uses 2 × MacTOPS × operation bit length, aggregated across processing units. A marketing TOPS rating needs these details before it can support that calculation.

**Wassenaar Arrangement**. A multilateral arrangement for coordinating controls on conventional arms and dual-use goods and technologies. Participating states agree control lists and related guidance. *Governance relevance:* National authorities decide whether to authorize a transfer. Membership identifies participation in the arrangement; assessing a particular export requires the applicable national rules.

### Methodology

**Authority tier**. Scrutica’s source classification: Tier 1 covers primary disclosures, filings and measurements; Tier 2 covers research databases and established derived metrics; Tier 3 covers press and analyst reporting; Tier 4 covers inferred or estimated values. *Governance relevance:* The tier describes the source category. Assess the cited claim’s scope and date separately, including whether it reports an observation, an announcement or a projection.

**BFS propagation (BFS)**. Breadth-first search explores a graph outward in stages from a starting point. Scrutica uses it to follow modeled disruptions through recorded business relationships. *Governance relevance:* Each transmission is weighted, and the simulator limits traversal depth and revisits. The resulting reach depends on these rules and the available relationships.

**Capex disaggregation**. Allocating capital expenditure among components such as buildings, power infrastructure and computing equipment. *Governance relevance:* Scrutica’s facility cost estimate applies an assumed accelerator share to reported investment, then divides by the selected GPU price. The estimate requires investment attributable to the facility and an appropriate hardware-price assumption.

**Confidence interval**. An interval produced by a statistical procedure with a specified repeated-sampling coverage rate. A 95% procedure would contain the true parameter in 95% of repeated applications under its assumptions. *Governance relevance:* Coverage requires a statistical model and the conditions on which it rests. Scrutica’s facility ranges vary assumed inputs; they have no assigned confidence level.

**Decay rate**. In Scrutica’s cascade model, a scenario-level multiplier applied when transmitting disruption across a relationship. Supply share and criticality contribute additional multipliers. *Governance relevance:* Changing the decay factor tests how far a disruption travels under that assumption. The scenario value is an editorial input; the separate category substitutability table does not set it.

**Estimation bounds**. Ranges obtained by varying assumptions in an estimate. Scrutica’s facility estimator varies inputs within each available hardware, power or cost calculation. *Governance relevance:* The calculations can share hardware and utilization assumptions, so agreement provides limited corroboration. The ranges describe sensitivity to the selected inputs and have no statistical coverage rate.

**is_estimated**. A per-record flag, set by the pipeline step that last wrote it, under that step’s own rule. True means that step judged the record estimated, on grounds that include a modeled, derived or inferred value, an unconfirmed record and a secondary source. False means no such judgment is recorded; it is not a verification. Read the flag with the record’s authority tier. *Governance relevance:* When the flag is true, retain the estimation method and assumptions with the value. A false flag supplies no independent verification of the underlying claim. A record holding a figure its source calculated can be flagged false.

**Petaflop-day (PF-day)**. One petaFLOP/s (10¹⁵ floating-point operations per second) sustained for 86,400 seconds: 8.64 × 10¹⁹ FLOP. *Governance relevance:* Scrutica’s cloud cost index prices this unit using dense BF16 peak throughput. The calculation excludes workload efficiency, so training time and cost require additional utilization assumptions.

**Price per petaflop-day**. Cloud rental price per petaflop-day = hourly instance price × 24 ÷ (GPU count × dense BF16 TFLOP/s per GPU ÷ 1,000). *Governance relevance:* Cloud rental prices use peak capacity for this normalization. On-premises cost estimates include an assumed utilization rate; compare those denominators along with the offer’s region and contract terms.

**Provenance**. Information about where a value came from and how it was transformed. It can include a source document, the date it was accessed and the calculation applied. *Governance relevance:* A citation must support the particular claim being made. For derived figures, the inputs and transformation are needed to reproduce the result.

---


## Selected companies

These descriptions explain what selected companies supply or develop, with primary-source links. Headquarters and ticker fields come from the company records. The [company directory](https://scrutica.com/entities/companies) provides the broader collection and recorded ownership relationships.

### NVIDIA Corporation (`org-nvidia`)

Supplies data-center GPU accelerators and the CUDA software platform for GPU programming. CUDA provides libraries, a compiler, debugging tools and a runtime used to develop accelerated applications. Sources: [1](https://developer.nvidia.com/cuda/gpus), [2](https://developer.nvidia.com/CUDA). Headquartered in the United States. Public-market ticker: NVDA. Detail page: https://scrutica.com/companies/org-nvidia.

### Taiwan Semiconductor Manufacturing Company (`org-tsmc`)

Manufactures semiconductors for chip-design customers and provides advanced packaging. Its CoWoS packages combine logic chips with high-bandwidth memory, so fabrication and packaging are separate parts of its contribution to AI hardware. Sources: [1](https://investor.tsmc.com/sites/ir/annual-report/2025/2025%20Annual%20Report_E.pdf), [2](https://3dfabric.tsmc.com/english/dedicatedFoundry/technology/cowos.htm). Headquartered in Taiwan. Public-market ticker: TSM. Detail page: https://scrutica.com/companies/org-tsmc.

### ASML (`org-asml`)

Supplies EUV and DUV lithography equipment for semiconductor manufacturing. ASML is the sole supplier of EUV systems; chipmakers also use DUV equipment for other layers in the same chip. Sources: [1](https://www.asml.com/en/products), [2](https://www.asml.com/en/en/products/euv-lithography-systems). Headquartered in the Netherlands. Public-market ticker: ASML. Detail page: https://scrutica.com/companies/org-asml.

### Samsung (`org-samsung`)

Produces high-bandwidth memory for AI and HPC systems and manufactures logic chips through Samsung Foundry. Its foundry services include process technology, design support and advanced packaging. Sources: [1](https://semiconductor.samsung.com/dram/hbm/), [2](https://semiconductor.samsung.com/foundry/about-samsung-foundry/company-info/). Headquartered in South Korea. Public-market ticker: 005930. Detail page: https://scrutica.com/companies/org-samsung.

### SK Hynix (`org-sk-hynix`)

Produces high-bandwidth memory for AI accelerators, with vertically stacked DRAM connected through silicon vias. It also supplies server DRAM and solid-state storage for data centers. Sources: [1](https://news.skhynix.com/en/exploring-the-ai-ecosystem-how-sk-hynixs-industry-leading-memory-fuels-ai-innovation/), [2](https://news.skhynix.com/en/dtf-2026/). Headquartered in South Korea. Public-market ticker: 000660. Detail page: https://scrutica.com/companies/org-sk-hynix.

### Micron (`org-micron`)

Supplies high-bandwidth memory for AI accelerators, DDR5 memory for servers and solid-state storage for data-center workloads. Sources: [1](https://www.micron.com/markets-industries/ai/ai-data-center). Headquartered in the United States. Public-market ticker: MU. Detail page: https://scrutica.com/companies/org-micron.

### Huawei (`org-huawei`)

Develops Ascend AI processors and Atlas computing systems built around them. Sources: [1](https://www.huawei.com/en/news/2025/9/hc-xu-keynote-speech). Headquartered in China. Detail page: https://scrutica.com/companies/org-huawei.

### SMIC (`org-smic`)

Provides semiconductor foundry and process-technology services to chip-design customers. Its disclosed process portfolio spans established technologies through FinFET manufacturing. Sources: [1](https://www.smics.com/uploads/64f6fb62/e00981%20Report.pdf). Headquartered in China. Detail page: https://scrutica.com/companies/org-smic.

### Hua Hong (`org-hua-hong`)

Operates wafer foundries with a focus on specialty integrated circuits and power semiconductors. Its 2024 annual report identifies embedded and standalone non-volatile memory among its specialty technologies. Sources: [1](https://www1.hkexnews.hk/listedco/listconews/sehk/2025/0327/2025032702315.pdf). Headquartered in China. Public-market ticker: 1347.HK. Detail page: https://scrutica.com/companies/org-hua-hong.

### Intel (`org-intel`)

Supplies Xeon server processors and Gaudi AI accelerators, and provides chip fabrication, packaging and testing through Intel Foundry. Sources: [1](https://www.intel.com/content/www/us/en/products/details/processors/xeon.html), [2](https://www.intel.com/content/www/us/en/products/details/processors/ai-accelerators/gaudi.html), [3](https://www.intel.com/content/www/us/en/foundry/overview.html). Headquartered in the United States. Public-market ticker: INTC. Detail page: https://scrutica.com/companies/org-intel.

### AMD (`org-amd`)

Supplies Instinct accelerators, EPYC server CPUs and ROCm software for AI and high-performance computing. Sources: [1](https://www.amd.com/en/products/accelerators/instinct.html), [2](https://www.amd.com/en/products/processors/server/epyc/ai/9005-host-cpu-gpu.html). Headquartered in the United States. Public-market ticker: AMD. Detail page: https://scrutica.com/companies/org-amd.

### Tokyo Electron (`org-tokyo-electron`)

Makes semiconductor production equipment for deposition, photoresist coating and development, etching and cleaning. Its product range also includes wafer probing and bonding equipment. Sources: [1](https://www.tel.com/product/), [2](https://www.tel.com/corporatesummary/). Headquartered in Japan. Detail page: https://scrutica.com/companies/org-tokyo-electron.

### Applied Materials (`org-applied-materials`)

Makes equipment that deposits, shapes, removes and measures materials during semiconductor production. Its process tools help form the transistor structures and interconnects on a chip. Sources: [1](https://www.appliedmaterials.com/us/en/semiconductor/products.html), [2](https://www.appliedmaterials.com/us/en/semiconductor/products/create/metals-deposition.html). Headquartered in the United States. Public-market ticker: AMAT. Detail page: https://scrutica.com/companies/org-applied-materials.

### KLA (`org-kla`)

Supplies inspection, metrology and data-analysis products for semiconductor process control. These tools detect defects and measure fabricated features as chipmakers develop processes and increase production yield. Sources: [1](https://ir.kla.com/sec-filings/annual-reports/content/0001193125-25-213412/0001193125-25-213412.pdf), [2](https://www.kla.com/wp-content/uploads/ChipManufacturing101.pdf). Headquartered in the United States. Public-market ticker: KLAC. Detail page: https://scrutica.com/companies/org-kla.

### Lam Research (`org-lam-research`)

Makes semiconductor fabrication equipment for etching, deposition and wafer cleaning. Etch tools remove selected material to form features such as transistors, contacts and wiring. Sources: [1](https://www.lamresearch.com/products/our-processes/), [2](https://www.lamresearch.com/product/reliant-etch-products/). Headquartered in the United States. Public-market ticker: LRCX. Detail page: https://scrutica.com/companies/org-lam-research.

### Equinix, Inc. (`org-equinix`)

Provides colocation space where customers install their servers, with physical connections to other networks and services in its IBX facilities. Sources: [1](https://docs.equinix.com/colocation/). Headquartered in the United States. Public-market ticker: EQIX. Detail page: https://scrutica.com/companies/org-equinix.

### Digital Realty (`org-digital-realty`)

Provides data-center space and colocation services across its international facility network. Sources: [1](https://www.digitalrealty.com/data-centers). Headquartered in the United States. Public-market ticker: DLR. Detail page: https://scrutica.com/companies/org-digital-realty.

### CoreWeave (`org-coreweave`)

Provides GPU cloud infrastructure for AI training and inference, including bare-metal compute, storage and cluster orchestration. Sources: [1](https://docs.coreweave.com/product-home), [2](https://coreweave.com/products/gpu-compute). Headquartered in the United States. Detail page: https://scrutica.com/companies/org-coreweave.

### Microsoft Corporation (`org-microsoft`)

Operates Azure and designs processors for its cloud infrastructure, including Maia AI accelerators and Cobalt CPUs. Maia 200 is designed for inference workloads. Sources: [1](https://azure.microsoft.com/en-us/explore/global-infrastructure/silicon-systems), [2](https://blogs.microsoft.com/blog/2026/01/26/maia-200-the-ai-accelerator-built-for-inference/). Headquartered in the United States. Public-market ticker: MSFT. Detail page: https://scrutica.com/companies/org-microsoft.

### Amazon.com Inc (`org-amazon`)

Operates AWS and designs the Trainium and Inferentia accelerators available through its cloud services. The Neuron software stack supports training and inference on these chips. Sources: [1](https://aws.amazon.com/silicon-innovation/), [2](https://aws.amazon.com/ai/machine-learning/neuron/). Headquartered in the United States. Public-market ticker: AMZN. Detail page: https://scrutica.com/companies/org-amazon.

### Alphabet Inc (Google) (`org-alphabet`)

Provides Google Cloud TPU capacity for model training and inference. TPU hardware has generation-specific configurations and software support, which matter when comparing its throughput with GPU systems. Sources: [1](https://cloud.google.com/tpu/docs/v6e-training), [2](https://docs.cloud.google.com/tpu/docs/tpu-inference?authuser=0). Headquartered in the United States. Public-market ticker: GOOGL. Detail page: https://scrutica.com/companies/org-alphabet.

### Meta Platforms Inc (`org-meta`)

Builds infrastructure to train and serve its AI models, including Llama. Its MTIA accelerator program includes chips for recommendation and ranking workloads. Sources: [1](https://engineering.fb.com/2025/09/29/data-infrastructure/metas-infrastructure-evolution-and-the-advent-of-ai/), [2](https://engineering.fb.com/2026/08/24/networking-traffic/mtia-300-meta-training-chip-built-in-nics/). Headquartered in the United States. Public-market ticker: META. Detail page: https://scrutica.com/companies/org-meta.

### OpenAI (`org-openai`)

Develops the GPT family of AI models and makes models available through its API. Sources: [1](https://developers.openai.com/api/docs/models). Headquartered in the United States. Detail page: https://scrutica.com/companies/org-openai.

### Anthropic (`org-anthropic`)

Develops Claude models. Its April 2026 agreement with Amazon describes additional Trainium capacity for training and deploying Claude; model access is also offered through cloud platforms. Sources: [1](https://www.anthropic.com/news/anthropic-amazon-compute?invite=1), [2](https://www.anthropic.com/claude/fable). Headquartered in the United States. Detail page: https://scrutica.com/companies/org-anthropic.

### Google DeepMind (`org-google-deepmind`)

Develops Gemini and Gemma models and publishes model cards describing their design and evaluation. Sources: [1](https://deepmind.google/models/model-cards/). Headquartered in the United Kingdom. Detail page: https://scrutica.com/companies/org-google-deepmind.

### xAI (`org-xai`)

Develops Grok models and uses the Colossus computing facility in South Memphis for model training. Sources: [1](https://x.ai/memphis/xai), [2](https://x.ai/news/grok-4?_bhlid=b9339d7816a05adeb52bae7050cc74d4fdc85352). Headquartered in the United States. Detail page: https://scrutica.com/companies/org-xai.

### Mistral (`org-mistral`)

Develops AI models and offers dedicated GPU capacity for training and inference through Mistral AI Cloud in Europe. Sources: [1](https://mistral.ai/), [2](https://mistral.ai/products/aicloud/). Headquartered in France. Detail page: https://scrutica.com/companies/org-mistral.

### Apple Inc (`org-apple`)

Designs Apple silicon and operates Private Cloud Compute for Apple Intelligence workloads that exceed on-device capacity. Its June 2026 disclosure describes an expansion of that system to Google Cloud. Sources: [1](https://security.apple.com/blog/private-cloud-compute/), [2](https://security.apple.com/blog/expanding-pcc/). Headquartered in the United States. Public-market ticker: AAPL. Detail page: https://scrutica.com/companies/org-apple.

### Oracle (`org-oracle`)

Provides GPU virtual machines and bare-metal instances through Oracle Cloud Infrastructure for AI training, inference and high-performance computing. Sources: [1](https://www.oracle.com/cloud/compute/gpu/). Headquartered in the United States. Public-market ticker: ORCL. Detail page: https://scrutica.com/companies/org-oracle.

### IBM (`org-ibm`)

Develops processors and accelerators for enterprise AI. Telum integrates inference into IBM Z transaction processing; Spyre provides a separate PCIe accelerator for inference workloads. Sources: [1](https://research.ibm.com/blog/spyre-for-z?_bhlid=0fed2e5662134506acde6a2fb3ded84ead5e0c0a), [2](https://research.ibm.com/publications/spyre-an-inference-optimized-scalable-ai-accelerator-for-enterprise-workloads). Headquartered in the United States. Public-market ticker: IBM. Detail page: https://scrutica.com/companies/org-ibm.

### ARM (`org-arm`)

Supplies CPU designs and system architecture for cloud and data-center processors through its Neoverse platform. The portfolio includes processor IP and integrated compute subsystems. Sources: [1](https://www.arm.com/products/cloud-datacenter), [2](https://www.arm.com/products/silicon-ip-cpu/neoverse/neoverse-v3), [3](https://www.arm.com/products/cloud-datacenter/neoverse-compute-subsystems). Headquartered in the United Kingdom. Public-market ticker: ARM. Detail page: https://scrutica.com/companies/org-arm.

### Cadence (`org-cadence`)

Provides electronic design automation tools for chip design, simulation and verification. These tools help engineers check circuit behavior and manufacturing requirements before fabrication. Sources: [1](https://www.cadence.com/en_US/home/explore/what-is-electronic-design-automation.html). Headquartered in the United States. Public-market ticker: CDNS. Detail page: https://scrutica.com/companies/org-cadence.

### Synopsys (`org-synopsys`)

Provides semiconductor design and verification software and licenses chip IP. Its tools cover digital and analog circuit design, simulation and emulation, with process-modeling software for manufacturing. Sources: [1](https://www.synopsys.com/silicon-design.html), [2](https://www.synopsys.com/verification.html). Headquartered in the United States. Public-market ticker: SNPS. Detail page: https://scrutica.com/companies/org-synopsys.

### Ansys (`org-ansys`)

Provides semiconductor simulation software for power, thermal and signal integrity. These analyses cover interactions among chips, packages and circuit boards. Sources: [1](https://www.ansys.com/en-gb/webinars/comprehensive-multiphysics-analysis-platform-3d-ic). Headquartered in the United States. Detail page: https://scrutica.com/companies/org-ansys.

### GlobalFoundries (`org-global-foundries`)

Provides semiconductor manufacturing and design support, with process technologies for CMOS devices and advanced packaging services that integrate multiple components. Sources: [1](https://gf.com/technologies/), [2](https://gf.com/technologies/cmos/). Headquartered in the United States. Public-market ticker: GFS. Detail page: https://scrutica.com/companies/org-global-foundries.

### Canon (`org-canon`)

Makes semiconductor lithography equipment that transfers circuit patterns onto wafers or panels. Its portfolio includes optical systems and nanoimprint equipment, which forms patterns with a mold. Sources: [1](https://global.canon/en/product/indtech/semicon/index.html), [2](https://global.canon/en/technology/canon-tech/tech/semicon/). Headquartered in Japan. Detail page: https://scrutica.com/companies/org-canon.

### Nikon (`org-nikon`)

Makes semiconductor lithography systems, including ArF immersion scanners, and equipment for alignment, inspection and advanced packaging. Sources: [1](https://www.nikon.com/business/semi/lineup/). Headquartered in Japan. Detail page: https://scrutica.com/companies/org-nikon.

### ASE (`org-ase`)

Provides semiconductor assembly, advanced packaging and testing services. Its tests cover wafers, packaged chips and complete systems. Sources: [1](https://ase.aseglobal.com/press-room/310x310/), [2](https://ase.aseglobal.com/test-services/). Headquartered in Taiwan. Detail page: https://scrutica.com/companies/org-ase.

### Amkor (`org-amkor`)

Provides outsourced semiconductor packaging and testing, including tests at wafer and package level. Sources: [1](https://amkor.com/), [2](https://amkor.com/test-services/). Headquartered in the United States. Detail page: https://scrutica.com/companies/org-amkor.

### Advantest (`org-advantest`)

Makes automated test equipment for semiconductor devices, including systems for memory and system-on-chip testing during design evaluation and production. Sources: [1](https://www.advantest.com/en/products/semiconductor-test-system/), [2](https://www.advantest.com/en/products/semiconductor-test-system/memory/). Headquartered in Japan. Detail page: https://scrutica.com/companies/org-advantest.

### Teradyne (`org-teradyne`)

Makes semiconductor test equipment for integrated circuits, memory, system-on-chip and system-in-package devices. Sources: [1](https://www.teradyne.com/semiconductor-testing/). Headquartered in the United States. Detail page: https://scrutica.com/companies/org-teradyne.

### Besi (`org-besi`)

Makes die-attach and hybrid-bonding equipment for semiconductor packaging. Its hybrid-bonding systems align and place components for dense interconnects and three-dimensional integration. Sources: [1](https://www.besi.com/products-technology/product-details/productgroup/die-attach/), [2](https://www.besi.com/products-technology/product-details/productgroup/hybrid-bonding/). Headquartered in the Netherlands. Detail page: https://scrutica.com/companies/org-besi.

### ASM International (`org-asm-international`)

Makes atomic-layer-deposition and epitaxy equipment. These processes form thin material layers and crystalline films used in semiconductor devices. Sources: [1](https://www.asm.com/our-technology-products/ald), [2](https://www.asm.com/our-technology-products/epitaxy). Headquartered in the Netherlands. Detail page: https://scrutica.com/companies/org-asm-international.

### JSR Corporation (`org-jsr-corporation`)

Supplies semiconductor materials, including photoresists for optical and EUV lithography and multilayer materials used with them. Sources: [1](https://www.jsr.co.jp/jsr_e/products/em/). Headquartered in Japan. Detail page: https://scrutica.com/companies/org-jsr-corporation.

### Tokyo Ohka Kogyo (`org-tokyo-ohka`)

Makes photoresists and high-purity chemicals for semiconductor and display photolithography, including EUV photoresists. Sources: [1](https://www.tok.co.jp/eng/company/outline), [2](https://www.tok.co.jp/eng/sustainability/social-activity/quality). Headquartered in Japan. Detail page: https://scrutica.com/companies/org-tokyo-ohka.

### Shin-Etsu (`org-shin-etsu`)

Supplies silicon wafers and photoresists for semiconductor manufacturing. The wafer provides the material on which devices are fabricated; the photosensitive resist transfers a circuit pattern during lithography. Sources: [1](https://www.shinetsu.co.jp/en/products/electronics-materials/silicon-wafers/), [2](https://www.shinetsu.co.jp/en/products/electronics-materials/photoresist/). Headquartered in Japan. Detail page: https://scrutica.com/companies/org-shin-etsu.

### SUMCO (`org-sumco`)

Manufactures silicon wafers for semiconductor devices, including polished wafers and epitaxial wafers with an added single-crystal silicon layer. Sources: [1](https://www.sumcosi.com/english/products/lineup.html). Headquartered in Japan. Detail page: https://scrutica.com/companies/org-sumco.

### Air Liquide (`org-air-liquide`)

Supplies high-purity gases and precursor materials for semiconductor manufacturing, together with equipment for their distribution, purification and handling. Sources: [1](https://electronics.airliquide.com/who-we-are/our-markets/semiconductors-0). Headquartered in France. Detail page: https://scrutica.com/companies/org-air-liquide.

### Linde (`org-linde`)

Supplies bulk and specialty gases for semiconductor manufacturing, including on-site gas production and delivery to fabrication plants. Sources: [1](https://assets.linde.com/-/media/celum-connect/2023/12/21/15/11/linde20electronics20brochure185963.pdf), [2](https://www.linde.com/news-and-media/2025/linde-to-expand-supply-of-industrial-gases-to-samsung-in-south-korea). Headquartered in Germany. Detail page: https://scrutica.com/companies/org-linde.

### Cerebras (`org-cerebras`)

Builds AI computing systems around its Wafer Scale Engine. The WSE-3 processor powers the CS-3 system used for model training and inference. Sources: [1](https://www.cerebras.ai/press-release/cerebras-announces-third-generation-wafer-scale-engine), [2](https://www.cerebras.ai/press-release/cerebras-and-core42-deliver-record-breaking-performance-for-openai-s-gpt-oss-120b-powering-ai). Headquartered in the United States. Detail page: https://scrutica.com/companies/org-cerebras.

### Groq (`org-groq`)

Provides inference computing through its Language Processing Unit and GroqCloud service. Its compiler distributes model operations across processors that store their assigned data in local SRAM. Sources: [1](https://groq.com/blog/the-groq-lpu-explained), [2](https://groq.com/wp-content/uploads/2024/05/GroqThoughts_OnePager-Power-VF.pdf). Headquartered in the United States. Detail page: https://scrutica.com/companies/org-groq.

### SambaNova Systems (`org-sambanova`)

Builds AI systems around its Reconfigurable Dataflow Unit. Programmable compute and memory units map model operations onto the hardware so data can pass between successive operations. Sources: [1](https://sambanova.ai/products/dataflow-architecture). Headquartered in the United States. Detail page: https://scrutica.com/companies/org-sambanova.

### Tenstorrent (`org-tenstorrent`)

Develops AI processors and RISC-V CPU technology. Its open-source TT-Metalium software gives developers access to the processors, on-chip network, and matrix and vector engines within a Tensix core. Sources: [1](https://tenstorrent.com/software/tt-metalium), [2](https://tenstorrent.com/en/newsroom/the-open-hardware-revolution). Headquartered in the United States. Detail page: https://scrutica.com/companies/org-tenstorrent.

### Graphcore (`org-graphcore`)

Developed the Intelligence Processing Unit, a parallel processor for machine learning. IPU systems connect to a host computer and use the Poplar software environment to program workloads across processors. Sources: [1](https://docs.graphcore.ai/projects/ipu-programmers-guide/en/3.3.0/ipu_introduction.html), [2](https://www.graphcore.ai/products/ipu). Headquartered in the United Kingdom. Detail page: https://scrutica.com/companies/org-graphcore.

### HiSilicon (`org-huawei-hisilicon`)

Designs chipsets for devices and media processing. Its Kirin system-on-chip designs integrate CPUs, graphics and neural-processing hardware for on-device applications. Sources: [1](https://www.hisilicon.com/en/products/Kirin/Kirin-flagship-chips/Kirin-9000), [2](https://www.hisilicon.com/en/products/smart-media). Headquartered in China. Detail page: https://scrutica.com/companies/org-huawei-hisilicon.

### Cambricon (`org-cambricon`)

Develops AI acceleration cards for cloud computing, edge-computing modules and processor IP for devices. Its software products include NeuWare and MagicMind. Sources: [1](https://www.cambricon.com/). Headquartered in China. Detail page: https://scrutica.com/companies/org-cambricon.

### ByteDance (`org-bytedance`)

Develops AI models through its Seed research teams. Its infrastructure work covers distributed training, reinforcement-learning frameworks, inference and compilation for different hardware architectures. Sources: [1](https://seed.bytedance.com/en/direction/llm), [2](https://seed.bytedance.com/direction/foundation). Headquartered in China. Detail page: https://scrutica.com/companies/org-bytedance.

### Alibaba (`org-alibaba`)

Provides cloud computing through Alibaba Cloud, including GPU instances for model training and inference and Arm-based instances using Yitian 710 processors. Sources: [1](https://www.alibabacloud.com/help/en/ecs/user-guide/instance-families/), [2](https://www.alibabacloud.com/help/en/ecs/user-guide/gpu-accelerated-compute-optimized-and-vgpu-accelerated-instance-families-1). Headquartered in China. Public-market ticker: BABA. Detail page: https://scrutica.com/companies/org-alibaba.

### Tencent Holdings Ltd. (`org-tencent`)

Develops the Tencent Hy model family, previously called Hunyuan, and distributes model access through Tencent Cloud TokenHub and its applications. Sources: [1](https://www.tencent.com/tencent-hy3-now-available-globally-extending-practical-ai-across-products-workflows-and-cloud-services/). Headquartered in China. Public-market ticker: 700. Detail page: https://scrutica.com/companies/org-tencent.

### IREN (`org-iren`)

Offers GPU cloud capacity, colocation and data-center development services for AI workloads. Its colocation service accommodates customers who supply their own hardware. Sources: [1](https://www.iren.com/), [2](https://iren.com/solutions/gpu-cloud/ai-cloud/). Headquartered in Australia. Public-market ticker: IREN. Detail page: https://scrutica.com/companies/org-iren.

### Applied Digital (`org-applied-digital`)

Develops and operates data-center campuses for AI and high-performance computing, including power and cooling infrastructure. Sources: [1](https://www.applieddigital.com/digital-infrastructure). Headquartered in the United States. Detail page: https://scrutica.com/companies/org-applied-digital.

### Crusoe (`org-crusoe`)

Develops data centers and provides GPU computing through Crusoe Cloud for AI training and inference. Sources: [1](https://www.crusoe.ai/resources/newsroom/crusoes-contracted-ai-infrastructure-capacity-approaches-5-gigawatts-across-data-centers-and-cloud), [2](https://crusoe.ai/cloud/). Headquartered in the United States. Detail page: https://scrutica.com/companies/org-crusoe.

### Nebius (`org-nebius`)

Provides GPU cloud infrastructure, from individual virtual machines to multi-node clusters, with networking and orchestration services for AI workloads. Sources: [1](https://nebius.com/compute), [2](https://docs.nebius.com/). Headquartered in the Netherlands. Public-market ticker: NBIS. Detail page: https://scrutica.com/companies/org-nebius.

### Lambda Labs (`org-lambda-labs`)

Provides GPU instances and interconnected clusters for model training, fine-tuning and deployment, with both on-demand access and longer-term capacity contracts. Sources: [1](https://lambda.ai/welcome/ai-cloud). Headquartered in the United States. Detail page: https://scrutica.com/companies/org-lambda-labs.

### Together AI (`org-together-ai`)

Provides GPU clusters for model training and inference, with managed orchestration, networking and shared storage. Sources: [1](https://www.together.ai/gpu-clusters), [2](https://www.together.ai/accelerated-compute). Headquartered in the United States. Detail page: https://scrutica.com/companies/org-together-ai.

### Cohere (`org-cohere`)

Develops the Command language-model family for enterprise applications, alongside Embed and Rerank models for search and retrieval. Sources: [1](https://docs.cohere.com/docs/models). Detail page: https://scrutica.com/companies/org-cohere.

### Inflection AI (`org-inflection-ai`)

Develops models for the Pi assistant and for instruction-following applications, with access documented through the Inflection API. Sources: [1](https://developers.inflection.ai/docs/introduction). Headquartered in the United States. Detail page: https://scrutica.com/companies/org-inflection-ai.

### Super Micro Computer, Inc. (`org-supermicro`)

Builds GPU servers for AI training and inference, with configurations that combine accelerators, host CPUs and networking. Sources: [1](https://www.supermicro.com/en/accelerators/nvidia/ai-factory), [2](https://www.supermicro.com/en/solutions/ngc). Headquartered in the United States. Public-market ticker: SMCI. Detail page: https://scrutica.com/companies/org-supermicro.

### Dell Technologies (`org-dell`)

Supplies PowerEdge GPU servers and rack systems for AI training, inference and high-performance computing. Configurations integrate NVIDIA or AMD accelerators with host processors, memory and storage. Sources: [1](https://www.dell.com/en-us/shop/servers/sf/poweredge-ai-servers). Headquartered in the United States. Public-market ticker: DELL. Detail page: https://scrutica.com/companies/org-dell.

### HPE (`org-hpe`)

Builds servers and Cray supercomputing systems for scientific computing and AI workloads. These systems integrate processors and accelerators with high-speed networking and cooling. Sources: [1](https://www.hpe.com/us/en/newsroom/press-release/2026/03/hpe-unveils-next-generation-ai-factory-and-supercomputing-advancements-with-nvidia.html). Headquartered in the United States. Public-market ticker: HPE. Detail page: https://scrutica.com/companies/org-hpe.

### Broadcom Inc (`org-broadcom`)

Designs custom AI accelerators and networking silicon for data centers, including Ethernet switching and connectivity products that link accelerators within computing clusters. Sources: [1](https://www.broadcom.com/solutions/ai-solutions/ai-infrastructure). Headquartered in the United States. Public-market ticker: AVGO. Detail page: https://scrutica.com/companies/org-broadcom.

### Marvell Technology (`org-marvell`)

Designs custom chips for cloud and AI systems and supplies networking and interconnect technology. Its custom-ASIC portfolio includes compute, memory, storage and high-speed-interface IP. Sources: [1](https://www.marvell.com/products/custom-asic.html), [2](https://www.marvell.com/ai.html). Headquartered in the United States. Public-market ticker: MRVL. Detail page: https://scrutica.com/companies/org-marvell.

### MediaTek (`org-mediatek`)

Designs system-on-chip products and co-designed the GB10 Grace Blackwell Superchip with NVIDIA for DGX Spark. Its contribution includes CPU, memory-subsystem and high-speed-interface design. Sources: [1](https://www.mediatek.com/tek-talk-blogs/designing-a-superchip-inside-mediatek-nvidias-collaboration-on-dgx-spark). Headquartered in Taiwan. Detail page: https://scrutica.com/companies/org-mediatek.

### Qualcomm Inc (`org-qualcomm`)

Designs processors and AI accelerators for devices and data centers. Its Hexagon neural-processing unit supports on-device inference; its data-center portfolio includes dedicated inference accelerators. Sources: [1](https://www.qualcomm.com/processors/hexagon), [2](https://www.qualcomm.com/data-center/expertise/ai-accelerators). Headquartered in the United States. Public-market ticker: QCOM. Detail page: https://scrutica.com/companies/org-qualcomm.

### Cloudflare, Inc. (`org-cloudflare`)

Provides serverless model inference through Workers AI, which runs models on GPUs across Cloudflare’s network and exposes them through application code and APIs. Sources: [1](https://developers.cloudflare.com/workers-ai/). Headquartered in the United States. Public-market ticker: NET. Detail page: https://scrutica.com/companies/org-cloudflare.

---


## Sovereign AI program records

Program records are ordered by headline announced funding. The figures distinguish funding stages; notes preserve qualifications about source scope and accounting. A status label can describe a program's progress without establishing that its computing capacity is operating. The [interactive register](https://scrutica.com/sovereign-ai) provides the source details.

Quantity listings include entries with a recorded count. Programme and partner quantities can overlap and include time-limited access.

### United States: NAIRR / CHIPS Act AI Components / Stargate Project

Programme record and sources: https://scrutica.com/sovereign-ai/program/sovereign-us

- Country: `us` · Status: operational (NAIRR Pilot) / legislation_pending (full NAIRR)
- Announced (headline): ~$2.60B
- Government funding: $30.0M
- Committed: $30.0M
- Timeline: 2024–?
- Interdependence: NVIDIA: high · US: low
- Key partners: NSF, DOE, Microsoft, NVIDIA, Google, Amazon, Meta
- Cited source URLs: 16

NAIRR gives researchers access to university and federal computing facilities, commercial cloud services, and AI tools. The $2.6B figure comes from the task force’s six-year cost estimate; the FY2026 appropriations agreement provides $30M for the pilot. Private Stargate investments and CHIPS semiconductor funding are separate from this research-access programme.

#### Funding and source notes

- The $30M is FY2026 funding for the NAIRR pilot. The cited appropriations agreement reports no amount paid out, so disbursement is unknown.
- The NAIRR task force estimated a six-year cost of approximately $2.6B in January 2023. Comparing one year’s pilot funding with that estimate does not measure completion of the full programme.
- CSET estimated 3.77 exaFLOPS for the pilot’s first six computing resources under full hardware utilisation. Researchers receive time-limited allocations on those systems; the estimate is neither dedicated year-round capacity nor a current inventory.
- OpenAI’s January 2025 Stargate announcement proposed $500B of private investment over four years in US AI infrastructure. CHIPS funding supports semiconductor manufacturing and research; neither amount is included in NAIRR’s government funding here.
- OpenAI reported a 1 June 2026 groundbreaking for The Barn, its 1 GW Stargate campus in Saline, Michigan. Related Digital identifies project financing from private equity and debt.
- The 208-GH200 allocation previously attributed to Purdue Anvil is unverified. Purdue’s NAIRR guide identifies A100 and H100 resources.

#### Programme quantities

- NVIDIA H100 SXM: 84. Status: operational. Physical hardware. Anvil AI’s H100 expansion: 21 servers with four GPUs each.
  [Count source](<https://www.rcac.purdue.edu/news/7345>) (authority tier 1; published 2025-08-22).
- NVIDIA H100: 136. Status: operational. Physical hardware. Expanse AI’s H100 partition: 34 servers with four GPUs each.
  [Count source](<https://www.sdsc.edu/systems/expanse/user_guide.html>) (authority tier 1; publication date not recorded).

### Spain: BSC AI Factory + MareNostrum 5 + AI Strategy 2024

Programme record and sources: https://scrutica.com/sovereign-ai/program/sovereign-es

- Country: `es` · Status: operational (MareNostrum 5, BSC AI Factory) / procurement_active (EUR 129M expansion)
- Announced (headline): ~$1.62B
- Government funding: ~$1.62B
- Timeline: 2024–2025
- Interdependence: NVIDIA: medium · US: medium
- Key partners: BSC-CNS, EuroHPC JU, Spanish Government, Catalan Government, FSAS Technologies, Telefonica, Fujitsu
- Cited source URLs: 12

BSC provides research and industry access to MareNostrum 5; its accelerated partition has 4,480 NVIDIA H100 GPUs. The BSC AI Factory has offered computing access and technical support since April 2025. ALIA uses MareNostrum 5 to train public language models for Spanish and its co-official languages. A €129M upgrade contract signed in January 2026 covers additional AI capacity.

#### Funding and source notes

- MareNostrum 5’s accelerated partition comprises 1,120 nodes with four NVIDIA Hopper GPUs each. Its general-purpose partition uses Intel CPUs.
- The June 2022 MareNostrum 5 procurement budget was €151.41M: half from EuroHPC and half from a consortium of Spain, Portugal and Türkiye. The published budget covers procurement; payments to suppliers are not itemized.
- The January 2026 AI upgrade contract covers acquisition, delivery, installation and maintenance for €129M. EuroHPC supplies half of the funding; Spain, Portugal and Türkiye supply the remainder.
- Spain’s AI Strategy 2024 budgeted €1.5B for 2024–25, including €90M for MareNostrum 5 and the Spanish Supercomputing Network. That €90M overlaps the strategy total. Its relationship to the later AI Factory contract is not reconciled.
- The dollar figure converts the €1.5B strategy budget at US$1.0818 per euro on 14 May 2024, BSC’s publication date, from the Federal Reserve’s H.10 series distributed by FRED. The three procurement components below have their own dates and carry no dollar equivalent.

#### Programme quantities

- NVIDIA H100: 4,480. Status: operational. Physical hardware. MareNostrum 5 accelerated partition: 1,120 nodes with four GPUs each.
  [Count source](<https://www.bsc.es/marenostrum/marenostrum-5>) (authority tier 1; publication date not recorded).

### India: IndiaAI Mission

Programme record and sources: https://scrutica.com/sovereign-ai/program/sovereign-in

- Country: `in` · Status: operational
- Announced (headline): ~$1.25B
- Government funding: ~$1.25B
- Committed: ~$1.25B
- Timeline: 2024–2029
- Interdependence: NVIDIA: high · US: high
- Key partners: MeitY, Digital India Corporation, NVIDIA, Google, G42, L&T, Yotta
- Cited source URLs: 20

IndiaAI pays for access to computing through empanelled service providers and supports datasets, models, training, and startup finance. The Cabinet approved ₹10,371.92 crore over five years. In October 2025, PIB reported more than 38,000 GPUs onboarded for shared access; it gave no government-owned fleet count or vendor breakdown.

#### Funding and source notes

- The Cabinet’s ₹10,371.92 crore approval covers the whole IndiaAI Mission. It is not a GPU-purchase budget.
- The compute programme uses a public-private partnership. The government’s aggregate GPU count does not establish how many devices it owns or how much capacity each user can reserve.
- Partner records span tender rounds, bids, and installation reports. They also include Google TPUs and AWS Trainium and Inferentia accelerators, so summing their counts does not reconstruct the government’s GPU total.
- The August 2025 third-tender report includes 1,050 Trillium TPUs within 3,850 additions to the existing 34,333-unit pool. Those TPUs are not added to the government’s later rounded 38,000 count; the sources do not establish a separate GPU-only total.
- IndiaAI received ₹21.79 crore in 2024–25 and ₹379.15 crore in 2025–26 through 9 February 2026, according to the parliamentary answer of 13 February.
- A separate parliamentary report records actual utilisation of ₹19.24 crore in 2024–25 and ₹256.86 crore in 2025–26 through 31 December 2025. These are expenditure figures at an earlier date than the reported fund transfers.
- The dollar estimate converts the Cabinet’s rupee allocation using 82.78 INR per USD on 7 March 2024, from the Federal Reserve’s H.10 series distributed by FRED. The displayed $1.25B is rounded.
- G42 and Cerebras announced an 8-exaFLOP system for India in February 2026. The announcement describes a future deployment and does not specify the precision needed to compare that performance with other systems.
- NVIDIA describes private expansion plans including L&T’s 30 MW Chennai and 40 MW Mumbai facilities and Yotta’s Blackwell Ultra cloud. These plans are separate from the mission’s government allocation.
- Adani and Jabil announced their intention to form an AI-hardware manufacturing alliance in June 2026, with formal documentation still in preparation. The release gives no alliance-specific investment amount or operating compute capacity.

#### Programme quantities

- GPUs (vendor breakdown undisclosed): >38,000. Status: Onboarded for shared access (PIB, October 2025). Shared access. Accelerators onboarded for subsidised shared access as of October 2025; vendor composition is unspecified.
  [Count source](<https://www.pib.gov.in/Pressreleaseshare.aspx?PRID=2178092&lang=2&reg=48>) (authority tier 1; published 2025-10-12).

#### Partner quantities

- Yotta Data Services (NVIDIA H100 + NVIDIA B200 Blackwell): 15,080. Participation stage: empanelled. Shared access. Yotta’s June 2025 allocations: 6,888 H100 plus 8,192 B200; delivery was incomplete.
  [Count source](<https://indianinfrastructure.com/2025/06/24/more-than-17000-gpus-installed-under-indiaai-mission-in-the-first-two-bidding-rounds/>) (authority tier 3; published 2025-06-24).
- Jio Platforms (NVIDIA H200 + AMD MI300X): 1,020. Participation stage: empanelled. Shared access. Jio’s June 2025 procurement: 752 H200 plus 268 MI300X accelerators.
  [Count source](<https://indianinfrastructure.com/2025/06/24/more-than-17000-gpus-installed-under-indiaai-mission-in-the-first-two-bidding-rounds/>) (authority tier 3; published 2025-06-24).
- Locuz Enterprise Solutions (NVIDIA H100): 1,300. Participation stage: empanelled. Shared access. Locuz’s additional third-tender H100 offer in August 2025.
  [Count source](<https://inc42.com/buzz/indiaai-to-add-over-3800-gpus-via-third-tender-report/>) (authority tier 3; published 2025-08-18).
- Vensysco Technologies (NVIDIA H100 + AWS Trainium 1 + AWS Inferentia 2): 2,300. Participation stage: empanelled. Shared access. Vensysco’s second-round bid: 2,000 H100, 100 Trainium1 and 200 Inferentia2 accelerators.
  [Count source](<https://egov.eletsonline.com/2025/05/indiaai-mission-receives-bids-for-18000-gpus-in-second-tender-phase/>) (authority tier 3; published 2025-05-15).
- Cyfuture India (NVIDIA H100 + L40S + A100, AMD MI300 + MI325, Intel Gaudi 2 + Gaudi 3): 1,184. Participation stage: shortlisted. Shared access. Cyfuture’s mixed-accelerator bid in the May 2025 second tender.
  [Count source](<https://egov.eletsonline.com/2025/05/indiaai-mission-receives-bids-for-18000-gpus-in-second-tender-phase/>) (authority tier 3; published 2025-05-15).
- Sify Digital Services (Google Trillium TPU + NVIDIA H200 + NVIDIA L4): 2,500. Participation stage: empanelled. Shared access. Sify’s third-tender offer: 1,000 Trillium TPUs, 800 H200 and 700 L4 GPUs.
  [Count source](<https://inc42.com/buzz/indiaai-to-add-over-3800-gpus-via-third-tender-report/>) (authority tier 3; published 2025-08-18).
- Ishan Infotech (Google Trillium TPU): 50. Participation stage: empanelled. Shared access. Ishan’s third-tender offer of 50 Trillium TPUs in August 2025.
  [Count source](<https://inc42.com/buzz/indiaai-to-add-over-3800-gpus-via-third-tender-report/>) (authority tier 3; published 2025-08-18).

### Singapore: National AI Strategy 2.0 / AISG / NSCC Sovereign Compute

Programme record and sources: https://scrutica.com/sovereign-ai/program/sovereign-sg

- Country: `sg` · Status: operational (ASPIRE 2A, 2A+ and 2B)
- Announced (headline): ~$744.0M
- Government funding: ~$744.0M
- Committed: ~$204.0M
- Timeline: 2023–2030
- Interdependence: NVIDIA: high · US: high
- Key partners: NRF, NSCC Singapore, NVIDIA, MDDI, AISG
- Cited source URLs: 12

NSCC launched ASPIRE 2A and 2A+ in October 2024, with 352 A100 and 320 H100 GPUs respectively. ASPIRE 2B launched in June 2026 with 1,536 H200 GPUs; NSCC included it in July resource allocations. The S$270 million grant supports computing infrastructure, research access and skills. Singapore's January 2026 National AI R&D Plan commits more than S$1 billion for 2025–2030. The Enterprise Compute Initiative provides cloud and consultancy support. IMDA's 2024 roadmap targets at least 300 MW of additional data-centre capacity.

#### Funding and source notes

- The S$270M grant covers infrastructure, research access and skills. Its announcement gives no disbursement figure.
- The 2024 five-year AI package totals S$1B, reported as approximately US$744M by Computer Weekly on 19 February 2024. The January 2026 National AI R&D Plan adds more than S$1B for 2025–2030. Their funding periods overlap; the announcements do not itemise the NSCC grant’s relationship to both packages.
- The NVIDIA partner allocation retains the 2024 ASPIRE 2A/2A+ scope: 672 GPUs. The programme hardware entries also include ASPIRE 2B.
- IMDA’s 300 MW figure is the additional-capacity target in its May 2024 roadmap; the cited statement does not establish an allocation.
- The committed dollar figure converts the S$270M grant at 1.3206 Singapore dollars per US dollar on 25 October 2024, the announcement’s date, from the Federal Reserve’s H.10 series distributed by FRED. The announced figure is Computer Weekly’s own US$744M, so this record carries no single conversion rate.

#### Programme quantities

- NVIDIA H100 SXM5: 320. Status: operational. Physical hardware. ASPIRE 2A+ at its October 2024 launch: 40 DGX H100 systems with eight GPUs each.
  [Count source](<https://www.nscc.sg/wp-content/uploads/2024/10/Embargoed-till-25-Oct-2024-4pm-Media-Release-for-NSCCs-ASPIRE-2A-2A-Launch.pdf>) (authority tier 1; published 2024-10-25).
- NVIDIA A100: 352. Status: operational. Physical hardware. ASPIRE 2A’s A100 partition at its October 2024 launch.
  [Count source](<https://www.nscc.sg/wp-content/uploads/2024/10/Embargoed-till-25-Oct-2024-4pm-Media-Release-for-NSCCs-ASPIRE-2A-2A-Launch.pdf>) (authority tier 1; published 2024-10-25).
- NVIDIA H200: 1,536. Status: operational. Physical hardware. ASPIRE 2B’s H200 partition.
  [Count source](<https://www.nscc.sg/aspire-2b/>) (authority tier 1; publication date not recorded).

#### Partner quantities

- NVIDIA (NVIDIA A100 (ASPIRE 2A) + NVIDIA H100 DGX SuperPOD (ASPIRE 2A+)): 672. Participation stage: deployed. Physical hardware. ASPIRE 2A and 2A+ at their October 2024 launch: 352 A100 plus 320 H100 GPUs.
  [Count source](<https://www.nscc.sg/wp-content/uploads/2024/10/Embargoed-till-25-Oct-2024-4pm-Media-Release-for-NSCCs-ASPIRE-2A-2A-Launch.pdf>) (authority tier 1; published 2024-10-25).

### Mexico: Coatlicue Supercomputer + National AI Plan

Programme record and sources: https://scrutica.com/sovereign-ai/program/sovereign-mx

- Country: `mx` · Status: Site and technical committee established by June 2026; operational completion unverified
- Announced (headline): ~$330.0M
- Government funding: ~$330.0M
- Timeline: 2026–2028
- Interdependence: NVIDIA: unknown · US: high
- Key partners: Barcelona Supercomputing Center (BSC)
- Cited source URLs: 10

Coatlicue is a planned public supercomputer for research and government services. Secihti’s November 2025 announcement specified 14,480 GPUs, an estimated 314 PFLOPS and approximately MXN 6B in investment. A June 2026 agreement placed it at IPN’s Zacatenco campus and established a technical committee led by Secihti and ATDT. Mexico’s research unit at Barcelona Supercomputing Center is working on meteorology, medical appointment scheduling and tax-fraud detection. The May 2026 presidential update said Coatlicue’s financing model was still being developed, including a special fund and paid services for commercial users.

#### Funding and source notes

- The 14,480 GPUs and 314 PFLOPS are design targets from November 2025. The announcement does not specify an accelerator supplier or floating-point precision.
- The original 24-month project timetable does not establish construction progress or a count of installed GPUs. The June 2026 agreement confirms the site and governance arrangements.
- The US$330M estimate is Aztec Reports’ approximation of the planned MXN 6B public investment (19 December 2025).
- The May 6, 2026 presidential statement describes financing still under development. No committed or disbursed amount is established by the cited project announcements.
- OECD’s National AI Agenda entry describes a separate 2024–2030 policy proposal developed through ANIA and UNESCO consultations.

### Netherlands: Dutch AI Factory (AIFNL) + Snellius/EuroHPC

Programme record and sources: https://scrutica.com/sovereign-ai/program/sovereign-nl

- Country: `nl` · Status: AI Factory procurement; supercomputer target 2027
- Announced (headline): ~$232.0M
- Government funding: ~$151.0M
- Committed: ~$232.0M
- Timeline: 2025–2027
- Interdependence: NVIDIA: high · US: medium
- Key partners: AIFNL Foundation, SURF, TNO, EuroHPC JU, Samenwerking Noord
- Cited source URLs: 8

The Dutch AI Factory will provide computing and technical support for AI research and development in Groningen. Its €200M budget combines €70M from EuroHPC, €70M from the national government and €60M from Groningen/North Drenthe. The expertise centre was scheduled to begin in 2026 and the supercomputer to reach full capacity in 2027. Dutch users can apply for time on other EuroHPC systems while it is built. SURF also operates the existing Snellius system, whose GPU partitions contain 288 A100 and 352 H100 accelerators.

#### Funding and source notes

- The NLAIF hardware tender opened on 11 March 2026 and closed on 30 April. It names EuroHPC as owner and Stichting NLAIF as the hosting consortium. The 2027 availability date is a target in the funding announcement.
- Snellius’s 640 GPUs come from SURF’s hardware inventory: 72 nodes with four A100s each and 88 nodes with four H100s each. These are installed Snellius accelerators; the planned Groningen machine has a separate procurement.
- The dollar figures convert the €200M project budget and the €130M Dutch public contribution at US$1.1613 per euro on 10 October 2025, SURF’s announcement date, from the Federal Reserve’s H.10 series distributed by FRED.

#### Programme quantities

- NVIDIA A100: 288. Status: operational. Physical hardware. Snellius’ A100 partition: 72 nodes with four GPUs each.
  [Count source](<https://servicedesk.surf.nl/wiki/spaces/WIKI/pages/30660208/Snellius+hardware>) (authority tier 1; publication date not recorded).
- NVIDIA H100 SXM5: 352. Status: operational. Physical hardware. Snellius’ H100 partition: 88 nodes with four GPUs each.
  [Count source](<https://servicedesk.surf.nl/wiki/spaces/WIKI/pages/30660208/Snellius+hardware>) (authority tier 1; publication date not recorded).

### Switzerland: Alps Supercomputer + Swiss AI Initiative

Programme record and sources: https://scrutica.com/sovereign-ai/program/sovereign-ch

- Country: `ch` · Status: Alps operational; Apertus released September 2025
- Announced (headline): ~$118.0M
- Government funding: ~$118.0M
- Timeline: 2024–?
- Interdependence: NVIDIA: high · US: high
- Key partners: CSCS, ETH Zurich, EPFL, HPE, NVIDIA
- Cited source URLs: 11

CSCS operates Alps in Lugano, with 10,752 NVIDIA Grace Hopper superchips reported at its September 2024 inauguration. The system supports the Swiss AI Initiative, led by ETH Zurich and EPFL. Their multilingual Apertus models were released on 2 September 2025 for public download. Alps also serves scientific workloads, including weather forecasting for MeteoSwiss and research at the Paul Scherrer Institute.

#### Funding and source notes

- The GPU count is the 10,752 Grace Hopper superchips reported by ETH Zurich at the Alps inauguration on 14 September 2024.
- SWI swissinfo reported Alps’ capital cost as CHF100M (approximately US$118M) on 3 October 2024, funded from the public purse.
- The ETH Board allocated CHF 20M to Swiss AI Initiative research for 2025–2028.

#### Programme quantities

- NVIDIA Grace Hopper GH200: 10,752. Status: operational. Physical hardware. Alps’ Grace Hopper superchips at its September 2024 inauguration.
  [Count source](<https://ethz.ch/en/news-and-events/eth-news/news/2024/09/press-release-new-research-infrastructure-alps-supercomputer-inaugurated.html>) (authority tier 1; published 2024-09-14).

#### Partner quantities

- Swiss National Supercomputing Centre (CSCS) (NVIDIA GH200 Grace Hopper Superchip): 10,752. Participation stage: commissioned. Physical hardware. Alps’ Grace Hopper superchips at its September 2024 inauguration.
  [Count source](<https://ethz.ch/en/news-and-events/eth-news/news/2024/09/press-release-new-research-infrastructure-alps-supercomputer-inaugurated.html>) (authority tier 1; published 2024-09-14).

### Czech Republic: CZAI AI Factory + IT4Innovations + KarolAIna

Programme record and sources: https://scrutica.com/sovereign-ai/program/sovereign-cz

- Country: `cz` · Status: CZAI selected October 2025; KarolAIna planned
- Announced (headline): ~$46.5M
- Government funding: ~$23.2M
- Committed: ~$46.5M
- Timeline: 2025–2027
- Interdependence: NVIDIA: unknown · US: high
- Key partners: VSB-Technical University of Ostrava, IT4Innovations, VUT Brno, Charles University, CVUT, EuroHPC JU
- Cited source URLs: 6

CZAI will combine the KarolAIna supercomputer with technical support for companies, public bodies and researchers. IT4Innovations announced a design of about 340 AI chips and 850 PFlop/s of AI performance; the announcement does not specify the chip model or numerical precision. The €40M project is funded equally by EuroHPC and Czechia. About half covers the machine’s procurement and operation, with the rest for services. Planned campuses in Prague and Brno will provide contact points for students and young professionals.

#### Funding and source notes

- EuroHPC selected CZAI in its third selection round on 10 October 2025. The approximately 340 chips are a planned specification in the IT4Innovations announcement.
- EuroHPC’s half of the funding provides access to half of KarolAIna’s computing time and related AI services. The €40M budget includes both computing infrastructure and services.
- The dollar figures convert the approximately €40M budget and the Czech half of it at US$1.1613 per euro on 10 October 2025, the announcement’s publication date, from the Federal Reserve’s H.10 series distributed by FRED.

#### Programme quantities

- AI accelerators (model undisclosed): ~340. Status: procurement pending. Physical hardware. Whole planned KarolAIna machine; EuroHPC receives half its computing time and services.
  [Count source](<https://www.it4i.cz/o-it4i/infoservis/novinky/cesko-ziskalo-svou-tovarnu-na-umelou-inteligenci-a-s-ni-i-novy-ai-superpocitac>) (authority tier 1; published 2025-10-10).

#### Partner quantities

- VSB – Technical University of Ostrava (model not recorded): ~340. Participation stage: empanelled. Physical hardware. Whole planned KarolAIna machine; EuroHPC receives half its computing time and services.
  [Count source](<https://www.it4i.cz/o-it4i/infoservis/novinky/cesko-ziskalo-svou-tovarnu-na-umelou-inteligenci-a-s-ni-i-novy-ai-superpocitac>) (authority tier 1; published 2025-10-10).

### UAE: Stargate UAE / G42 / Core42 / TII Sovereign AI + Microsoft Cumulative + Abu Dhabi Digital Strategy

Programme record and sources: https://scrutica.com/sovereign-ai/program/sovereign-ae

- Country: `ae` · Status: procurement_active / construction_active / Oracle Supercluster operational (Nov 2025)
- Timeline: 2024–2029
- Interdependence: NVIDIA: high · US: high
- Key partners: Microsoft, OpenAI, NVIDIA, Oracle, Cisco, SoftBank, Technology Innovation Institute (TII), G42, Core42, Khazna Data Centers, Abu Dhabi Department of Government Enablement (DGE)
- Cited source URLs: 14

Stargate UAE is a planned 1 GW cluster within the larger 5 GW UAE–US AI Campus in Abu Dhabi. G42’s October 2025 update described construction of the first 200 MW for delivery in 2026. Separate initiatives include Microsoft’s UAE investment programme, Abu Dhabi’s government digital strategy, and Oracle’s Blackwell supercluster for regional cloud customers.

#### Funding and source notes

- OpenAI’s $500B Stargate investment target applies to the United States. G42’s May and October 2025 announcements specify the UAE campus’s planned capacity but give no total construction budget.
- Microsoft’s spending plan and the government digital strategy cover different activities and periods. The sources do not establish how much government expenditure may pay for commercial cloud services, so these amounts are not added into a combined total.
- Microsoft’s $15.2B covers investment and expenditure in the UAE from 2023 through 2029, including its $1.5B equity stake in G42, data-centre construction, and operating costs. It combines past spending with future plans.
- Abu Dhabi’s AED13B digital strategy runs from 2025 to 2027 and covers government services, cloud migration, AI applications, and a unified resource-planning system. The full budget cannot be attributed to GPU purchases.
- Oracle announced deployment of an Abu Dhabi supercluster with more than 4,000 NVIDIA Blackwell GPUs in November 2025. Its release describes services for government and commercial customers across the region; it gives neither the cluster’s price nor a government-owned GPU count.
- Core42’s H100 inventory in the UAE is not quantified by the cited material. Access to GPUs at a European partner facility would not establish an equivalent quantity installed in the UAE.
- G42’s November 2025 announcement reports US approval for advanced-chip exports under its Regulated Technology Environment. The approval permits deployment; the announcement does not establish that the full Stargate campus is operational.
- MGX Fund I is an investment fund with a global portfolio. Its reported fundraising is excluded from these UAE infrastructure figures.

#### Programme quantities

- NVIDIA Blackwell (Oracle OCI Supercluster, Abu Dhabi region): >4,000. Status: operational. Physical hardware. Oracle’s Abu Dhabi supercluster serves government and commercial customers.
  [Count source](<https://www.oracle.com/middleeast/news/announcement/oracle-accelerates-sovereign-ai-capabilities-with-the-launch-of-the-first-supercloud-cluster-2025-11-24/>) (authority tier 1; published 2025-11-24).

### Africa (Regional): Kenya, Rwanda, Nigeria + Continental Initiatives

Programme record and sources: https://scrutica.com/sovereign-ai/program/sovereign-africa-regional

- Country: `africa-regional` · Status: Strategies published; implementation varies by country and provider
- Interdependence: NVIDIA: high · US: high
- Key partners: UNDP, Cassava Technologies, NVIDIA, Google, AWS, Microsoft Azure
- Cited source URLs: 18

The African Union’s continental strategy, endorsed in July 2024, calls for coordinated national AI policies and regional cooperation. Implementation combines public research infrastructure with commercial services: South Africa’s CHPC operates a research GPU cluster, while Cassava offers a South African AI service and plans expansion into Nigeria, Kenya, Egypt and Morocco.

#### Funding and source notes

- Regional totals for public spending and installed GPUs are unknown.
- South Africa’s CHPC operates a research cluster with nine GPU nodes containing 30 NVIDIA V100 devices.
- Cassava’s AI Factory expansion is a commercial project. The cited company announcement describes deployment in South Africa and planned expansion into four other countries.
- UNESCO’s Rwanda assessment describes US$76.5M as the cost of implementing the country’s AI policy over 2023–2028. That budget spans policy priorities including skills and AI adoption across public services and industry.

### Australia: National AI Plan + GovAI Sovereign Compute

Programme record and sources: https://scrutica.com/sovereign-ai/program/sovereign-au

- Country: `au` · Status: Plan published (December 2025); NCI and Pawsey systems operational
- Government funding: ~$26.2M
- Timeline: 2025–?
- Interdependence: NVIDIA: high · US: high
- Key partners: CSIRO, Department of Industry, Science and Resources, OpenAI, NextDC, AWS, Microsoft, NCI, Pawsey Supercomputing Research Centre
- Cited source URLs: 15

The December 2025 National AI Plan brings together existing funding for AI research and adoption with measures to expand local infrastructure. It describes GovAI as an Australian-hosted service through which public agencies can develop AI applications. NCI and Pawsey operate research computing systems; Microsoft, AWS and Firmus have announced separate commercial infrastructure investments.

#### Funding and source notes

- The plan combines more than A$460M of existing AI and related funding. Its A$39.9M measure supports AI policy and capability, including the National AI Centre.
- Gadi has 120 H200, 16 A100 and 640 V100 GPUs. Setonix has 768 physical MI250X cards across 192 nodes; software exposes their 1,536 GPU dies separately. Both systems support research workloads, including AI.
- Microsoft, AWS and Firmus announce investment in their own commercial infrastructure. Their budgets have different dates and scopes, including construction, operations and expansion.
- The A$39.9M policy and capability measure is approximately US$26.2M at the Federal Reserve H.10 rate of US$0.6556 per Australian dollar on 2 December 2025.

#### Programme quantities

- NVIDIA H200: 120. Status: operational. Physical hardware. Gadi’s H200 partition: 30 nodes with four GPUs each.
  [Count source](<https://www.nci.org.au/infrastructure/hpc-systems>) (authority tier 1; publication date not recorded).
- NVIDIA A100: 16. Status: operational. Physical hardware. Gadi’s A100 partition: two nodes with eight GPUs each.
  [Count source](<https://www.nci.org.au/infrastructure/hpc-systems>) (authority tier 1; publication date not recorded).
- NVIDIA V100: 640. Status: operational. Physical hardware. Gadi’s V100 partition: 160 nodes with four GPUs each.
  [Count source](<https://www.nci.org.au/infrastructure/hpc-systems>) (authority tier 1; publication date not recorded).
- AMD Instinct MI250X: 768. Status: operational. Physical hardware. Setonix’s 192 GPU nodes each contain four physical MI250X modules; each module has two compute dies.
  [Count source](<https://pawsey.atlassian.net/wiki/spaces/US/pages/51928618/Setonix%2BGPU%2BPartition%2BQuick%2BStart/>) (authority tier 1; publication date not recorded).

### Brazil: Brazilian Artificial Intelligence Plan (PBIA) 2024-2028

Programme record and sources: https://scrutica.com/sovereign-ai/program/sovereign-br

- Country: `br` · Status: Plan published; Santos Dumont upgrade operational (July 2025)
- Timeline: 2024–2028
- Interdependence: NVIDIA: high · US: medium
- Key partners: LNCC, Eviden (Atos Group), BNDES, FNDCT, MCTI, Inria Brasil
- Cited source URLs: 12

PBIA sets out R$23.03B of planned investment for 2024–2028, including R$5.79B for infrastructure and AI development. Enterprise support combines loans and non-repayable funding. LNCC’s Santos Dumont upgrade was completed in July 2025; Eviden reported processing capacity of 18.85 PFLOPS.

#### Funding and source notes

- PBIA’s headline investment spans infrastructure, research, skills, business support and public services. The final plan describes projected investment over 2024–2028, with separate financing mechanisms for individual actions.
- The July 2025 supplier announcement supports 248 H100 GPUs (62 × 4), 144 Grace Hopper superchips (36 × 4) and 36 MI300A APUs (18 × 2). LNCC’s installed-system specifications independently report the same configurations.
- Eviden reports 18.85 PFLOPS for the upgraded system. Its release does not identify the benchmark or numerical precision for that figure.

#### Programme quantities

- NVIDIA H100: 248. Status: operational. Physical hardware. Santos Dumont’s 248-H100 partition after the July 2025 upgrade.
  [Count source](<https://www.bull.com/en/press-releases/santos-dumont-lncc-supercomputer-receives-fourfold-upgrade>) (authority tier 1; published 2025-07-09).
- NVIDIA GH200 Grace Hopper: 144. Status: operational. Physical hardware. Santos Dumont’s 144 Grace Hopper superchips after the July 2025 upgrade.
  [Count source](<https://www.bull.com/en/press-releases/santos-dumont-lncc-supercomputer-receives-fourfold-upgrade>) (authority tier 1; published 2025-07-09).
- AMD MI300A: 36. Status: operational. Physical hardware. Santos Dumont’s 36 MI300A APUs after the July 2025 upgrade.
  [Count source](<https://www.bull.com/en/press-releases/santos-dumont-lncc-supercomputer-receives-fourfold-upgrade>) (authority tier 1; published 2025-07-09).

#### Partner quantities

- Laboratório Nacional de Computação Científica (LNCC) (NVIDIA Hopper H100 (HGX)): 248. Participation stage: commissioned. Physical hardware. Santos Dumont’s H100 partition; GH200 and MI300A devices belong to separate partitions.
  [Count source](<https://www.bull.com/en/press-releases/santos-dumont-lncc-supercomputer-receives-fourfold-upgrade>) (authority tier 1; published 2025-07-09).

### Canada: Canadian Sovereign AI Compute Strategy / Pan-Canadian AI Strategy / AI for All

Programme record and sources: https://scrutica.com/sovereign-ai/program/sovereign-ca

- Country: `ca` · Status: funding_approved / procurement_active
- Timeline: 2024–2030
- Interdependence: NVIDIA: high · US: high
- Key partners: Digital Research Alliance of Canada, CIFAR, ISED, Canada Infrastructure Bank, NRC
- Cited source URLs: 15

Canada’s compute strategy combines public supercomputing with support for commercial data centres and subsidies for smaller businesses’ compute purchases. The June 2026 AI for All strategy adds C$700M to the Compute Access Fund and describes proposed commercial partnerships that could supply 850 MW by 2030, with potential expansion to 2.3 GW.

#### Funding and source notes

- Budget 2024 provided up to C$2B over five years: C$700M for the AI Compute Challenge, C$1B for public computing infrastructure, and C$300M for the Compute Access Fund.
- Budget 2025’s C$925.6M includes C$800M previously provided for in the fiscal framework. Its overlap with the Budget 2024 allocation is unresolved.
- The March 2025 breakdown assigned up to C$705M to the public supercomputer and C$200M to existing infrastructure, both within the C$1B public-infrastructure envelope. A smaller secure facility was also planned, with no separate amount specified.
- The 2026 supercomputer call proposes approximately C$890M over seven fiscal years from 2026–27 for design, construction and operation. Funding remains subject to spending authority, negotiation and a contribution agreement. Applications closed on 1 June 2026.
- Narval’s existing configuration comprises 159 nodes with four A100 40 GB GPUs each, or 636 GPUs. The new supercomputer call gives no device count.
- The 850 MW proposed for 2030 and potential expansion to 2.3 GW depend on commercial partnerships still being finalised in June 2026.

#### Programme quantities

- NVIDIA A100 SXM4 (40GB): 636. Status: operational. Physical hardware. Narval’s A100 partition: 159 nodes with four GPUs each.
  [Count source](<https://docs.alliancecan.ca/wiki/Narval/en>) (authority tier 1; publication date not recorded).

### China: National AI Programs / Big Fund III / Eastern Data Western Computing

Programme record and sources: https://scrutica.com/sovereign-ai/program/sovereign-cn

- Country: `cn` · Status: operational at scale / procurement_active
- Timeline: 2022–2025
- Interdependence: NVIDIA: unknown · US: high
- Key partners: Huawei, Biren Technology, Alibaba, ByteDance, Tencent, Baidu, MIIT
- Cited source URLs: 8

China’s national compute plans combine regional data-centre development with investment in domestic chips and AI companies. Big Fund III has RMB344B in registered capital and helped establish the RMB60B National AI Industry Investment Fund. Its contribution to the latter is not separately disclosed, so the two fund sizes have no additive total here.

#### Funding and source notes

- Big Fund III helped establish the RMB60B AI fund. Xuhui’s January 2025 announcement gives no contribution split, so the two fund sizes are shown separately without an additive total.
- Big Fund III’s RMB344B is registered capital. The six state-owned banks’ May 2024 announcements included contributions to be made over ten years; registration does not establish that the money has been paid in or invested.
- The cited material gives no fund-level disbursement amount. Provincial and municipal programmes also finance AI, so the two funds do not describe all public support.
- MIT Technology Review reported unused AI data-centre capacity in March 2025. Its investigation does not establish a national utilisation rate.
- The cited material does not provide an installed count for the Ascend 910B or 910C.
- Bloomberg reported a draft RMB2T nationwide AI data-centre plan in June 2026. This record carries no enacted primary document or approved budget for that proposal; it is excluded from the funding figures.

### Germany: National AI Strategy / Gaia-X / Gauss Centre for Supercomputing

Programme record and sources: https://scrutica.com/sovereign-ai/program/sovereign-de

- Country: `de` · Status: funding_approved / procurement_active / operational (existing GCS sites)
- Government funding: ~$3.62B
- Committed: ~$3.62B
- Timeline: 2019–2030
- Interdependence: NVIDIA: high · US: medium
- Key partners: BMBF, BMWK, BMAS, Eviden, NVIDIA, Deutsche Telekom, Gaia-X Association
- Cited source URLs: 13

Germany’s national AI strategy funds research and adoption alongside computing infrastructure. OECD reports €3.38B allocated to specific projects by June 2024. At the computing centres, Jülich operates JUPITER; LRZ’s Blue Lion is planned to enter service in 2027. HammerHAI has offered services on existing systems since April 2025 and signed a contract for its new HPE supercomputer in March 2026.

#### Funding and source notes

- OECD reports €3.38B allocated to specific projects across the national AI strategy by June 2024. It gives no payments to recipients.
- Blue Lion’s €250M budget includes acquisition and operation through 2032. HammerHAI’s approximately €85M project budget covers an AI Factory led by HLRS.
- EuroHPC’s March 2026 HammerHAI contract announcement scheduled delivery for the second quarter and operation for the second half of 2026. HammerHAI had already opened its services on existing partner systems in April 2025.
- Gaia-X develops rules and specifications for interoperable data exchange. JUPITER’s hardware also appears in the EuroHPC regional record, so country and regional hardware totals require deduplication.
- In November 2025, the federal government planned approximately €5.5B from the Infrastructure and Climate Neutrality Special Fund for the broader Hightech Agenda during the legislative term; spending remains subject to parliamentary budget provision.
- The dollar figures convert the €3.38B allocation at US$1.0711 per euro on 28 June 2024, the last rate published in the month the allocation is dated to, from the Federal Reserve’s H.10 series distributed by FRED. They restate an allocation, not a payment.

### EU-Wide: EuroHPC Joint Undertaking

Programme record and sources: https://scrutica.com/sovereign-ai/program/sovereign-eurohpc

- Country: `eurohpc` · Status: JUPITER operational; LISA inaugurated June 2026, user access targeted for summer
- Timeline: 2021–2027
- Interdependence: NVIDIA: high · US: high
- Key partners: EU Commission, EuroHPC participating countries, NVIDIA, AMD, HPE, Eviden
- Cited source URLs: 16

EuroHPC co-funds supercomputers and access services with participating countries. Its three AI Factory selections, from December 2024 to October 2025, named 19 sites. JUPITER reached one exaflop on the LINPACK benchmark and ranked fourth in the November 2025 TOP500. LISA adds 1,328 H100 GPUs to Leonardo in Bologna; its June 2026 inauguration announcement targeted user access that summer. The wider EuroHPC budget is at least €8.2 billion for 2021–27 and includes supercomputing and quantum activities.

#### Funding and source notes

- The €8.2 billion budget includes €4.1 billion from EU programmes and a similar contribution from participating countries. Private members are expected to contribute a further €900 million; these amounts cover different funding sources within the wider EuroHPC programme.
- The September 2024 AI supercomputer call offered up to €800 million in EU funding, matched by participating countries.
- LISA’s €50 million budget covers acquisition, installation, operation and support. Its relationship to the wider budget is not itemised in the inauguration announcement.
- The GPU entries cover JUPITER and LISA. EuroHPC describes JUPITER’s accelerator count as approximately 24,000 GH200 superchips.
- The 13 AI Factory Antennas selected in October 2025 had an EU funding contribution of around €55 million, matched by participating states.
- The partner list includes selected AI Factory hosts and consortium members. Selection announcements do not specify a GPU allocation for every member.

#### Programme quantities

- NVIDIA GH200 Grace Hopper Superchip (JUPITER, Eviden BullSequana XH3000): ~24,000. Status: operational. Physical hardware. Whole JUPITER system, described by EuroHPC in November 2025.
  [Count source](<https://www.eurohpc-ju.europa.eu/jupiter-officially-propels-europe-exascale-era-2025-11-17_en>) (authority tier 1; published 2025-11-17).
- NVIDIA H100 (LISA upgrade, 166 8-way BullSequana AI 600 servers): 1,328. Status: commissioning. Physical hardware. Whole LISA extension: 166 servers with eight H100 GPUs each, also described under Italy and EuroHPC.
  [Count source](<https://www.eurohpc-ju.europa.eu/inauguration-sol-new-quantum-computer-together-lisa-upgrade-leonardo-supercomputer-2026-06-11_en>) (authority tier 1; published 2026-06-11).

#### Partner quantities

- NAISS (Mimer AI Factory, Sweden — Round 1) (NVIDIA GB200): 400. Participation stage: awarded. Physical hardware. Whole MIMER design: 100 nodes with 400 GB200 GPUs. Sweden and EuroHPC each allocate half the computing resources.
  [Count source](<https://www.naiss.se/news/2026/04/bull-to-deliver-ai-optimised-supercomputer-for-mimer-ai-factory/>) (authority tier 1; published 2026-04-21).
- VSB-TUO / IT4Innovations (CZAI, Czech Republic — Round 3) (model not recorded): ~340. Participation stage: empanelled. Physical hardware. Whole planned KarolAIna machine; EuroHPC receives half its computing time and services.
  [Count source](<https://www.it4i.cz/o-it4i/infoservis/novinky/cesko-ziskalo-svou-tovarnu-na-umelou-inteligenci-a-s-ni-i-novy-ai-superpocitac>) (authority tier 1; published 2025-10-10).

### Finland: LUMI + LUMI-AI (EuroHPC AI Factory)

Programme record and sources: https://scrutica.com/sovereign-ai/program/sovereign-fi

- Country: `fi` · Status: LUMI operational; LUMI-AI contract awarded August 2026
- Timeline: 2022–2027
- Interdependence: NVIDIA: low · US: medium
- Key partners: CSC, EuroHPC JU, SRV, Finnish Government
- Cited source URLs: 14

LUMI’s GPU partition has 2,978 nodes with four AMD MI250X modules each, or 11,912 modules. Its measured Linpack performance is 379.7 PFlop/s. The LUMI AI Factory began services on this existing system while its successor was procured. EuroHPC signed the LUMI-AI supply contract with Bull in August 2026; deployment is scheduled for the second half of 2027. CSC hosts the systems in Kajaani for a multinational consortium.

#### Funding and source notes

- The GPU count follows LUMI’s operator documentation and the architecture specification updated in autumn 2023. Each MI250X module contains two compute dies, which Slurm exposes separately; the count here is physical modules.
- LUMI’s original procurement had a €144.5M budget in 2020, shared equally by EuroHPC and the ten-country LUMI consortium.
- Finland pledged up to €250M for LUMI-AI in December 2024.
- The installed GPU count covers the existing LUMI-G partition. LUMI-AI has a signed procurement contract and a 2027 deployment target.

#### Programme quantities

- AMD Instinct MI250X: 11,912. Status: operational. Physical hardware. LUMI-G’s 2,978 nodes each contain four physical MI250X modules; each module has two compute dies.
  [Count source](<https://docs.lumi-supercomputer.eu/hardware/lumig/>) (authority tier 1; publication date not recorded).

### France: France 2030 / AI Action Summit / GENCI / Mistral Compute / SoftBank Hauts-de-France

Programme record and sources: https://scrutica.com/sovereign-ai/program/sovereign-fr

- Country: `fr` · Status: Jean Zay operational; Mistral Compute acquisition financed March 2026; SoftBank campuses announced for 2031
- Government funding: ~$2.72B
- Committed: ~$2.72B
- Timeline: 2018–2031
- Interdependence: NVIDIA: high · US: medium
- Key partners: GENCI, Eviden (Atos), CNRS, OVHcloud, Scaleway, Mistral AI, ASML, SoftBank
- Cited source URLs: 17

France combines public research computing with commercial cloud and data-centre projects. GENCI’s Jean Zay, operated by CNRS at IDRIS, has 1,456 H100, 416 A100 and 1,832 V100 GPUs. Mistral AI secured financing in March 2026 to acquire 13,800 GB300 GPUs for Bruyères-le-Châtel. SoftBank’s separate campus plan targets 3.1 GW in Hauts-de-France by 2031 as a first phase.

#### Funding and source notes

- ANR identifies a €2.5B national AI strategy budget under France 2030, which also funds research and training.
- The Jean Zay extension received a €40M grant and was inaugurated on 13 May 2025. The grant announcement does not itemize payments to suppliers.
- Mistral’s transaction adviser describes approximately $830M of financing for the acquisition of 13,800 NVIDIA GB300 GPUs. Availability of that specific cohort is unconfirmed. NVIDIA’s June 2025 plan for 18,000 Blackwell and Blackwell Ultra GPUs describes an earlier, broader deployment plan.
- Mistral AI, MGX, Bpifrance and NVIDIA announced a joint venture for a Paris-region campus in May 2025, targeting 1.4 GW. Construction was expected in the second half of 2026, with operations from 2028.
- SoftBank’s May 2026 plan proposes up to €75B and 5 GW in France. Its first phase specifies €45B for 3.1 GW at Dunkirk, Bosquel and Bouchain by 2031; these are planned investments. The €45B is included in the €75B.
- France 2030’s additional €655M, announced on 16 June 2026, supports AI Gigafactories and the European common-interest AI project. Its overlap with earlier strategy allocations has not been reconciled.
- The recorded private pledges differ in scope and date; their combined programme total is not established.
- The US$87.345B equivalent of SoftBank’s €75B upper-limit plan uses the ECB reference rate of US$1.1646 per euro on 1 June 2026.
- The February 2025 Franco-Emirati framework proposed up to €50B for a 1-GW AI campus. A consolidated funding schedule across this framework and the May 2025 joint venture is unavailable.
- Brookfield’s February 2025 plan allocates up to €15B to Data4 data centres and a further €5B to associated infrastructure, including data transfer, chip storage and energy generation, through 2030. Dollar equivalents for this plan and the Franco-Emirati framework lack a dated conversion source.
- The dollar figures convert the €2.5B France 2030 budget at US$1.0879 per euro on 28 May 2024, ANR’s publication date, from the Federal Reserve’s H.10 series distributed by FRED. SoftBank’s component below uses a different, later rate.

#### Programme quantities

- NVIDIA H100: 1,456. Status: operational. Physical hardware. Jean Zay’s H100 extension: 364 servers with four GPUs each.
  [Count source](<https://www.genci.fr/en/news/genci-and-cnrs-choose-eviden-make-jean-zay-supercomputer-one-most-powerful-france>) (authority tier 1; published 2024-03-28).
- NVIDIA A100: 416. Status: operational. Physical hardware. Jean Zay’s A100 partition, described alongside its March 2024 expansion.
  [Count source](<https://www.genci.fr/en/news/genci-and-cnrs-choose-eviden-make-jean-zay-supercomputer-one-most-powerful-france>) (authority tier 1; published 2024-03-28).
- NVIDIA V100: 1,832. Status: operational. Physical hardware. Jean Zay’s V100 partition, described alongside its March 2024 expansion.
  [Count source](<https://www.genci.fr/en/news/genci-and-cnrs-choose-eviden-make-jean-zay-supercomputer-one-most-powerful-france>) (authority tier 1; published 2024-03-28).
- NVIDIA GB300: 13,800. Status: procurement active. Physical hardware. GB300 acquisition financed for Mistral’s Bruyères-le-Châtel data centre.
  [Count source](<https://www.gide.com/news-insights/gide-conseille-mistral-ai-dans-le-cadre-de-son-premier-financement-dun-montant-de-c-830-millions-de-dollars-americains-afin-doperer-un-data-center-en-region-parisienne-a-bruyeres-le-chatel-franc/>) (authority tier 1; published 2026-03-30).

#### Partner quantities

- CNRS (NVIDIA H100): 1,456. Participation stage: commissioned. Physical hardware. Jean Zay’s H100 extension; its A100 and V100 partitions are separate.
  [Count source](<https://www.genci.fr/en/news/genci-and-cnrs-choose-eviden-make-jean-zay-supercomputer-one-most-powerful-france>) (authority tier 1; published 2024-03-28).

### United Kingdom: AI Opportunities Action Plan / UK Compute Roadmap / AIRR / Sovereign AI Unit / AI Hardware Plan

Programme record and sources: https://scrutica.com/sovereign-ai/program/sovereign-gb

- Country: `gb` · Status: Isambard-AI operational; Sovereign AI launched April 2026; Edinburgh service planned
- Timeline: 2025–2030
- Interdependence: NVIDIA: high · US: high
- Key partners: NVIDIA, Nscale, CoreWeave, Microsoft, British Business Bank, EPCC, University of Bristol, Balderton Capital (James Wise, Sovereign AI Unit chair), AI Safety Institute
- Cited source URLs: 10

Isambard-AI’s full 5,448-GH200 system opened in Bristol on 17 July 2025, providing research compute through the AI Research Resource alongside Cambridge’s Dawn. The July 2025 Compute Roadmap targets 420 AI ExaFLOPS of public capacity by 2030, up from 21 in 2025. Sovereign AI launched in April 2026 with a £500M fund, direct equity investment and access to public supercomputers for startups. The June 2026 hardware plan includes a mixed-chip AI machine targeted for 2030; Edinburgh’s planned £750M national supercomputing service is a separate project.

#### Funding and source notes

- The July 2025 Compute Roadmap commits up to £2B for public research compute through 2030, including over £1B for AIRR expansion and up to £750M for the Edinburgh service. The 2025 AIRR plan budgeted approximately £250M for AI compute infrastructure and scheduled procurement competitions.
- The £1.1B June 2026 hardware plan includes £750M for a new AI machine, with £400M of that for chips; £120M for hardware innovation; £45M for skills; and up to £150M for a British Business Bank investment fund. The plan is additional to Edinburgh’s £750M service; its overlap with other parts of the £2B Compute Roadmap budget is unspecified.
- The Roadmap, published on 17 July 2025, reported £44B of private investment in AI data centres over the preceding 12 months.
- The installed Isambard-AI count is the full system reported by Bristol in July 2025. The cited Edinburgh announcements do not specify an accelerator count.

#### Programme quantities

- NVIDIA GH200: 5,448. Status: operational. Physical hardware. Whole Isambard-AI system. Bristol and NVIDIA describe the same machines as host and chip supplier.
  [Count source](<https://www.bristol.ac.uk/news/2025/july/isambard-launch.html>) (authority tier 1; published 2025-07-17).

#### Partner quantities

- University of Bristol (NVIDIA GH200 Grace Hopper Superchip): 5,448. Participation stage: commissioned. Physical hardware. Whole Isambard-AI system. Bristol and NVIDIA describe the same machines as host and chip supplier.
  [Count source](<https://www.bristol.ac.uk/news/2025/july/isambard-launch.html>) (authority tier 1; published 2025-07-17).
- University of Cambridge (Intel Data Center GPU Max 1550): 1,024. Participation stage: commissioned. Physical hardware. Dawn’s 256 servers each contain four Intel Max 1550 GPUs; each GPU has two internal stacks.
  [Count source](<https://docs.hpc.cam.ac.uk/hpc/user-guide/pvc.html>) (authority tier 1; publication date not recorded).
- NVIDIA (NVIDIA GH200 Grace Hopper Superchip): 5,448. Participation stage: awarded. Physical hardware. Whole Isambard-AI system. Bristol and NVIDIA describe the same machines as host and chip supplier.
  [Count source](<https://www.bristol.ac.uk/news/2025/july/isambard-launch.html>) (authority tier 1; published 2025-07-17).

### Indonesia: National AI Roadmap + Danantara Sovereign AI Fund (Planned)

Programme record and sources: https://scrutica.com/sovereign-ai/program/sovereign-id

- Country: `id` · Status: Roadmap published; AI fund proposed; HPC infrastructure launched at BRIN in June 2026
- Timeline: 2025–2029
- Interdependence: NVIDIA: high · US: high
- Key partners: BRIN, KOMDIGI, Danantara, BDx Indonesia
- Cited source URLs: 8

Indonesia’s 2025 AI roadmap proposed a sovereign AI fund for 2027–2029, with Danantara as its main operator and public and private financing. The proposal did not specify a fund size. Infrastructure projects have proceeded separately: Komdigi’s AI Center of Excellence with Indosat, Cisco and NVIDIA includes a planned Blackwell-based AI factory; KISTI reported the June 2026 launch of Korea–ASEAN-funded HPC infrastructure at BRIN for regional research.

#### Funding and source notes

- The proposed 2027–2029 AI fund is described in Indonesia Business Post’s August 2025 report on the roadmap; its funding amount is unspecified.
- NVIDIA’s July 2025 Center of Excellence announcement names Blackwell GPUs but gives no system count, specific Blackwell model or purchase value.
- KISTI’s June 2026 announcement places the BRIN infrastructure within a Korea–ASEAN cooperation project. Its $10M project budget also covers information services and training across ASEAN; it is not an Indonesian AI fund allocation.

### Ireland: CASPIr Supercomputer (EuroHPC) + AIF IRL-Antenna

Programme record and sources: https://scrutica.com/sovereign-ai/program/sovereign-ie

- Country: `ie` · Status: CASPIr procurement; tender closed May 2026
- Timeline: 2025–2027
- Interdependence: NVIDIA: unknown · US: high
- Key partners: ICHEC, University of Galway, EuroHPC JU, Department of Enterprise, Trade and Employment
- Cited source URLs: 7

CASPIr is planned to provide more than 15 PFlop/s for Irish and European researchers, public services and smaller businesses. EuroHPC and the University of Galway will jointly own the system; ICHEC will operate it. The acquisition budget is up to €25M, with 65% from Ireland and 35% from EuroHPC. The tender opened in March 2026 and closed in May. CASPIr will also support Ireland’s AI Factory Antenna, which provides technical resources and training for AI adoption.

#### Funding and source notes

- The cited procurement documents describe a future system and do not identify the accelerator model or installed GPU count.
- Ireland’s 65% share of the €25M acquisition ceiling is €16.25M. The tender gives no payment figures.

### Israel: National AI Program (Telem) & Supercomputer

Programme record and sources: https://scrutica.com/sovereign-ai/program/sovereign-il

- Country: `il` · Status: operational
- Timeline: 2024–2027
- Interdependence: NVIDIA: high · US: high
- Key partners: Israel Innovation Authority, Nebius, NVIDIA
- Cited source URLs: 7

The Israel Innovation Authority opened applications for subsidised use of Nebius’s live B200 infrastructure on 18 January 2026. The programme allocates computing resources totalling 1,000 B200 accelerators over the coming years, with 70% for companies training AI models and 30% for basic research.

#### Funding and source notes

- Nebius won the tender on 14 May 2025 and committed to beginning operations by early 2026. The January 2026 announcement confirmed live infrastructure and opened applications for access.
- Industry applications require at least 16 B200 accelerators and academic applications at least eight, for periods of one to six months. Use begins no earlier than the month after committee approval.
- The May 2025 award described project investment above NIS500M, including NIS160M in government support, without payment figures.
- The separate Phase II national AI programme received a NIS500M budget through 2027, announced in September 2024. It also funds research infrastructure, a national AI research institute and human-capital development.

#### Programme quantities

- NVIDIA B200: 1,000. Status: operational. Shared access. B200 access allocated over coming years; individual grants last one to six months.
  [Count source](<https://innovationisrael.org.il/en/press_release/supercomputer-access-2026/>) (authority tier 1; published 2026-01-18).

#### Partner quantities

- Nebius (NVIDIA B200): 1,000. Participation stage: awarded. Shared access. B200 access allocated over coming years; individual grants last one to six months.
  [Count source](<https://innovationisrael.org.il/en/press_release/supercomputer-access-2026/>) (authority tier 1; published 2026-01-18).

### Italy: IT4LIA AI Factory + Leonardo + CINECA

Programme record and sources: https://scrutica.com/sovereign-ai/program/sovereign-it

- Country: `it` · Status: Leonardo operational; LISA inaugurated June 2026 with user access expected that summer; IT4LIA supercomputer contracted April 2026
- Timeline: 2022–?
- Interdependence: NVIDIA: high · US: high
- Key partners: CINECA, EuroHPC JU, MUR, ACN, INFN, Leonardo S.p.A., IQM
- Cited source URLs: 13

CINECA operates Leonardo at Bologna’s Tecnopolo DAMA. Its Booster partition has 13,824 NVIDIA A100 GPUs and entered full production in July 2023. The LISA extension adds a separate 1,328-GPU system; at its June 2026 inauguration, user access was expected during the summer. IT4LIA’s new AI supercomputer was contracted to E4 Computer Engineering and Dell in April 2026.

#### Funding and source notes

- Leonardo’s Booster partition has 3,456 nodes with four A100 GPUs each. LISA’s 166 eight-GPU servers are a separate extension.
- The ministry’s February 2022 financing statement assigns €120M to Leonardo over 2019–25, alongside €120M allocated by EuroHPC. These allocations do not itemize payments to suppliers.
- EuroHPC’s June 2026 LISA announcement gives a €50M budget for acquisition, installation, operation and technical support. It dates the inauguration to 11 June and expects user availability during summer 2026.
- The April 2026 IT4LIA procurement contract covers acquisition, delivery, installation and maintenance for €290M, split equally between EuroHPC and Italy’s Ministry of University and Research.
- E4 describes an approximately €420M overall IT4LIA project, which has a wider scope than the €290M equipment contract.

#### Programme quantities

- NVIDIA A100 (Leonardo partition): 13,824. Status: operational. Physical hardware. Leonardo’s Booster partition: 3,456 nodes with four A100 GPUs each.
  [Count source](<https://docs.hpc.cineca.it/hpc/leonardo.html>) (authority tier 1; publication date not recorded).
- NVIDIA Hopper (LISA): 1,328. Status: commissioning. Physical hardware. Whole LISA extension: 166 servers with eight H100 GPUs each, also described under Italy and EuroHPC.
  [Count source](<https://www.eurohpc-ju.europa.eu/inauguration-sol-new-quantum-computer-together-lisa-upgrade-leonardo-supercomputer-2026-06-11_en>) (authority tier 1; published 2026-06-11).

### Japan: METI ¥10T AI Strategy / RIKEN FugakuNEXT / Rapidus

Programme record and sources: https://scrutica.com/sovereign-ai/program/sovereign-jp

- Country: `jp` · Status: operational / early_access (ROQUO / RIKYU)
- Government funding: ~$7.90B
- Committed: ~$7.90B
- Timeline: 2024–2030
- Interdependence: NVIDIA: high · US: high
- Key partners: NVIDIA, Fujitsu, Sakura Internet, KDDI, SoftBank, GMO, Highreso, RUTILEA, RIKEN, MEXT, Rapidus
- Cited source URLs: 31

RIKEN's RIKYU system has 1,600 Blackwell GPUs and is in restricted early access through September 2026. Its separate ROQUO system began operation on 19 June 2026 with 540 Blackwell GPUs for hybrid quantum and supercomputer research. FugakuNEXT targets operation around 2030 and more than 600 EFLOPS at sparse FP8 precision. METI's AI and semiconductor framework provides for more than ¥10 trillion in public support through FY2030. Rapidus reported working 2 nm GAA transistor prototypes in July 2025; it targets mass production in 2027.

#### Funding and source notes

- The hardware entries count RIKYU and ROQUO. Partner projects include dated expansion plans and temporary compute access; they do not form a national installed-GPU inventory.
- The FY2026 budget, enacted on 7 April 2026, includes ¥1.239T for AI and semiconductors. It provides ¥150B in equity for IPA to support next-generation semiconductor companies and ¥387.3B for multimodal foundation-model development. Qazinform, citing Jiji Press on 29 December 2025, reported the ¥1.239T proposal as approximately US$7.9B.
- METI’s April 2024 cloud-project approvals set maximum subsidies in yen. Data Center Dynamics reported the awards as ¥72.5B (approximately US$470M). NVIDIA’s May 2024 account described US$740M (approximately ¥114.6B) in planned assistance to six firms over the preceding two months. These accounts cover different award groups; neither provides a payment total.
- ROQUO access initially covers project participants. RIKYU remains in Early Access Phase 2 through September 2026.

#### Programme quantities

- NVIDIA Blackwell: 1,600. Status: delivered. Physical hardware. RIKYU’s 400 GB200 NVL4 nodes; restricted early access through September 2026.
  [Count source](<https://docs.r-ccs.riken.jp/rikyu/en/>) (authority tier 1; publication date not recorded).
- NVIDIA Blackwell: 540. Status: operational. Physical hardware. ROQUO’s 135 GB200 NVL4 nodes; access initially limited to project participants.
  [Count source](<https://www.riken.jp/pr/news/2026/20260619_2/index.html>) (authority tier 1; published 2026-06-19).

#### Partner quantities

- SAKURA Internet (NVIDIA H100 + NVIDIA HGX B200): ~10,800. Participation stage: awarded. Physical hardware. SAKURA’s planned total after Hopper and B200 expansion, described in November 2024.
  [Count source](<https://nvidianews.nvidia.com/news/japan-cloud-leaders-build-nvidia-ai-infrastructure-to-transform-industries>) (authority tier 1; published 2024-11-12).
- Highreso (NVIDIA H200): 1,600. Participation stage: awarded. Physical hardware. Planned GPUs across Highreso’s two AI data centres, described in November 2024.
  [Count source](<https://nvidianews.nvidia.com/news/japan-cloud-leaders-build-nvidia-ai-infrastructure-to-transform-industries>) (authority tier 1; published 2024-11-12).
- RUTILEA (NVIDIA Hopper (H100/H200)): >1,000. Participation stage: awarded. Physical hardware. RUTILEA’s Hopper capacity described in November 2024.
  [Count source](<https://nvidianews.nvidia.com/news/japan-cloud-leaders-build-nvidia-ai-infrastructure-to-transform-industries>) (authority tier 1; published 2024-11-12).
- Ubitus K.K. (NVIDIA H100): 256. Participation stage: empanelled. Project allocation. H100 capacity for a six-month GENIAC project announced in October 2024.
  [Count source](<https://ubitus.ai/geniac/>) (authority tier 1; published 2024-10-11).

### South Korea: National AI Initiative & K-Chips Act

Programme record and sources: https://scrutica.com/sovereign-ai/program/sovereign-kr

- Country: `kr` · Status: NHN Cloud system operational; Naver operation reported by press; Kakao system commissioning
- Committed: ~$1.10B
- Timeline: 2025–2030
- Interdependence: NVIDIA: high · US: high
- Key partners: NVIDIA, Samsung Electronics, SK Group, Hyundai, MSIT, Naver, SK Telecom, LG, NCSoft, Upstage, KEPCO
- Cited source URLs: 16

The 2025 supplementary-budget procurement allocated 13,136 GPUs across NHN Cloud, Kakao and Naver Cloud, with 10,456 reserved for government use. NHN Cloud announced completion of its 7,656 B200 system in May 2026; Kakao reported 2,040 GPUs installed and testing in December 2025 within its 2,424-GPU allocation. ZDNet Korea reported Naver’s 3,056 H200 system in operation in May 2026. Korea’s longer-term targets combine public acquisitions with private-company plans. A separate National AI Computing Center is planned at Haenam.

#### Funding and source notes

- MSIT’s July 2025 procurement announcement allocates 10,080 B200 and 3,056 H200 GPUs across three cloud providers. The KRW 1.46T supplementary-budget allocation was reported as approximately US$1.1B. The government-use quantities are subsets of the provider totals.
- NHN Cloud announced the completed 7,656-B200 system on 13 May 2026, with 6,120 GPUs reserved for government use. Some government-funded resources had been available since 1 April.
- Kakao’s 29 December 2025 announcement describes 2,040 installed B200 GPUs under final testing, within its 2,424-GPU allocation. It scheduled beta access for 2 January 2026 and gives no status for the remaining 384.
- Naver’s operational status follows ZDNet Korea’s 14 May 2026 report on its 3,056 H200 GPUs. Provider confirmation of that status is unavailable.
- Korea Herald reported in December 2025 a 52,000-GPU public-sector target for 2028 within a larger 260,000-GPU public/private plan for 2030. NVIDIA’s October 2025 announcement includes about 50,000 public-sector GPUs and more than 200,000 for private companies, without a completion date. These targets include multiple acquisition programmes.
- The sovereign-model programme selected five consortia in August 2025. At the January 2026 evaluation, LG AI Research, SK Telecom and Upstage advanced; Naver Cloud and NC AI were dropped, and a further selection round was planned to fill a fourth place.
- ZDNet Korea reported on 16 June 2026 that the Korea AI Computing Center company had registered the previous week, with its headquarters at Haenam. It reported Samsung SDS’s stake as 30%, the government’s as 29% and Naver Cloud’s as 26.1%. This corporate formation is separate from the 2025 cloud-provider GPU procurement.
- The AI Basic Act and its Enforcement Decree took effect on 22 January 2026.

#### Programme quantities

- Vendor unspecified (public-sector 2028 target): 52,000. Status: announced. Mixed or unspecified scope. Public-sector GPU target for 2028 reported by Korea Herald.
  [Count source](<https://www.koreaherald.com/article/10636733>) (authority tier 3; published 2025-12-15).
- NVIDIA GPUs (public/private 2030 target): 260,000. Status: announced. Mixed or unspecified scope. Public/private GPU target for 2030 reported by Korea Herald, including public-sector acquisitions.
  [Count source](<https://www.koreaherald.com/article/10636733>) (authority tier 3; published 2025-12-15).
- NVIDIA B200: 7,656. Status: operational. Physical hardware. Whole NHN B200 system, including 6,120 GPUs reserved for government use.
  [Count source](<https://inside.nhn.com/news/939>) (authority tier 1; published 2026-05-13).
- NVIDIA B200: 2,424. Status: commissioning. Physical hardware. Kakao’s 2,424-GPU allocation; 2,040 were installed and testing in December 2025.
  [Count source](<https://www.kakaocorp.com/page/detail/11917>) (authority tier 1; published 2025-12-29).
- NVIDIA H200: 3,056. Status: operational. Physical hardware. Whole Naver H200 system; operation was reported by ZDNet Korea in May 2026.
  [Count source](<https://zdnet.co.kr/view/?no=20260513173011>) (authority tier 3; published 2026-05-14).

#### Partner quantities

- NHN Cloud (NVIDIA B200): 7,656. Participation stage: deployed. Physical hardware. Whole NHN B200 system, including 6,120 GPUs reserved for government use.
  [Count source](<https://inside.nhn.com/news/939>) (authority tier 1; published 2026-05-13).
- Naver Cloud (NVIDIA H200): 3,056. Participation stage: deployed. Physical hardware. Whole Naver H200 system; operation was reported by ZDNet Korea in May 2026.
  [Count source](<https://zdnet.co.kr/view/?no=20260513173011>) (authority tier 3; published 2026-05-14).
- Kakao (NVIDIA B200): 2,424. Participation stage: awarded. Physical hardware. Kakao’s 2,424-GPU allocation; 2,040 were installed and testing in December 2025.
  [Count source](<https://www.kakaocorp.com/page/detail/11917>) (authority tier 1; published 2025-12-29).

### Malaysia: Public and private AI cloud initiatives

Programme record and sources: https://scrutica.com/sovereign-ai/program/sovereign-my

- Country: `my` · Status: funding_approved (Sovereign AI Cloud) / operational (private YTL AI Cloud)
- Government funding: ~$474.0M
- Timeline: 2024–2030
- Interdependence: NVIDIA: high · US: high
- Key partners: National AI Office (NAIO), YTL Power International, NVIDIA, Microsoft
- Cited source URLs: 12

Malaysia’s RM2B Sovereign AI Cloud is to be implemented through the Malaysian Communications and Multimedia Commission. YTL’s operating GB200 cloud and Skyvast’s Huawei-based initiative are private projects with separate funding and hardware.

#### Funding and source notes

- Budget 2026 included RM2B for the Sovereign AI Cloud through MCMC. Parliament approved the budget on 15 December 2025.
- The RM20B national-roadmap budget is an unverified World AI Show claim.
- The Sovereign AI Cloud budget announcement gives no GPU count or hardware specification.
- YTL reported its GB200 cloud operational in October 2025. The 600 MW figure covers the wider Green Data Center Park at Kulai; the AI cluster’s power is unspecified.
- Developing Telecoms reported an initial eight Huawei Ascend GPUs and a plan for 3,000 by 2026 at Skyvast in May 2025. On 21 May, MITI identified the initiative as privately driven and outside any government-to-government agreement or nationally mandated technology programme.
- The dollar figures convert the RM2B budget at 4.22 ringgit per US dollar on 10 October 2025, the last rate published before the minister’s statement, from the Federal Reserve’s H.10 series distributed by FRED.

### NATO/AUKUS: AI procurement and joint research

Programme record and sources: https://scrutica.com/sovereign-ai/program/sovereign-nato-aukus

- Country: `nato-aukus` · Status: AI systems procured; joint research and accelerator programmes
- Interdependence: NVIDIA: unknown · US: unknown
- Key partners: NATO DIANA, NATO Innovation Fund, AUKUS AIA Working Group
- Cited source URLs: 9

NATO has procured AI systems for shared institutions. Allied Command Operations acquired Palantir’s Maven Smart System in March 2025; in November, NCIA selected Google’s air-gapped Distributed Cloud for the Joint Analysis, Training and Education Centre. AUKUS trials reported in August 2024 tested shared AI models on uncrewed aircraft. DIANA funds and supports dual-use technology companies, while the NATO Innovation Fund invests in startups under a mandate of at least €1 billion over 15 years.

#### Funding and source notes

- The Maven and Google announcements give no GPU inventory. Google describes a multimillion-dollar contract and planned integration, without an exact price or commissioning date.
- The NATO Innovation Fund’s capital commitment and DIANA’s accelerator grants finance companies and programme activity; neither source itemises a shared GPU installation.
- DIANA’s 2024 financial statements describe €100,000 for each member of the 2025 accelerator cohort. SandboxAQ’s selection announcement identifies its project but gives no dollar-denominated allocation.

### Norway: Sigma2/NRIS National Research Infrastructure + LUMI-AI Participation

Programme record and sources: https://scrutica.com/sovereign-ai/program/sovereign-no

- Country: `no` · Status: operational (Olivia, since 2025) / funding committed (LUMI-AI)
- Timeline: 2025–2027
- Interdependence: NVIDIA: medium · US: medium
- Key partners: Sigma2 AS, HPE, Norwegian Research Council, CSC (Finland), Universities of Bergen, Oslo, Tromso, NTNU
- Cited source URLs: 11

Sigma2 operates Olivia at Lefdal Mine Datacenter and represents Norway in the LUMI-AI consortium. Olivia’s NOK225M contract and Norway’s NOK215M hardware commitment to LUMI-AI cover separate procurements.

#### Funding and source notes

- In January 2025, the Research Council recommended NOK3.4B of supercomputing investment over five years.
- In October 2025, Sigma2 reported a proposed NOK380M allocation over two years in the 2026 budget. This would supplement annual baseline funding of NOK50M and NOK20.7M for work with the National Library on Norwegian and Sami language models.
- Olivia’s NOK225M award records the contract value without payment figures.
- Olivia entered research use in 2025 with 304 GH200 GPUs. Sigma2’s documentation dates the addition of 144 GPUs to February 2026, bringing the system to 448.
- Norway’s NOK215M LUMI-AI contribution purchases a share of hardware hosted in Finland. A separate NOK30M Research Council contribution supports the Norwegian AI Factory, with matching EuroHPC funding.

#### Programme quantities

- NVIDIA GH200 Grace Hopper Superchip: 448. Status: operational. Physical hardware. Whole Olivia system after the February 2026 expansion to 448 GH200 GPUs.
  [Count source](<https://www.mynewsdesk.com/no/sigma2/pressreleases/nasjonalt-loeft-for-kunstig-intelligens-naa-blir-olivia-enda-kraftigere-3434764>) (authority tier 1; published 2026-03-02).

#### Partner quantities

- Sigma2 AS (NVIDIA GH200 Grace Hopper Superchip): 448. Participation stage: deployed. Physical hardware. Whole Olivia system after the February 2026 expansion to 448 GH200 GPUs.
  [Count source](<https://www.mynewsdesk.com/no/sigma2/pressreleases/nasjonalt-loeft-for-kunstig-intelligens-naa-blir-olivia-enda-kraftigere-3434764>) (authority tier 1; published 2026-03-02).

### Philippines: NAICRI (National AI Center for Research and Innovation) + DOST-ASTI COARE/Saliksik HPC

Programme record and sources: https://scrutica.com/sovereign-ai/program/sovereign-ph

- Country: `ph` · Status: NAICRI launched February 2026; Saliksik operational
- Interdependence: NVIDIA: high · US: high
- Key partners: DOST (Department of Science and Technology), DOST-ASTI (Advanced Science and Technology Institute), NAICRI, COARE (Computing and Archiving Research Environment), PREGINET, NVIDIA
- Cited source URLs: 4

DOST launched the National Artificial Intelligence Center for Research and Innovation (NAICRI) on 26 February 2026 to coordinate AI research, computing infrastructure and shared platforms. DOST-ASTI’s COARE service provides research compute through Saliksik: 16 A100 GPUs across two nodes and six P40 GPUs across six nodes, alongside 36 CPU nodes. The launch also presented PREGINET, the research and education network; NAIRA, an AI service platform; and DIMER, a shared model repository.

#### Funding and source notes

- The Saliksik counts describe the cluster listed in DOST-ASTI’s operator documentation. That inventory gives no installation dates and does not establish a nationwide total for government-owned GPUs.
- The cited NAICRI launch reports do not give a consolidated program budget.

#### Programme quantities

- NVIDIA Tesla A100: 16. Status: operational. Physical hardware. Saliksik’s A100 nodes: two nodes with eight GPUs each.
  [Count source](<https://coarewiki.asti.dost.gov.ph/en/using-coare/hpc>) (authority tier 1; publication date not recorded).
- NVIDIA Tesla P40: 6. Status: operational. Physical hardware. Saliksik’s P40 nodes: six nodes with one GPU each.
  [Count source](<https://coarewiki.asti.dost.gov.ph/en/using-coare/hpc>) (authority tier 1; publication date not recorded).

### Poland: PIAST-AI Factory (EuroHPC) + PSNC

Programme record and sources: https://scrutica.com/sovereign-ai/program/sovereign-pl

- Country: `pl` · Status: PIAST-AI new-infrastructure procurement planned for 2027; full operation targeted for 2028
- Timeline: 2025–2028
- Interdependence: NVIDIA: unknown · US: high
- Key partners: PSNC, Poznan University of Technology, Adam Mickiewicz University, Nicolaus Copernicus University, EuroHPC JU, Polish Ministry of Digital Affairs
- Cited source URLs: 7

PSNC leads PIAST-AI with Poznań University of Technology, Adam Mickiewicz University and Nicolaus Copernicus University. A June 2026 account by PSNC staff describes a €100 million investment, with new-infrastructure procurement planned for 2027 and full operation for 2028. Existing PSNC computing platforms already support AI models and services. Poland also participates in the LUMI AI Factory consortium.

#### Funding and source notes

- PSNC’s June 2025 hosting announcement planned more than 1,000 GPUs without naming a model. The June 2026 account gives a later implementation schedule but no revised GPU specification.
- CORDIS lists €12.5 million in total project costs and a €6.25 million EU contribution for PIAST AIF, running from May 2026 to April 2029. The sources do not itemise how this grant relates to the €100 million infrastructure investment described by PSNC staff.
- The June 2025 hosting agreement establishes PSNC’s role. The May 2026 consortium update discusses GPU procurement without naming an awarded hardware supplier.

### Saudi Arabia: HUMAIN / Project Transcendence

Programme record and sources: https://scrutica.com/sovereign-ai/program/sovereign-sa

- Country: `sa` · Status: procurement_active
- Timeline: 2025–2030
- Interdependence: NVIDIA: high · US: high
- Key partners: NVIDIA, AWS, Google Cloud, PIF, SDAIA, HUMAIN, xAI
- Cited source URLs: 15

HUMAIN and NVIDIA announced an initial 18,000-GPU GB300 cluster in May 2025. HUMAIN's November expansion targeted up to 600,000 NVIDIA GPUs across Saudi Arabia and the United States over three years. AWS separately announced up to 150,000 NVIDIA and Trainium accelerators for a Riyadh AI Zone. In July 2026, HUMAIN and Cohere announced at least 50 MW of dedicated capacity, expected online in Q4 2027. KAUST's operating Shaheen III GPU partition has 2,800 GH200 Grace Hopper Superchips.

#### Funding and source notes

- The May and November 2025 HUMAIN releases describe the first cluster as an announcement and provide no commissioning date.
- The November 2025 target of up to 600,000 NVIDIA GPUs covers Saudi Arabia and the United States; the release gives no country allocation or uniform model for the whole target.
- SDAIA hardware counts remain unverified. The KAUST entry covers Shaheen III’s GPU partition, with no university-wide or national GPU total.
- The AWS target combines NVIDIA GPUs and Trainium accelerators without a published split. The Cohere announcement gives no overlap with HUMAIN’s other capacity targets.
- Middle East AI News reported a proposed Project Transcendence investment of up to US$100B in November 2024, citing Bloomberg. Semafor’s May 2025 account of HUMAIN’s CEO interview put the company’s data-centre buildout at up to US$77B, with investment sought from US technology companies. Neither account specifies a government funding contribution.
- US Commerce authorized HUMAIN in November 2025 for up to the equivalent of 35,000 GB300 GPUs, subject to security and reporting conditions. The authorization provides no evidence of delivery.
- Google Cloud announced a US$10B joint investment with PIF in a Saudi AI hub on 13 May 2025. The investors’ shares and the project’s overlap with HUMAIN’s broader buildout estimate are unspecified.

#### Programme quantities

- NVIDIA GB300: 18,000. Status: announced. Physical hardware. HUMAIN’s initial GB300 cluster announced in May 2025.
  [Count source](<https://nvidianews.nvidia.com/news/humain-and-nvidia-announce-strategic-partnership-to-build-ai-factories-of-the-future-in-saudi-arabia>) (authority tier 1; published 2025-05-13).
- NVIDIA GH200 Grace Hopper: 2,800. Status: operational. Physical hardware. Shaheen III’s GH200 GPU partition.
  [Count source](<https://docs.hpc.kaust.edu.sa/systems/shaheen3/index.html>) (authority tier 1; publication date not recorded).

### Sweden: Berzelius + MIMER AI Factory (EuroHPC)

Programme record and sources: https://scrutica.com/sovereign-ai/program/sovereign-se

- Country: `se` · Status: Berzelius operational; MIMER hardware contract awarded April 2026
- Timeline: 2024–2026
- Interdependence: NVIDIA: high · US: high
- Key partners: Knut and Alice Wallenberg Foundation, Linköping University, Eviden, AI Sweden, EuroHPC JU, Swedish Research Council
- Cited source URLs: 8

Berzelius at Linköping University has 752 A100 GPUs and 128 H200 GPUs across 110 DGX nodes. The Knut and Alice Wallenberg Foundation funded the system and its expansions. MIMER’s April 2026 procurement selected Bull to supply a separate system with 400 GB200 GPUs, with installation planned later in 2026. Its approximately €30 million budget covers acquisition, delivery, installation and maintenance; Sweden and EuroHPC each fund half.

#### Funding and source notes

- The Swedish government’s September 2025 budget proposal included SEK479 million for AI and data reforms in 2026. This broader package does not provide a dollar-denominated programme total.
- The Berzelius inventory comprises 94 eight-A100 nodes and 16 eight-H200 nodes. Its operator dates the latest expansion to 2025.
- Ericsson, AstraZeneca, SAAB, SEB and Wallenberg Investments announced a separate company-led AI Factory in May 2025, with two planned DGX SuperPODs using GB300 systems.

#### Programme quantities

- NVIDIA A100: 752. Status: operational. Physical hardware. Berzelius’ A100 partition: 94 nodes with eight GPUs each.
  [Count source](<https://www.nsc.liu.se/systems/berzelius/>) (authority tier 1; publication date not recorded).
- NVIDIA H200: 128. Status: operational. Physical hardware. Berzelius’ H200 partition: 16 nodes with eight GPUs each.
  [Count source](<https://www.nsc.liu.se/systems/berzelius/>) (authority tier 1; publication date not recorded).

#### Partner quantities

- National Academic Infrastructure for Supercomputing in Sweden (NAISS) (NVIDIA GB200): 400. Participation stage: awarded. Physical hardware. Whole MIMER design: 100 nodes with 400 GB200 GPUs. Sweden and EuroHPC each allocate half the computing resources.
  [Count source](<https://www.naiss.se/news/2026/04/bull-to-deliver-ai-optimised-supercomputer-for-mimer-ai-factory/>) (authority tier 1; published 2026-04-21).

### Thailand: National AI Strategy and Action Plan (2022-2027) + ThaiSC/LANTA + ThaiLLM

Programme record and sources: https://scrutica.com/sovereign-ai/program/sovereign-th

- Country: `th` · Status: LANTA operational since November 2022; ThaiLLM released April 2026
- Timeline: 2022–2027
- Interdependence: NVIDIA: high · US: high
- Key partners: NSTDA (National Science and Technology Development Agency), ThaiSC (NSTDA Supercomputer Center), NECTEC, MHESI (Ministry of Higher Education, Science, Research and Innovation), MDES (Ministry of Digital Economy and Society), Big Data Institute (BDI), Digital Economy and Society Development Fund (DEF), HPE (LANTA vendor), NVIDIA
- Cited source URLs: 5

ThaiSC, NSTDA’s supercomputing centre, operates LANTA at Thailand Science Park in Pathum Thani. Its 176 GPU nodes contain 704 NVIDIA A100s; another 160 CPU nodes and ten high-memory nodes support scientific computing. The 346-node system has 8.15 PFlop/s peak performance and a 200 Gbps Slingshot interconnect. In April 2026, NSTDA announced the release of ThaiLLM’s 8B- and 30B-parameter models, trained on more than 100B Thai tokens and served on ThaiSC infrastructure. The project brings together MDES, BDI, MHESI and NECTEC, with support from the Digital Economy and Society Development Fund.

#### Funding and source notes

- ThaiSC dates LANTA’s operation to November 2022. NSTDA’s February 2026 article describes this existing system.
- NSTDA’s ThaiLLM announcement names the funding body but gives no funding amount. The cited program sources do not supply a consolidated budget.
- ThaiSC lists 10 PB of parallel storage plus 945 TB of NVMe storage. The earlier research recorded 12 PB from the National AI Strategy portal; that discrepancy remains unresolved.

#### Programme quantities

- NVIDIA A100: 704. Status: operational. Physical hardware. LANTA’s A100 partition: 176 nodes with four GPUs each.
  [Count source](<https://thaisc.io/thaisc-resorces/lanta>) (authority tier 1; publication date not recorded).

### Taiwan: AI New Ten Major Construction + NCHC Sovereign Compute

Programme record and sources: https://scrutica.com/sovereign-ai/program/sovereign-tw

- Country: `tw` · Status: operational (NCHC Nano4 service from June 2026)
- Timeline: 2025–2029
- Interdependence: NVIDIA: high · US: high
- Key partners: NCHC, NSTC, NDC, NVIDIA, AMD, Hon Hai (Foxconn), WiAdvance, Visionbay.ai
- Cited source URLs: 15

NCHC opened Nano4 for service on 1 June 2026. Its configuration includes 1,760 H200 GPUs and two GB200 NVL72 systems with 144 Blackwell GPUs between them. NVIDIA's May 2025 announcement also included an HGX B300 system; NCHC’s Nano4 specification lists H200 and GB200 NVL72 hardware, with no B300 entry. Taiwan has released TAIDE language models, beginning with TAIDE-LX-7B in April 2024. The AI Basic Act took effect on 14 January 2026. The Executive Yuan reported completion of the 2026 central-budget review in August; the statement gives no programme-specific AI appropriation.

#### Funding and source notes

- The Nano4 counts use accelerator units: 220 nodes with eight H200 GPUs each, plus two GB200 NVL72 systems with 72 Blackwell GPUs each.
- A current installed GPU count for the HGX B300 system in NVIDIA’s May 2025 announcement has not been established; it is excluded from the hardware total.
- TechSoda reported a draft of approximately NT$200B (US$6.2B) in June 2025, including a proposed NT$90B (US$2.8B) expansion from science, construction and National Development Fund budgets. The report anticipated submission for Executive Yuan approval later that month.
- The Cabinet’s August 2025 draft for FY2026 assigned NT$31.1B to the Ten AI Initiatives: NT$28B in the central budget and NT$3.1B from business and non-business funds. It then went to the legislature. The August 2026 central-budget completion statement gives no final programme amount.
- The hardware entries cover the two named Nano4 configurations. They exclude other NCHC systems.

#### Programme quantities

- NVIDIA H200: 1,760. Status: operational. Physical hardware. Nano4’s H200 partition: 220 nodes with eight GPUs each.
  [Count source](<https://www.nchc.org.tw/Page?itemid=58&mid=109>) (authority tier 1; publication date not recorded).
- NVIDIA Blackwell (GB200 NVL72): 144. Status: operational. Physical hardware. Nano4’s Blackwell partition: two GB200 NVL72 systems with 72 GPUs each.
  [Count source](<https://www.nchc.org.tw/Page?itemid=58&mid=109>) (authority tier 1; publication date not recorded).

#### Partner quantities

- NVIDIA (NVIDIA H200 + NVIDIA Blackwell (GB200 NVL72)): 1,904. Participation stage: deployed. Physical hardware. NVIDIA’s Nano4 accelerator supply: 1,760 H200 plus 144 Blackwell GPUs.
  [Count source](<https://www.nchc.org.tw/Page?itemid=58&mid=109>) (authority tier 1; publication date not recorded).

### Ukraine: National LLM / Diia AI / NVIDIA Sovereign AI Partnership

Programme record and sources: https://scrutica.com/sovereign-ai/program/sovereign-ua

- Country: `ua` · Status: Diia.AI entered public beta in September 2025; national LLM development reported January 2026
- Timeline: 2025–2030
- Interdependence: NVIDIA: high · US: high
- Key partners: Kyivstar (VEON), Google (Gemma, Cloud/Vertex AI), NVIDIA, Ministry of Digital Transformation, WINWIN AI Center of Excellence
- Cited source URLs: 6

Kyivstar finances and leads the national model, based on Google’s Gemma family with Vertex AI computing support. The January 2026 update set a spring beta target, followed by transfer to the state and an open-source release after validation. Diia.AI had already entered public beta in September 2025, with an income-certificate service available through chat. NVIDIA’s November partnership supports a further Diia-focused model through software and technical advice. The planned AI Factory would host government AI services in Ukraine; the cited announcements give no GPU procurement quantity or infrastructure budget.

#### Funding and source notes

- The January 2026 statement that no public budget funds were being used applies to Kyivstar’s national language-model development at that stage.
- VEON identifies the Gemma model family and Vertex AI computing support. It names neither the Gemma version nor the physical location of the training hardware.
- The spring 2026 national-model beta date was a target in the January update; it does not confirm a release.
- The August 2025 AI Factory plan describes domestic GPU clusters, cooling and data infrastructure. Its announcement does not establish that this equipment was installed.
- Eurasia Foundation and East Europe Foundation are credited for support to the Diia.AI service. Their cited roles do not establish AI Factory financing or hardware procurement.

### Vietnam: National AI Strategy & NVIDIA Partnership

Programme record and sources: https://scrutica.com/sovereign-ai/program/sovereign-vn

- Country: `vn` · Status: Viettel and FPT AI services operating; National Data Development Fund established in 2025
- Timeline: 2025–2027
- Interdependence: NVIDIA: high · US: high
- Key partners: FPT Corporation, NVIDIA, Ministry of Science and Technology, Vietnam National Innovation Center
- Cited source URLs: 12

Vietnam’s AI infrastructure includes Viettel’s DGX B200 system at Hoa Lac and FPT’s commercial H100/H200 services. The science ministry reported Viettel’s system entering operation in February 2026; FPT’s $200M AI Factory investment plan dates to April 2024. Decree 160/2025 established the National Data Development Fund with VND 1T in initial state capital for data technologies and services, including AI. A proposed National AI Development Fund is a different institution. Reports of more than $7B in proposed AI and data-centre investments describe the wider market, without a common programme budget.

#### Funding and source notes

- The National Data Development Fund has VND 1T in initial state capital for data technologies and services, including AI. A dated exchange-rate source has not been established for a USD conversion.
- The separately proposed National AI Development Fund has no amount in the cited VietnamPlus report.
- TNGlobal’s July 2026 report attributes the more-than-$7B figure to BMI’s aggregate of announced and proposed AI and data-centre investments. It is not a government allocation or a programme total.
- VietnamPlus reports 22 DGX B200 systems at Viettel. This is a system count; the record does not convert it into an inferred GPU count.
- FPT’s March 2026 announcement describes HGX B300 as a forthcoming service. The operating H100/H200 services and proposed B300 expansion have different delivery stages.

---


## Facility reference: largest recorded power figures, operating records first

This selection takes the forty largest positive power figures on operational records, then the twenty largest on records with any other status (announced, permitted, under construction, expanding, retired or unstated), from the facility reference extract bundled with this guide. The two lists are separate because a figure on a record that is not operating describes planned or past plant: never add it to an operating figure. Both include manufacturing plants. It is a ranking within this file; the [facility register](https://scrutica.com/entities) has broader coverage.

Read the type, status and source alongside each power figure. Even on an operational record the recorded power can describe a whole campus's build-out or an earlier phase of a cluster whose later phase includes it; the facility's page says which, and the site's operating totals hold those figures out. GPU inventories, H100-equivalent capacity, hardware ownership and building ownership are separate fields.

An organisation in a facility's owner or operator field is listed below as recorded there, not as the site's owner or operator: those fields hold whatever the source recorded (a registry listing, a map tag, a named party), and each facility record at scrutica.com states the relationship its own source establishes, or why the attribution is withheld.

Records distinguish “includes estimates”, “not marked estimated” and a missing flag. The flag can concern capacity, coordinates or another field; the source notes describe its scope and method where recorded.

### Operational records

#### Colossus 2 (`fac-epoch-xai-colossus-2`)

Memphis, Tennessee, United States — 1276 MW · IT load 946 MW · H100-equivalent 1,111,672 · type: AI Training · status: Operational · AI hardware owner `org-xai`.
Estimate flag: includes estimates.

Record notes: Total capital cost (2025 USD): $35.84B | Epoch owner label changed 'xAI' -> 'SpaceXAI' (2026-07). Denotes the same xAI entity: X.AI Holdings Corp. was acquired by SpaceX effective 2026-02-02 per SpaceX SEC Form S-1 (filed 2026-05-20, 'the xAI Merger'); 'SpaceXAI' is the post-merger brand (press, 2026-07-06/07). Mapped to org-xai; alias recorded in organization_aliases. | Epoch estimation methods: our cooling equipment power model, company disclosures, drone imagery.. · Source: https://epoch.ai/data/data-centers

Detail page: https://scrutica.com/facilities/fac-epoch-xai-colossus-2.

#### Anthropic-Amazon New Carlisle (`fac-epoch-anthropic-amazon-new-carlisle`)

New Carlisle, Indiana, United States — 1092 MW · IT load 910 MW · H100-equivalent 687,215 · type: AI Training · status: Operational · AI hardware owner `org-amazon`.
Estimate flag: includes estimates.

Record notes: Total capital cost (2025 USD): $34.47B | Epoch estimation methods: regulatory filings, chip efficiency, third-party reporting.. · Source: https://epoch.ai/data/data-centers

Detail page: https://scrutica.com/facilities/fac-epoch-anthropic-amazon-new-carlisle.

#### Microsoft Fairwater Atlanta (`fac-epoch-microsoft-fairwater-atlanta`)

Fayetteville, Georgia, United States — 859 MW · IT load 636 MW · H100-equivalent 768,064 · type: AI Training · status: Operational · AI hardware owner `org-microsoft`.
Estimate flag: includes estimates.

Record notes: Total capital cost (2025 USD): $24.09B | Epoch estimation methods: our cooling equipment power model, drone imagery, company disclosures, chip availability, chip efficiency.. · Source: https://epoch.ai/data/data-centers

Detail page: https://scrutica.com/facilities/fac-epoch-microsoft-fairwater-atlanta.

#### Meta Prometheus (`fac-epoch-meta-prometheus`)

New Albany, Ohio, United States — 820 MW · IT load 631 MW · H100-equivalent 763,011 · type: AI Training · status: Operational · AI hardware owner `org-meta`.
Estimate flag: includes estimates.

Record notes: Total capital cost (2025 USD): $23.90B | Epoch estimation methods: our cooling equipment power model, regulatory filings, chip availability, chip efficiency, land use maps.. · Source: https://epoch.ai/data/data-centers

Detail page: https://scrutica.com/facilities/fac-epoch-meta-prometheus.

#### OpenAI Stargate Abilene (`fac-epoch-openai-stargate-abilene`)

Abilene, Texas, United States — 590 MW · IT load 421 MW · H100-equivalent 510,358 · type: AI Training · status: Operational · AI hardware owner `org-oracle`.
Estimate flag: includes estimates.

Record notes: Total capital cost (2025 USD): $15.95B | Epoch estimation methods: our cooling equipment power model, company disclosures, regulatory filings, drone imagery.. · Source: https://epoch.ai/data/data-centers

Detail page: https://scrutica.com/facilities/fac-epoch-openai-stargate-abilene.

#### TeraWulf Lake Mariner Campus (`fac-epoch-fluidstack-lake-mariner`)

Barker, New York, United States — 510 MW · type: AI Training · status: Operational · AI hardware owner `org-fluidstack`.
Estimate flag: includes estimates.

Record notes: Epoch names Fluidstack as AI hardware owner. The separate TeraWulf campus correction does not establish hardware ownership. | CONTRADICTION (differing scopes): Epoch's AI-building figures, 100 MW IT power and 117 MW facility power (2026-05-31), vs 510 MW disclosed multi-tenant total — Epoch tracks only the AI buildings (3-5, Fluidstack lease); 510 MW spans both tenant blocks incl. Core42's buildings 1-2. | H100-equivalent count kept null: Epoch's ~161,192 figure is anchored to the AI-building scope (100 MW IT). Capacity snapshots carry Epoch's per-date satellite measurements. | Total capital cost (2025 USD): $3.79B. | Epoch estimation methods: our cooling equipment power model, company disclosures, chip availability, chip efficiency.. · Source: https://epoch.ai/data/data-centers

Detail page: https://scrutica.com/facilities/fac-epoch-fluidstack-lake-mariner.

#### Microsoft Fairwater Wisconsin (`fac-epoch-microsoft-fairwater-wisconsin`)

Mount Pleasant, Wisconsin, United States — 442 MW · IT load 369 MW · H100-equivalent 445,679 · type: AI Training · status: Operational · AI hardware owner `org-microsoft`.
Estimate flag: includes estimates.

Record notes: Total capital cost (2025 USD): $13.98B | Epoch estimation methods: company disclosures, regulatory filings, chip availability, chip efficiency.. · Source: https://epoch.ai/data/data-centers

Detail page: https://scrutica.com/facilities/fac-epoch-microsoft-fairwater-wisconsin.

#### Colossus 1 (`fac-epoch-xai-colossus-1`)

Memphis, Tennessee, United States — 425 MW · IT load 340 MW · H100-equivalent 275,795 · type: AI Training · status: Operational · AI hardware owner `org-xai`.
Estimate flag: includes estimates.

Record notes: Total capital cost (2025 USD): $12.88B | Epoch owner label changed 'xAI' -> 'SpaceXAI' (2026-07). Denotes the same xAI entity: X.AI Holdings Corp. was acquired by SpaceX effective 2026-02-02 per SpaceX SEC Form S-1 (filed 2026-05-20, 'the xAI Merger'); 'SpaceXAI' is the post-merger brand (press, 2026-07-06/07). Mapped to org-xai; alias recorded in organization_aliases. | Epoch estimation methods: our cooling equipment power model, company disclosures, regulatory filings.. · Source: https://epoch.ai/data/data-centers

Detail page: https://scrutica.com/facilities/fac-epoch-xai-colossus-1.

#### Google New Albany (`fac-epoch-google-new-albany`)

New Albany, Ohio, United States — 407 MW · IT load 339 MW · H100-equivalent 351,692 · type: AI Training · status: Operational · AI hardware owner `org-google`.
Estimate flag: includes estimates.

Record notes: Total capital cost (2025 USD): $12.84B | Epoch estimation methods: our cooling equipment power model, chip availability, chip efficiency.. · Source: https://epoch.ai/data/data-centers

Detail page: https://scrutica.com/facilities/fac-epoch-google-new-albany.

#### CoreWeave Denton TX (`fac-epoch-coreweave-denton-tx`)

Denton, Texas, United States — 366.8 MW · IT load 262 MW · H100-equivalent 252,652 · type: AI Training · status: Operational · AI hardware owner `org-coreweave`.
Estimate flag: not marked estimated.

Record notes: Total capital cost (2025 USD): $9.93B | Epoch estimation methods: company disclosures.. · Source: https://epoch.ai/data/data-centers

Detail page: https://scrutica.com/facilities/fac-epoch-coreweave-denton-tx.

#### Google Columbus (`fac-epoch-google-columbus`)

Columbus, Ohio, United States — 364 MW · IT load 303 MW · H100-equivalent 416,877 · type: AI Training · status: Operational · AI hardware owner `org-google`.
Estimate flag: includes estimates.

Record notes: Total capital cost (2025 USD): $11.48B | Epoch estimation methods: our cooling equipment power model, regulatory filings, chip efficiency, chip availability.. · Source: https://epoch.ai/data/data-centers

Detail page: https://scrutica.com/facilities/fac-epoch-google-columbus.

#### Amazon Madison Mega Site (`fac-epoch-amazon-madison-mega-site`)

Canton, Mississippi, United States — 341 MW · IT load 284 MW · H100-equivalent 214,348 · type: AI Training · status: Operational · AI hardware owner `org-amazon`.
Estimate flag: includes estimates.

Record notes: Total capital cost (2025 USD): $10.76B | Epoch estimation methods: regulatory filings, chip efficiency, third-party reporting.. · Source: https://epoch.ai/data/data-centers

Detail page: https://scrutica.com/facilities/fac-epoch-amazon-madison-mega-site.

#### DayOne Nusajaya (`fac-dayone-johor-nusajaya`)

Johor Bahru, Johor, Malaysia — 336 MW · IT load 240 MW · H100-equivalent 281,455 · type: AI Training · status: Operational.
Estimate flag: includes estimates.

Record notes: Total capital cost (2025 USD): $9.09B | Epoch estimation methods: our cooling equipment power model, company disclosures, chip availability, chip efficiency.. · Source: https://epoch.ai/data/data-centers

Detail page: https://scrutica.com/facilities/fac-dayone-johor-nusajaya.

#### QTS Richmond 1 (`fac-epoch-qts-richmond-1`)

Sandston, Virginia, United States — 333 MW · IT load 238 MW · H100-equivalent 242,546 · type: AI Training · status: Operational.
Estimate flag: includes estimates.

Record notes: Total capital cost (2025 USD): $9.02B | Epoch estimation methods: company disclosures, regulatory filings, chip availability, chip efficiency.. · Source: https://epoch.ai/data/data-centers

Detail page: https://scrutica.com/facilities/fac-epoch-qts-richmond-1.

#### Google Council Bluffs (East) (`fac-epoch-google-council-bluffs-east`)

Council Bluffs, Iowa, United States — 284 MW · IT load 237 MW · H100-equivalent 346,639 · type: AI Training · status: Operational · AI hardware owner `org-google`.
Estimate flag: includes estimates.

Record notes: Total capital cost (2025 USD): $8.98B | Epoch estimation methods: our cooling equipment power model, chip availability, chip efficiency.. · Source: https://epoch.ai/data/data-centers

Detail page: https://scrutica.com/facilities/fac-epoch-google-council-bluffs-east.

#### Google Omaha (`fac-epoch-google-omaha`)

Omaha, Nebraska, United States — 284 MW · IT load 237 MW · H100-equivalent 275,391 · type: AI Training · status: Operational · AI hardware owner `org-google`.
Estimate flag: includes estimates.

Record notes: Total capital cost (2025 USD): $8.98B | Epoch estimation methods: our cooling equipment power model, regulatory filings, chip availability, chip efficiency.. · Source: https://epoch.ai/data/data-centers

Detail page: https://scrutica.com/facilities/fac-epoch-google-omaha.

#### Google Papillion (`fac-epoch-google-papillion`)

Papillion, Nebraska, United States — 284 MW · IT load 237 MW · H100-equivalent 80,343 · type: AI Training · status: Operational · AI hardware owner `org-google`.
Estimate flag: includes estimates.

Record notes: Total capital cost (2025 USD): $8.98B | Epoch estimation methods: our cooling equipment power model, regulatory filings, chip availability, chip efficiency.. · Source: https://epoch.ai/data/data-centers

Detail page: https://scrutica.com/facilities/fac-epoch-google-papillion.

#### Huawei Horinger (`fac-epoch-huawei-horinger`)

Hohhot, Inner Mongolia, China — 278 MW · IT load 241.8 MW · H100-equivalent 123,799 · type: AI Training · status: Operational · AI hardware owner `org-huawei`.
Estimate flag: includes estimates.

Record notes: Total capital cost (2025 USD): $9.16B | Distinct campus from fac-huawei-ulanqab (Huawei Cloud Ulanqab DC): Horinger New Area is under Hohhot, ~133-145 km from Ulanqab city (separate prefecture-level cities); Horinger groundbreaking 2023-10 (Xinhua 2023-11-29) vs Ulanqab campus construction from 2017. Convergent tier-3 evidence; verified 2026-07-10. | Epoch estimation methods: our cooling equipment power model, company disclosures, chip availability, chip efficiency.. · Source: https://epoch.ai/data/data-centers

Detail page: https://scrutica.com/facilities/fac-epoch-huawei-horinger.

#### Amazon Ridgeland (`fac-epoch-amazon-ridgeland`)

Ridgeland, Mississippi, United States — 273 MW · IT load 228 MW · H100-equivalent 171,298 · type: AI Training · status: Operational · AI hardware owner `org-amazon`.
Estimate flag: includes estimates.

Record notes: Total capital cost (2025 USD): $8.64B | Epoch estimation methods: regulatory filings, chip efficiency, third-party reporting.. · Source: https://epoch.ai/data/data-centers

Detail page: https://scrutica.com/facilities/fac-epoch-amazon-ridgeland.

#### Microsoft Goodyear (`fac-epoch-microsoft-goodyear`)

Goodyear, Arizona, United States — 263 MW · IT load 202 MW · H100-equivalent 205,154 · type: AI Training · status: Operational · AI hardware owner `org-microsoft`.
Estimate flag: includes estimates.

Record notes: Total capital cost (2025 USD): $7.65B | Epoch estimation methods: our cooling equipment power model, company disclosures, chip availability, chip efficiency.. · Source: https://epoch.ai/data/data-centers

Detail page: https://scrutica.com/facilities/fac-epoch-microsoft-goodyear.

#### QTS Richmond 2 (`fac-epoch-qts-richmond-2`)

Sandston, Virginia, United States — 252 MW · IT load 180 MW · H100-equivalent 182,920 · type: AI Training · status: Operational.
Estimate flag: includes estimates.

Record notes: Total capital cost (2025 USD): $6.82B | Epoch estimation methods: company disclosures, regulatory filings, chip availability, chip efficiency.. · Source: https://epoch.ai/data/data-centers

Detail page: https://scrutica.com/facilities/fac-epoch-qts-richmond-2.

#### VNET Bayin Ulanqab (`fac-epoch-vnet-bayin-ulanqab`)

Ulanqab, Inner Mongolia, China — 252 MW · IT load 221 MW · H100-equivalent 122,789 · type: AI Training · status: Operational · AI hardware owner `org-vnet-group`.
Estimate flag: includes estimates.

Record notes: Total capital cost (2025 USD): $8.37B | Distinct campus from fac-alibaba-ulanqab (Alibaba Cloud Ulanqab Super DC): VNET's own 2024-04-18 press release sites this project in the Bayin district (巴音片区) of Chahar High-Tech Zone (vnet.com/portal/article/index/cid/14/id/879.html, tier 1); Alibaba's campus sits in the separately-named Yiwu sub-district. Verified 2026-07-10. | Epoch estimation methods: our cooling equipment power model, chip availability, chip efficiency.. · Source: https://epoch.ai/data/data-centers

Detail page: https://scrutica.com/facilities/fac-epoch-vnet-bayin-ulanqab.

#### Meta Rosemount (`fac-epoch-meta-rosemount`)

Rosemount, Minnesota, United States — 241 MW · IT load 178 MW · H100-equivalent 215,765 · type: AI Training · status: Operational · AI hardware owner `org-meta`.
Estimate flag: includes estimates.

Record notes: Total capital cost (2025 USD): $6.74B | Epoch estimation methods: our cooling equipment power model, structural similarity to known buildings, chip availability, chip efficiency.. · Source: https://epoch.ai/data/data-centers

Detail page: https://scrutica.com/facilities/fac-epoch-meta-rosemount.

#### Meta Jeffersonville (`fac-epoch-meta-jeffersonville`)

Jeffersonville, Indiana, United States — 241 MW · IT load 178 MW · H100-equivalent 203,132 · type: AI Training · status: Operational · AI hardware owner `org-meta`.
Estimate flag: includes estimates.

Record notes: Total capital cost (2025 USD): $6.74B | Epoch estimation methods: our cooling equipment power model, structural similarity to known buildings, chip availability, chip efficiency.. · Source: https://epoch.ai/data/data-centers

Detail page: https://scrutica.com/facilities/fac-epoch-meta-jeffersonville.

#### Microsoft Project Osmium (`fac-epoch-microsoft-project-osmium`)

Cumming, Iowa, United States — 228 MW · IT load 190 MW · H100-equivalent 151,591 · type: AI Training · status: Operational · AI hardware owner `org-microsoft`.
Estimate flag: includes estimates.

Record notes: Total capital cost (2025 USD): $7.20B | Epoch estimation methods: our cooling equipment power model, company disclosures, chip availability, chip efficiency.. · Source: https://epoch.ai/data/data-centers

Detail page: https://scrutica.com/facilities/fac-epoch-microsoft-project-osmium.

#### Google Storey County (`fac-epoch-google-storey-county`)

Clark, Nevada, United States — 214 MW · IT load 161 MW · H100-equivalent 153,612 · type: AI Training · status: Operational · AI hardware owner `org-google`.
Estimate flag: includes estimates.

Record notes: Total capital cost (2025 USD): $6.10B | Epoch estimation methods: our cooling equipment power model, chip availability, chip efficiency.. · Source: https://epoch.ai/data/data-centers

Detail page: https://scrutica.com/facilities/fac-epoch-google-storey-county.

#### Alibaba Zhangbei (`fac-epoch-alibaba-zhangbei`)

Zhangbei, Hebei, China — 203 MW · IT load 169 MW · H100-equivalent 132,895 · type: AI Training · status: Operational · AI hardware owner `org-alibaba`.
Estimate flag: includes estimates.

Record notes: Total capital cost (2025 USD): $6.40B | Epoch estimation methods: our cooling equipment power model, chip efficiency.. · Source: https://epoch.ai/data/data-centers

Detail page: https://scrutica.com/facilities/fac-epoch-alibaba-zhangbei.

#### QTS Richmond 3 (`fac-epoch-qts-richmond-3`)

Sandston, Virginia, United States — 202 MW · IT load 144 MW · H100-equivalent 174,835 · type: AI Training · status: Operational.
Estimate flag: includes estimates.

Record notes: Total capital cost (2025 USD): $5.46B | Epoch estimation methods: company disclosures, regulatory filings, chip availability, chip efficiency.. · Source: https://epoch.ai/data/data-centers

Detail page: https://scrutica.com/facilities/fac-epoch-qts-richmond-3.

#### Coreweave Helios (`fac-epoch-coreweave-helios`)

Afton, Texas, United States — 200 MW · IT load 132 MW · H100-equivalent 159,551 · type: AI Training · status: Operational · AI hardware owner `org-coreweave`.
Estimate flag: includes estimates.

Record notes: Total capital cost (2025 USD): $5.00B | Epoch estimation methods: our cooling equipment power model, company disclosures, regulatory filings.. · Source: https://epoch.ai/data/data-centers

Detail page: https://scrutica.com/facilities/fac-epoch-coreweave-helios.

#### Meta Montgomery (`fac-epoch-meta-montgomery`)

Montgomery, Alabama, United States — 199 MW · IT load 153 MW · H100-equivalent 174,330 · type: AI Training · status: Operational · AI hardware owner `org-meta`.
Estimate flag: includes estimates.

Record notes: Total capital cost (2025 USD): $5.80B | Epoch estimation methods: our cooling equipment power model, structural similarity to known buildings, chip availability, chip efficiency.. · Source: https://epoch.ai/data/data-centers

Detail page: https://scrutica.com/facilities/fac-epoch-meta-montgomery.

#### Meta Kuna (`fac-epoch-meta-kuna`)

Kuna, Idaho, United States — 198 MW · IT load 152 MW · H100-equivalent 173,319 · type: AI Training · status: Operational · AI hardware owner `org-meta`.
Estimate flag: includes estimates.

Record notes: Total capital cost (2025 USD): $5.76B | Epoch estimation methods: our cooling equipment power model, structural similarity to known buildings, chip availability, chip efficiency.. · Source: https://epoch.ai/data/data-centers

Detail page: https://scrutica.com/facilities/fac-epoch-meta-kuna.

#### Meta Temple (`fac-epoch-meta-temple`)

Temple, Texas, United States — 198 MW · IT load 152 MW · H100-equivalent 173,247 · type: AI Training · status: Operational · AI hardware owner `org-meta`.
Estimate flag: includes estimates.

Record notes: Total capital cost (2025 USD): $5.76B | Epoch estimation methods: our cooling equipment power model, company disclosures, chip availability, chip efficiency.. · Source: https://epoch.ai/data/data-centers

Detail page: https://scrutica.com/facilities/fac-epoch-meta-temple.

#### Google The Dalles (`fac-epoch-google-the-dalles`)

The Dalles, Oregon, United States — 185 MW · IT load 154 MW · H100-equivalent 201,616 · type: AI Training · status: Operational · AI hardware owner `org-google`.
Estimate flag: includes estimates.

Record notes: Total capital cost (2025 USD): $5.83B | Epoch estimation methods: our cooling equipment power model, chip availability, chip efficiency.. · Source: https://epoch.ai/data/data-centers

Detail page: https://scrutica.com/facilities/fac-epoch-google-the-dalles.

#### Google Lincoln (`fac-epoch-google-lincoln`)

Lincoln, Nebraska, United States — 170 MW · IT load 141 MW · H100-equivalent 300,656 · type: AI Training · status: Operational · AI hardware owner `org-google`.
Estimate flag: includes estimates.

Record notes: Total capital cost (2025 USD): $5.34B | Epoch estimation methods: our cooling equipment power model, chip availability, chip efficiency.. · Source: https://epoch.ai/data/data-centers

Detail page: https://scrutica.com/facilities/fac-epoch-google-lincoln.

#### Google Lancaster (`fac-epoch-google-lancaster`)

Lancaster, Ohio, United States — 164 MW · IT load 137 MW · H100-equivalent 130,368 · type: AI Training · status: Operational · AI hardware owner `org-google`.
Estimate flag: includes estimates.

Record notes: Total capital cost (2025 USD): $5.19B | Epoch estimation methods: our cooling equipment power model, chip availability, chip efficiency.. · Source: https://epoch.ai/data/data-centers

Detail page: https://scrutica.com/facilities/fac-epoch-google-lancaster.

#### Google Mesa (`fac-epoch-google-mesa`)

United States — 126 MW · IT load 90 MW · H100-equivalent 125,821 · type: AI Training · status: Operational · AI hardware owner `org-google`.
Estimate flag: includes estimates.

Record notes: Total capital cost (2025 USD): $3.41B | Epoch estimation methods: our cooling equipment power model, chip availability, chip efficiency.. · Source: https://epoch.ai/data/data-centers

Detail page: https://scrutica.com/facilities/fac-epoch-google-mesa.

#### Google Waltham Cross (`fac-epoch-google-waltham-cross`)

Cheshunt, Hertfordshire, United Kingdom — 124 MW · IT load 88 MW · H100-equivalent 123,799 · type: AI Training · status: Operational · AI hardware owner `org-google`.
Estimate flag: includes estimates.

Record notes: Total capital cost (2025 USD): $3.33B | Epoch estimation methods: our cooling equipment power model, chip availability, chip efficiency.. · Source: https://epoch.ai/data/data-centers

Detail page: https://scrutica.com/facilities/fac-epoch-google-waltham-cross.

#### Google Midlothian (`fac-epoch-google-midlothian`)

Midlothian, Texas, United States — 123 MW · IT load 103 MW · H100-equivalent 199,090 · type: AI Training · status: Operational · AI hardware owner `org-google`.
Estimate flag: includes estimates.

Record notes: Total capital cost (2025 USD): $3.90B | Epoch estimation methods: our cooling equipment power model, chip availability, chip efficiency, regulatory filings.. · Source: https://epoch.ai/data/data-centers

Detail page: https://scrutica.com/facilities/fac-epoch-google-midlothian.

#### STACK Infrastructure NVA02 (`fac-epoch-stack-infrastructure-nva02`)

Manassas, Virginia, United States — 118 MW · IT load 85 MW · H100-equivalent 61,141 · type: AI Training · status: Operational.
Estimate flag: includes estimates.

Record notes: Total capital cost (2025 USD): $3.22B | Epoch estimation methods: our cooling equipment power model, chip availability, chip efficiency, company disclosures.. · Source: https://epoch.ai/data/data-centers

Detail page: https://scrutica.com/facilities/fac-epoch-stack-infrastructure-nva02.

#### CoreWeave Chester VA (`fac-epoch-coreweave-chester-va`)

Chester, Virginia, United States — 114 MW · IT load 82 MW · H100-equivalent 142,496 · type: AI Training · status: Operational · AI hardware owner `org-coreweave`.
Estimate flag: not marked estimated.

Record notes: Total capital cost (2025 USD): $3.11B | Epoch estimation methods: company disclosures.. · Source: https://epoch.ai/data/data-centers

Detail page: https://scrutica.com/facilities/fac-epoch-coreweave-chester-va.

### Records not operating, retired or without a status

#### 1 Gig Data Center East Fishkill, NY (`gridstatus-nyiso-1738`)

NY, United States — 1000 MW · type: Hyperscale DC · status: Announced.
Estimate flag: not marked estimated.

Record notes: Queue ID: 1738. Gen type: Load. Proposed commercial operation date not recorded.. · Source: https://www.nyiso.com/documents/20142/1407078/NYISO-Interconnection-Queue.xlsx

Detail page: https://scrutica.com/facilities/gridstatus-nyiso-1738.

#### Micron Fab 2 (`gridstatus-nyiso-1627`)

NY, United States — 576 MW · type: Memory Fab · status: Announced · recorded in owner field `org-micron`.
Estimate flag: not marked estimated.

Record notes: NYISO queue request 1627, Micron Fab 2, lists a peak load of 576 MW and end use M-CH (Microchip fabrication) in the workbook read on 8 September 2026 (Load Projects, row 21; Load Project Tracking, row 47). Developer: Micron New York Semiconductor Manufacturing LLC. https://www.nyiso.com/documents/20142/1407078/NYISO-Interconnection-Queue.xlsx NIST identifies the first two Clay fabs as DRAM manufacturing. https://www.nist.gov/chips/micron-new-york-clay. · Source: https://www.nyiso.com/documents/20142/1407078/NYISO-Interconnection-Queue.xlsx

Detail page: https://scrutica.com/facilities/gridstatus-nyiso-1627.

#### North East Data LLC Data Center (`gridstatus-nyiso-1741`)

NY, United States — 500 MW · type: Hyperscale DC · status: Announced.
Estimate flag: not marked estimated.

Record notes: Queue ID: 1741. Gen type: Load. Proposed commercial operation date not recorded.. · Source: https://www.nyiso.com/documents/20142/1407078/NYISO-Interconnection-Queue.xlsx

Detail page: https://scrutica.com/facilities/gridstatus-nyiso-1741.

#### Arsenal Data Site 1000 (`gridstatus-nyiso-1730`)

NY, United States — 467 MW · type: Hyperscale DC · status: Announced.
Estimate flag: not marked estimated.

Record notes: Queue ID: 1730. Gen type: Load. Proposed commercial operation date not recorded.. · Source: https://www.nyiso.com/documents/20142/1407078/NYISO-Interconnection-Queue.xlsx

Detail page: https://scrutica.com/facilities/gridstatus-nyiso-1730.

#### Data & Technology Campus (`gridstatus-nyiso-1726`)

NY, United States — 300 MW · type: Hyperscale DC · status: Announced.
Estimate flag: not marked estimated.

Record notes: Queue ID: 1726. Gen type: Load. Proposed commercial operation date not recorded.. · Source: https://www.nyiso.com/documents/20142/1407078/NYISO-Interconnection-Queue.xlsx

Detail page: https://scrutica.com/facilities/gridstatus-nyiso-1726.

#### New York State Artificial Intelligence Data Center (`gridstatus-nyiso-1731`)

NY, United States — 300 MW · type: Hyperscale DC · status: Announced.
Estimate flag: not marked estimated.

Record notes: Queue ID: 1731. Gen type: Load. Proposed commercial operation date not recorded.. · Source: https://www.nyiso.com/documents/20142/1407078/NYISO-Interconnection-Queue.xlsx

Detail page: https://scrutica.com/facilities/gridstatus-nyiso-1731.

#### Lake Mariner Data II (`gridstatus-nyiso-1670`)

NY, United States — 250 MW · type: Hyperscale DC · status: Announced.
Estimate flag: not marked estimated.

Record notes: Queue ID: 1670. Gen type: Load. Proposed commercial operation date not recorded.. · Source: https://www.nyiso.com/documents/20142/1407078/NYISO-Interconnection-Queue.xlsx

Detail page: https://scrutica.com/facilities/gridstatus-nyiso-1670.

#### Wulf Compute Data Center II (`gridstatus-nyiso-1732`)

NY, United States — 250 MW · type: Hyperscale DC · status: Announced.
Estimate flag: not marked estimated.

Record notes: Queue ID: 1732. Gen type: Load. Proposed commercial operation date not recorded.. · Source: https://www.nyiso.com/documents/20142/1407078/NYISO-Interconnection-Queue.xlsx

Detail page: https://scrutica.com/facilities/gridstatus-nyiso-1732.

#### Pontoon Bridge Road Data Center (`gridstatus-nyiso-1745`)

NY, United States — 250 MW · type: Hyperscale DC · status: Announced.
Estimate flag: not marked estimated.

Record notes: Queue ID: 1745. Gen type: Load. Proposed commercial operation date not recorded.. · Source: https://www.nyiso.com/documents/20142/1407078/NYISO-Interconnection-Queue.xlsx

Detail page: https://scrutica.com/facilities/gridstatus-nyiso-1745.

#### Broome County Tech Park (`gridstatus-nyiso-1752`)

NY, United States — 250 MW · type: Hyperscale DC · status: Announced.
Estimate flag: not marked estimated.

Record notes: Queue ID: 1752. Gen type: Load. Proposed commercial operation date not recorded.. · Source: https://www.nyiso.com/documents/20142/1407078/NYISO-Interconnection-Queue.xlsx

Detail page: https://scrutica.com/facilities/gridstatus-nyiso-1752.

#### Arsenal Data Site 250 (`gridstatus-nyiso-1728`)

NY, United States — 233 MW · type: Hyperscale DC · status: Announced.
Estimate flag: not marked estimated.

Record notes: Queue ID: 1728. Gen type: Load. Proposed commercial operation date not recorded.. · Source: https://www.nyiso.com/documents/20142/1407078/NYISO-Interconnection-Queue.xlsx

Detail page: https://scrutica.com/facilities/gridstatus-nyiso-1728.

#### Arsenal Data Site 500 (`gridstatus-nyiso-1729`)

NY, United States — 233 MW · type: Hyperscale DC · status: Announced.
Estimate flag: not marked estimated.

Record notes: Queue ID: 1729. Gen type: Load. Proposed commercial operation date not recorded.. · Source: https://www.nyiso.com/documents/20142/1407078/NYISO-Interconnection-Queue.xlsx

Detail page: https://scrutica.com/facilities/gridstatus-nyiso-1729.

#### St Lawrence Data and Agricultural Center (`gridstatus-nyiso-1213`)

NY, United States — 200 MW · type: Hyperscale DC · status: Announced.
Estimate flag: not marked estimated.

Record notes: Queue ID: 1213. Gen type: Load. Proposed commercial operation date not recorded.. · Source: https://www.nyiso.com/documents/20142/1407078/NYISO-Interconnection-Queue.xlsx

Detail page: https://scrutica.com/facilities/gridstatus-nyiso-1213.

#### Proposed Datacenters at 450 Broadway, Buchanan, NY, 10511 (`gridstatus-nyiso-1717`)

NY, United States — 200 MW · type: Hyperscale DC · status: Announced.
Estimate flag: not marked estimated.

Record notes: Queue ID: 1717. Gen type: Load. Proposed commercial operation date not recorded.. · Source: https://www.nyiso.com/documents/20142/1407078/NYISO-Interconnection-Queue.xlsx

Detail page: https://scrutica.com/facilities/gridstatus-nyiso-1717.

#### Greenidge 200 MW Data Center Project (`gridstatus-nyiso-1725`)

NY, United States — 200 MW · type: Hyperscale DC · status: Announced.
Estimate flag: not marked estimated.

Record notes: Queue ID: 1725. Gen type: Load. Proposed commercial operation date not recorded.. · Source: https://www.nyiso.com/documents/20142/1407078/NYISO-Interconnection-Queue.xlsx

Detail page: https://scrutica.com/facilities/gridstatus-nyiso-1725.

#### Globe Digital Holdings - 1 (`gridstatus-nyiso-1747`)

NY, United States — 200 MW · type: Hyperscale DC · status: Announced.
Estimate flag: not marked estimated.

Record notes: Queue ID: 1747. Gen type: Load. Proposed commercial operation date not recorded.. · Source: https://www.nyiso.com/documents/20142/1407078/NYISO-Interconnection-Queue.xlsx

Detail page: https://scrutica.com/facilities/gridstatus-nyiso-1747.

#### North Country Data Center (`gridstatus-nyiso-0979`)

NY, United States — 176 MW · type: Hyperscale DC · status: Announced.
Estimate flag: not marked estimated.

Record notes: Queue ID: 0979. Gen type: Load. Proposed commercial operation date not recorded.. · Source: https://www.nyiso.com/documents/20142/1407078/NYISO-Interconnection-Queue.xlsx

Detail page: https://scrutica.com/facilities/gridstatus-nyiso-0979.

#### Cayuga Data (`gridstatus-nyiso-1733`)

NY, United States — 162 MW · type: Hyperscale DC · status: Announced.
Estimate flag: not marked estimated.

Record notes: Queue ID: 1733. Gen type: Load. Proposed commercial operation date not recorded.. · Source: https://www.nyiso.com/documents/20142/1407078/NYISO-Interconnection-Queue.xlsx

Detail page: https://scrutica.com/facilities/gridstatus-nyiso-1733.

#### Niagara Digital Campus (`gridstatus-nyiso-1681`)

NY, United States — 140 MW · type: Hyperscale DC · status: Announced.
Estimate flag: not marked estimated.

Record notes: Queue ID: 1681. Gen type: Load. Proposed commercial operation date not recorded.. · Source: https://www.nyiso.com/documents/20142/1407078/NYISO-Interconnection-Queue.xlsx

Detail page: https://scrutica.com/facilities/gridstatus-nyiso-1681.

#### Massena Load (`gridstatus-nyiso-0909`)

NY, United States — 100 MW · type: Hyperscale DC · status: Decommissioned.
Estimate flag: not marked estimated.

Record notes: Queue ID: 0909. Gen type: Load. Proposed commercial operation date not recorded.. · Source: https://www.nyiso.com/documents/20142/1407078/NYISO-Interconnection-Queue.xlsx

Detail page: https://scrutica.com/facilities/gridstatus-nyiso-0909.


---


## Reference: per-country operating facility power

These figures come from the facility reference extract bundled with this guide, grouped by country and ordered by operating MW. Operating MW sums the positive power figures on operational records only. Figures on records under construction or expanding, announced or permitted, decommissioned or cancelled, or without a status are listed separately by status and never added to it, because they describe planned or past plant. All facility types contribute, including manufacturing plants; countries outside this file's coverage are absent.

Two holds the site applies cannot be applied to this extract: a campus build-out figure filed on an operating record, and an earlier phase of a cluster whose later phase includes it. Both can sit inside the operating MW below; each country's page (scrutica.com/country/<code>) prints the operating figure with them held out. Campus envelopes and individual project values may also overlap, so these sums do not establish installed computing capacity. Operators are ranked by the operating MW attributed to them within this selection. A linked sovereign-program announcement is a separate financial record and need not fund the facilities counted here.

### United States

- Operating power: 13,367.6 MW across 47 operational facility records
- Not operating, never added above: 6,025 MW announced or permitted (21 records) · 170 MW decommissioned or cancelled (3 records)
- Tracked sovereign-AI program: [United States: NAIRR / CHIPS Act AI Components / Stargate Project](https://scrutica.com/sovereign-ai/program/sovereign-us)
- Drill-down: https://scrutica.com/entities?country=US

### China

- Operating power: 733 MW across 3 operational facility records
- Tracked sovereign-AI program: [China: National AI Programs / Big Fund III / Eastern Data Western Computing](https://scrutica.com/sovereign-ai/program/sovereign-cn)
- Drill-down: https://scrutica.com/entities?country=CN

### Malaysia

- Operating power: 336 MW across 1 operational facility record
- Tracked sovereign-AI program: [Malaysia: Public and private AI cloud initiatives](https://scrutica.com/sovereign-ai/program/sovereign-my)
- Drill-down: https://scrutica.com/entities?country=MY

### United Kingdom

- Operating power: 124 MW across 1 operational facility record
- Tracked sovereign-AI program: [United Kingdom: AI Opportunities Action Plan / UK Compute Roadmap / AIRR / Sovereign AI Unit / AI Hardware Plan](https://scrutica.com/sovereign-ai/program/sovereign-gb)
- Drill-down: https://scrutica.com/entities?country=GB

### Portugal

- Operating power: 40 MW across 1 operational facility record
- Drill-down: https://scrutica.com/entities?country=PT

---


## Reference: reading accelerator specifications

The public register uses the source-backed static catalogue in `src/lib/data/chip-classifications.ts`. The [chip register](https://scrutica.com/export-controls/chips) records precision-specific throughput, memory capacity and bandwidth, hardware families, and China-access classifications. It has no per-chip detail routes.

**Match the units and hardware scope.** Throughput depends on precision and on whether the specification assumes sparsity. Compare dense BF16 with dense BF16, and check whether a value describes an individual accelerator, a board or a complete system. Memory capacity and bandwidth impose additional workload constraints that a peak FLOP/s figure cannot represent.

**Keep product identity explicit.** Variants within an architecture can have different throughput, memory and interconnect specifications. A system name can combine host CPUs and several accelerators. Use the exact catalogue entry and its source when converting a hardware count into capacity.

**Read announcement status.** Preliminary vendor specifications and proposed system-release dates remain qualified by their source. For example, NVIDIA's HGX page supplies preliminary Rubin specifications; its published throughput table separates the numerical precision and dense or sparse operation ([NVIDIA HGX](https://www.nvidia.com/en-us/data-center/hgx/)). The following roadmap statement comes from NVIDIA's March 2025 announcement:

- **Rubin Ultra** — NVIDIA's announced follow-on Rubin GPU; systems built on it are due in the second half of 2027 ([source](https://blogs.nvidia.com/blog/nvidia-keynote-at-gtc-2025-ai-news-live-updates/)).

**Read the control assessment separately.** The catalogue gives an assessment basis, source and conditions alongside its China-access category. Hardware performance alone cannot determine whether a particular shipment needs a licence. Check the item, destination, end user and applicable rule.

The MCP `scrutica_estimate_flops` tool has a smaller accepted hardware set: H100-SXM5, H200, B100, B200, A100-80GB, MI300X and TPU-v5p. A model's presence in the wider chip catalogue does not make it an accepted input to that tool.

---


## Regional scenarios

The [scenario register](https://scrutica.com/scenarios) separates each analysis's data vintage from its recorded review date. A newer input date leaves the review date unchanged until a review is recorded. The five pages use different methods, so their outputs need to be read in the context of the specific case.

**[Taiwan Strait](https://scrutica.com/scenarios/taiwan-strait).** Four cases consider maritime blockade, a targeted strike, grid failure and an export restriction. This analysis connects the selected disruption assumptions to the supply-chain simulator and compares alternative foundry capacity. The server simulation receives the case’s severity and propagation-decay factor. Its duration describes the case context and buffer comparison; it does not enter the cascade calculation.

**[IRGC missile range](https://scrutica.com/scenarios/iran-threat).** Published weapon-range envelopes are shown alongside recorded regional compute locations. Some locations represent approximate cloud regions. This geographic comparison establishes which recorded points lie within the selected range envelope; it provides no estimate of attack likelihood, targeting accuracy or defensive performance.

**[Saudi grid](https://scrutica.com/scenarios/abqaiq-2).** The analysis examines links between petroleum infrastructure, grid zones and data centers. The analysis distinguishes assigned substations, shared grid zones and geographic proximity. Some substation assignments are inferred from corridor proximity. Actual supply arrangements and backup would determine whether a disruption affects both sites.

**[Tokyo earthquake](https://scrutica.com/scenarios/tokyo-earthquake).** The analysis combines seismic scenarios with memory-fab, materials and equipment records. It includes a separate assessment of Japanese inputs to Korean memory production. The proposed propagation to SK Hynix extends beyond the Japanese-input relationships recorded in the graph and has no inventory-based depletion timescale. Hazard assumptions and source dates belong with the affected-facility comparison.

**[Cable chokepoints](https://scrutica.com/scenarios/south-china-sea).** The analysis examines cable routes, landing stations, chokepoint zones and historical incidents on the Asia–Europe corridor. Capacity-loss estimates use the design capacity of affected cables relative to the stated corridor total. Design capacity and rerouting assumptions do not establish the traffic actually carried or spare capacity available during an outage.

The MCP `scrutica_get_scenario` tool returns selected evidence from the same source modules, including assumptions, facilities, fabs, cables or weapon systems as applicable. Lists can be capped; accompanying totals state their complete source counts. Simulation outputs remain on the interactive pages.

Implementation: the scenario registry in `src/lib/scenarios/registry.ts` and `buildScenarioSubstrate` in `src/app/api/mcp/[transport]/scenario-substrate.ts`.

---


## Connecting an agent

The public MCP endpoint is `https://scrutica.com/api/mcp`. It uses Streamable HTTP, protocol version 2025-11-25, and requires no API key. Client configuration differs by application; use the [setup instructions](https://scrutica.com/agents) for the relevant client.

The server exposes 10 tools:

- scrutica_search — Search Scrutica.
- scrutica_get_facility — Facility record.
- scrutica_get_company — Company record.
- scrutica_query_export_controls — Export-control designations.
- scrutica_entity_list_changes — Entity List change log.
- scrutica_get_supply_chain — Supply-chain edges.
- scrutica_get_sovereign_program — Sovereign AI program.
- scrutica_get_scenario — Risk scenario.
- scrutica_estimate_flops — FLOP estimator.
- scrutica_get_methodology — Methodology reference.

The [MCP reference](https://scrutica.com/api/mcp/llms.txt) documents schemas and example calls. Search for candidate records before requesting details, then use the returned identifiers. Respect each tool's limits and inspect the returned totals before treating a bounded list as complete. Scenario evidence and throughput estimates have the limits described in the method sections above.

**Freshness.** The export-control and Entity List change tools return `regime_status`. Read its last Federal Register check date alongside the latest recorded addition or removal, notices not yet reflected in the dataset, and the sources and time window of the listed rules. A newest designation date describes an addition; it omits later removals and says nothing by itself about when the source was checked. A stale or unverified check leaves current membership unresolved.

**Request limits.** The MCP limit is 60 requests per minute per IP. Public JSON endpoints have route-specific limits. On HTTP 429, follow the response's `Retry-After` header.

**Other access.** The [OpenAPI specification](https://scrutica.com/api/openapi.json) documents public JSON endpoints. Appending `.md` to a public page URL requests its Markdown version, for example [methodology.md](https://scrutica.com/methodology.md). Facility Markdown has a dedicated rendering path; other pages are converted from their HTML. The [site index](https://scrutica.com/llms.txt) provides links for discovery.

[Query Builder](https://scrutica.com/query) accepts natural-language questions, selects tools and synthesizes their results. Its default model chain is GPT-6 Sol (failing over to GPT-5.6 Terra then Claude Opus 4.8 then GPT-5.5); deployment configuration can override the primary model. Inspect cited records and calculations before relying on an answer. Its tool collection is separate from the public MCP catalog, and its [privacy policy](https://scrutica.com/privacy) explains how submitted questions are handled.

---


## Published datasets and downloads

The [dataset directory](https://scrutica.com/datasets) lists the releases available for download. The current manifest records data version 2026.4, released 2026-09-25. Documentation revised: 2026-09-25.

- [sovereign-ai-index](https://scrutica.com/datasets/sovereign-ai-index): 36 records; CSV SHA-256 52c2385a010922b05759600ddd7867f6199f651b4bec91d2bfb546b9f847b844.
- [compute-cost-index](https://scrutica.com/datasets/compute-cost-index): 508 records; CSV SHA-256 9d4ad130620644787a23f75d6b4e8b84ffee19e38782383a8d3ba80dc07180d5.

Each published bundle includes CSV data, a schema, a data dictionary, methodology and Croissant metadata. The manifest's SHA-256 identifies the CSV bytes. Documentation can be corrected separately from the data release; preserve both dates when describing a revised dictionary or methodology.

The Sovereign AI Index contains program funding and source fields. The Compute Cost Index contains cloud-price observations and their peak-capacity normalization. These published bundles use CC BY 4.0. The full supply-chain collection is outside the public release manifest; its presence elsewhere in this guide does not make it downloadable.

Use the [citation page](https://scrutica.com/datasets/cite) for the release's citation and DOI status. A DOI is available only when the citation metadata provides one. Public API responses and on-page exports have their own field selections and limits, which can differ from a packaged release.

For updates, the [Entity List RSS feed](https://scrutica.com/export-controls/changes/feed.xml) groups changes by notice. The [Record](https://scrutica.com/changelog) reports additions and updates to its covered collections. Their dates describe the events they record; a dataset download remains the version stated in its manifest.

---


## Data sources and methodological references

Scrutica combines imported research datasets with corporate disclosures, public registries and individual project sources. Epoch AI infrastructure, hardware and model datasets and CSET supply-chain records are among the imported collections. RAND's compute-infrastructure work supplies a methodological reference. These roles differ: citing a research paper does not imply that its dataset is imported.

The [About page](https://scrutica.com/about) describes source coverage; [Related Work](https://scrutica.com/related-work) explains the role of named research organizations. Individual records and dataset dictionaries identify the sources used for a particular field. Licensed-source records have separate public-access restrictions, as described above.

---


## Contact

Send corrections or methodological questions to davidgringras@hsph.harvard.edu. Include the Scrutica URL, the specific field or calculation, and the source and date that support the correction. Published amendments appear in the [corrections register](https://scrutica.com/corrections).

[About](https://scrutica.com/about) describes the project and its author. The [short index](https://scrutica.com/llms.txt) and [agent manual](https://scrutica.com/agents) provide the current navigation and connection instructions.

---

*End of reference. Coverage snapshot: 2026-10-03; snapshot generated at 2026-10-03T15:26:31.775Z.*