NCI Australia Gadi

TypeAI TrainingStatusOperationalWhereCanberra ACT 2601, AUIndexed fromEpoch AI GPU Clusters

Operational; National Computational Infrastructure Australia is recorded as owning the AI hardware inside this site

1/3site roles identified0 site relationships0 findings in the company chain

Date in site records: Jan 23, 2020. Site data: Epoch AI GPU Clusters.

Ownership and operation

Plate I. Site owner, operator and hardware ownerscrutica.com · data to Jan 23, 2020 (derived)

AI hardware owner · per catalog

National Computational Infrastructure Australia parent ownership unresolved

Which organisation owns the AI hardware inside it?

NCI Australia GadiSite

AI hardware owner · per catalog National Computational Infrastructure Australia. Epoch AI's field for who owns the AI hardware, which it defines as not necessarily the owner or operator of the facility. source

Companies are grouped by their role at the site, with recorded parents above them. No parent companies were found in the ownership data. Further ownership remains unresolved where labelled. Record source: Epoch AI GPU Clusters; authority tier 2. Each role is drawn only as the relationship its own source establishes, listed above; none implies exclusive control of the full site.

Export controls and government links

The site’s companies and their recorded parents were checked; no export-control designation, sovereign investment or classified jurisdiction was found. Headquarters country is unrecorded for National Computational Infrastructure Australia, leaving its jurisdiction check incomplete.

Capacity

Reading. Power and acceleratorsscrutica.com · data to Jan 23, 2020 (derived)

419 kW

Facility power

Source-reported power.

not recorded

Power for IT equipment

776

Accelerators

Present inventory by partition

withheld NVIDIA H200 141GB does not identify an H200 form factor.

640 × NVIDIA Tesla V100 SXM2
installed (from the facility status)
125 TFLOP/s dense FP16 per accelerator · partition peak 80 PFLOP/s · rate citation
Catalogue row; no partition-level source recorded
120 × NVIDIA H200 141GB
installed (from the facility status)
Held: Variant not stated; rate depends on it. NVIDIA H200 141GB does not identify an H200 form factor. Documented variants: H200 SXM 989.5; H200 NVL 835.5 TFLOP/s; none applied.
NCI HPC systems specification and NCI Annual Report 2025 director's message · 2026-07-20 · tier 1 · Gadi H200 partition: 30 nodes × 4 NVIDIA H200 141GB; the 2025 annual report dates the expansion to 2025 ('30 new GPU nodes, each equipped with four NVIDIA H200 GPUs'). Installation day not stated.
16 × NVIDIA A100 80GB
installed (from the facility status)
312 TFLOP/s dense FP16 per accelerator · partition peak 5 PFLOP/s · rate citation
NCI HPC systems specification and NCMAS 2024 applicant information (2023-11-09) · 2023-11-09 · tier 1 · Gadi A100 partition: 2 DGX A100 nodes × 8 NVIDIA A100 80GB; listed as existing in the NCMAS 2024 guide (2023-11-09). Installation day not stated.

One partition is held, so no facility figure is derived; the held partition’s count and source remain visible and nothing is substituted for its rate. Rates are the cited dense FP16 figure per accelerator; the sum applies the methodology’s utilization and interconnect assumptions and is a theoretical aggregate, not one job’s throughput. The catalogue figures above are the record’s scalar fields (the source’s transcription, or its recorded correction) and are a different basis.

Power comparison

419 kW is at the 15th percentile among the 310 operating AI Training records with a reported power figure. These represent 84% of the 369 operating AI Training records tracked; announced, under-construction and campus build-out figures are not ranked. The scale ends at the largest operating figure in this group, 1.276 GW.

Data quality · 1 flag

The accelerator inventory includes partitions taken from the operator’s own documentation that the catalogue source does not list. The catalogue source lists Gadi's 640 V100 alone. NCI's own specification also lists 30 nodes with four H200 141GB (120, an expansion NCI dates to 2025) and two DGX A100 nodes with eight A100 80GB (16, present by November 2023). The displayed chip count covers all three partitions, each with its own source. No facility throughput or H100-equivalent is derived, because NCI does not state the H200's form factor and the SXM and NVL versions differ in dense FP16 rate. Ownership is not inferred from funding or hosting, and the installation day of the H200 expansion is not stated.

Methods and sources

Ownership roles

Title identifies the legal owner of the site; operation identifies who runs it. AI-hardware ownership follows the catalog’s attribution of the machines inside, whose owner may lease the facility.

Each role requires its own attribution. An operator named in a source supplies evidence of operation only.

Parent companies

Ownership checks start with each company named at the site and follow recorded parent relationships. A company holding several roles is checked once. Its linked company page lists the relationships and their sources; further parents may remain undocumented.

Level 1 identifies a company at the site, level 2 its parent, and each further level another parent. Export-control designations and government investments are assigned to the company concerned.

Missing values

Unrecorded roles remain blank. Findings whose company position is unresolved stay in the list with that position labelled unknown.

Publication limits

Subscription corporate data helps identify parent relationships and sovereign investments. Publication is limited to relationship classes and supporting public sources; licensed records, identifiers and values are withheld.

Epoch AI estimates capacity using satellite imagery, permits and corporate disclosures. Its catalog is the source for this site record.

Related analysis

Browse the facilities directory or search the company directory.