COMPUTE ATLAS Supercomputer supply-chain graph
1153 systems 482 sites Sourced data

Supercomputers/Analysis/Data trust

Analysis 12

How far to trust this data

234 of 1153 systems (20%) are marked verified, meaning two independent organisations directly confirm the same specific fact. The rest are reported: one credible source, cited, and not contradicted. Nothing in the dataset is guessed, but most of it rests on a single source, and this page says where. When every cited source was re-read, 1092 systems had at least one that names the system, and only 40 had readable sources that all failed to.

What the tiers mean. Verified is a claim about independence: two different organisations, not two pages from the same one. Reported is a claim about provenance: a real source, read, that says this. Estimated is ours, and is used only where the method is published. A reported row is not a doubtful row. It is one whose second source we have not yet found, and the house rule is to publish that rather than to omit it or to upgrade it.

  • Verified
  • Reported or estimated
How well each sector is sourced Horizontal bars of the number of systems in each operator sector, split into those marked verified by two independent sources and those that are reported or estimated. Government labs 233 systems Government labs 42 verified 191 reported 0 estimated Government labs 42 verified 191 reported 0 estimated 18% verified Research HPC 568 systems Research HPC 127 verified 441 reported 0 estimated Research HPC 127 verified 441 reported 0 estimated 22% verified Big tech 99 systems Big tech 20 verified 79 reported 0 estimated Big tech 20 verified 79 reported 0 estimated 20% verified AI labs 19 systems AI labs 4 verified 15 reported 0 estimated AI labs 4 verified 15 reported 0 estimated 21% verified Neoclouds 85 systems Neoclouds 14 verified 71 reported 0 estimated Neoclouds 14 verified 71 reported 0 estimated 16% verified Industry 125 systems Industry 22 verified 103 reported 0 estimated Industry 22 verified 103 reported 0 estimated 18% verified Finance 7 systems Finance 2 verified 5 reported 0 estimated Finance 2 verified 5 reported 0 estimated 29% verified Life sciences 17 systems Life sciences 3 verified 14 reported 0 estimated Life sciences 3 verified 14 reported 0 estimated 18% verified 0142284426568
Systems by operator sector, split by confidence tier. A low verified share means a sector is less independently documented, not less accurate; the small sectors move on a single row.
Table view: How well each sector is sourced
SectorSystemsVerifiedReportedEstimatedShare verified
Government labs23342191018%
Research HPC568127441022%
Big tech992079020%
AI labs19415021%
Neoclouds851471016%
Industry12522103018%
Finance725029%
Life sciences17314018%
How this was calculated

Each system carries a confidence tier set when it was added: verified only if two independent organisations confirm the same specific fact and both links are recorded, reported for one credible cited source, estimated for our own bounded figures. The same three tiers are recorded on every component edge, so a system can be reported while one of its parts is verified, and the validator checks each edge for its own second source rather than inheriting the system's.

The source re-check, pnpm verify, fetches every cited source for every system and asks whether the page names the system and mentions the parts and quantities the row claims. It is a name-mention signal, not a fact check: a page can name a system and still disagree with the row, and a paywalled or blocking page tells us nothing either way. It never changes a tier. It proposes a shortlist and a human decides.

By status, by country

StatusSystemsVerifiedReportedEstimatedShare verified
operational809183626023%
decommissioned21840178018%
under construction6245806%
planned41437010%
upgraded23320013%
CountrySystemsVerifiedReportedEstimatedShare verified
USA40662344015%
Japan1063967037%
Germany872562029%
China58850014%
United Kingdom57750012%
France41734017%
Canada2822607%
South Korea28919032%
Australia24321013%
India24816033%
Italy22913041%
Saudi Arabia22418018%

The component graph

The edge table, the part of the dataset that says which supplier's part is in which machine, has 4258 rows: 724 verified (17%), 3532 reported and 2 estimated. Tiers are set edge by edge, so the graph is better sourced in places than the system rows above it.

What the source re-check found

ResultSystemsWhat it means
Corroborated1092At least one cited source, re-read, names the system.
Mixed10One source does not name it, another could not be read.
Unsupported40Every readable cited source no longer names the system.
Unreadable11No cited source could be read, which tells us nothing either way.
Not yet checked0Added since the last full re-check.

Of the 1142 systems with at least one readable source, 96% are corroborated. “Unsupported” is evidence about the citation, not about the fact: a page can be rewritten, moved behind a login or trimmed after we cited it.

Ready to be promoted

27 systems are still marked reported although two different hosts each independently name the system, a part and a quantity. That is the shortlist the re-check proposes for a human to look at. A name-mention cannot establish that two sources agree in the sense verified requires, so none is promoted automatically.

SystemCountryHosts that name it
AI-Farabium Kazakhstan b2b.telecom.kz, top500.org
ARIS Greece doc.aris.grnet.gr, top500.org
ASPIRE 2A+ Singapore nsccsg.github.io, top500.org
Athena Poland oferta-badawcza.agh.edu.pl, top500.org
Barbora Czechia docs.it4i.cz, it4i.cz
Betzy Norway documentation.sigma2.no, top500.org
Bridges-2 USA api.nsf.gov, psc.edu
bwUniCluster 3.0 Germany nemo.uni-freiburg.de, wiki.bwhpc.de
CM-2 USA cisl.ucar.edu, en.wikipedia.org
Cray X-MP USA cisl.ucar.edu, en.wikipedia.org
Cray-2 USA cs.lbl.gov, en.wikipedia.org
DeltaAI USA delta.ncsa.illinois.edu, ncsa.illinois.edu
Google TPU v4 supercomputer USA arxiv.org, cloud.google.com
Govorun Russia hlit.jinr.ru, rscgroup.ru
Guacolda Chile en.wikipedia.org, nlhpc.cl
IARA Brazil sidi.org.br, top500.org
Imperial CX3 United Kingdom bebeez.eu, icl-rcs-user-guide.readthedocs.io
Iridis 5 United Kingdom southampton.ac.uk, top500.org
Jairu Brazil inforchannel.com.br, telesintese.com.br
KEK VPP500/80 Japan scwww.kek.jp, top500.org
Minerva USA labs.icahn.mssm.edu, top500.org
Naver B200 4K Cluster South Korea koreatechtoday.com, koreatimes.co.kr
Project Ceiba USA aws.amazon.com, nvidianews.nvidia.com
Santos Dumont Brazil baguete.com.br, top500.org
Stanage United Kingdom alces-flight.com, docs.hpc.shef.ac.uk
Sunrise United Kingdom top500.org, ukaea.org
Together AI / Hypertec Cloud GB200 Cluster USA prnewswire.com, together.ai

Needs a second look

40 systems have cited sources that could be read and that do not name them. That does not mean the fact is wrong. Some are naming mismatches, where the page describes the same hardware under another name (the last column shows what the check did find), and some are citations that should be replaced:

SystemCountryWhat the cited pages did show
Abe USA nothing matching the row
Abel (UiO) Norway Mellanox InfiniBand FDR; Intel Xeon E5-2670
Aire (Leeds) United Kingdom NVIDIA L40S; quantity 84
ALPS (NCHC) Taiwan Mellanox InfiniBand QDR
Baku (HLRS) Germany Windows HPC Server 2008
Blue Horizon USA nothing matching the row
ByteDance MegaScale Cluster China nothing matching the row
Cab USA Intel Xeon E5-2670; Mellanox InfiniBand QDR
CARL and EDDY Germany Intel Xeon E5-2650 v4; Mellanox InfiniBand FDR
Chebyshev Russia Mellanox InfiniBand DDR; CentOS
Chengdu Intelligent Computing Center China nothing matching the row
Chongqing AI Innovation Center China nothing matching the row
CNIC Era China Mellanox InfiniBand FDR
CSAR Cray T3E-1200E United Kingdom nothing matching the row
CZI GPU Cluster USA NVIDIA H100 SXM5; VAST Data Platform; quantity 1024
Digital Government AI Cluster Uzbekistan nothing matching the row
DOST Blue Gene/P Philippines nothing matching the row
Eka India nothing matching the row
Gaïa (EDF) France Omni-Path Architecture
Hawk (Cardiff) United Kingdom Intel Xeon Gold 6148; quantity 26; quantity 30
IBM Vela USA nothing matching the row
Jade (CINES) France Intel Xeon E5450
Jade (LLNL) USA Intel Xeon E5-2695 v4; Omni-Path Architecture; Tri-Lab Operating System Stack
Jean (ARL DSRC) USA NVIDIA InfiniBand HDR; CentOS
Jean Zay V100 France NVIDIA Tesla V100; Omni-Path Architecture
JSS FX1 Japan Mellanox InfiniBand DDR
Julich Cray T3E Germany nothing matching the row
Kay Ireland Intel Xeon Gold 6148; NVIDIA Tesla V100; quantity 64
Lynx (Arizona) USA nothing matching the row
Maia (NAS) USA Intel Xeon E5-2670; Mellanox InfiniBand FDR
MCR USA Quadrics QsNet
Neo USA NVIDIA H200 SXM; Tri-Lab Operating System Stack; quantity 32
NIPA (Kakao) South Korea Intel Xeon 6 6900P (Granite Rapids); NVIDIA InfiniBand NDR (Quantum-2); NVIDIA B200 SXM
Puma (Arizona) USA nothing matching the row
Qingdao AI Computing Center China nothing matching the row
Raijin Japan NEC SX-Aurora TSUBASA Vector Engine; quantity 4320; NVIDIA InfiniBand HDR
Sol (ASU) USA Rocky Linux
TSUBAME1.0 Japan nothing matching the row
Vayu (NCI) Australia Mellanox InfiniBand QDR
Zin USA Intel Xeon E5-2670; Mellanox InfiniBand QDR

The read

The honest summary is that this is a reported dataset with a verified minority. About one system in four (20%) and one component edge in four (17%) has a second independent source behind it. Among the sectors with at least fifteen systems the verified share runs from 16% (Neoclouds) to 22% (Research HPC), and the small sectors move on a single row.

The tier is a column, not a badge of quality, so it should be used as a filter: take the verified rows where a wrong number is expensive, the reported rows for scale and direction, and treat any planned system's capacity as a claim. The fastest way to raise the verified share is not more systems but a second source for the ones already here, and the shortlist above is where that work starts.