Large-scale compute, told as a supply chain.
Rankings tell you how fast a machine is. They do not tell you who built it, whose fabric it runs on, which of those vendors still exists under its own name, or which listed company carries the exposure. Compute Atlas is the layer underneath: a parts graph with 1540 systems, 545 sites and 836 companies as first-class nodes, every fact carrying its own source.
Where the machines are
Full mapPin area follows the summed Rmax of the systems running at the site today.
- Runs systems today
- Historical systems only
- Country hosts a site
Click a pin for details. Drag to pan; use the buttons, or Ctrl or Cmd with scroll, to zoom. Pins mark the host city, not the building: coordinates are city-level. In time travel, a system with no recorded end year is shown only in its first year, and systems with no recorded start year are left out.
- Systems
- 1540
- Component edges
- 5581
- Parts
- 275
- Companies
- 836
- Countries
- 74
- Aggregate Rmax
- 19.02 EFlop/s
Explore
All toolsWho really supplies a machine?
Follow a part to its supplier, its owner today and the ticker.
CompareTwo systems side by side
Differences highlighted, with a link to share.
BoardWhere sources disagree
51 open conflicts, both sides linked.
ForecastSystems to watch
Planned machines and the windows their operators state.
- Corrected Cerebras: Cerebras is now a listed company (CBRS) 2026-10-20
- Added Mesabi: Mesabi added from its source pages 2026-10-18
- Added Cumulus (Cambridge): Cumulus (Cambridge) added from its source pages 2026-10-18
- Added Hera (LLNL): Hera (LLNL) added from its source pages 2026-10-18
- Added Edge (LLNL): Edge (LLNL) added from its source pages 2026-10-18
- Added Fission (INL): Fission (INL) added from its source pages 2026-10-18
What this answers that a ranking cannot
All analysesAccelerator share, weighted by FLOPS
Machine-count share understates whoever supplies the largest systems. Weighting by capability reorders it.
Interconnect share over time
Two acquisitions in twelve months ended the independent fabric business. Invisible without entity resolution.
Domestic substitution index
The European line sits at zero in every year. Not one European system is fully domestically supplied.
The shadow list
113 clusters on no ranking anywhere come to 96 to 394% of every ranked system here combined.
Announcement to first light
Median 6 months. The worst case, Aurora, took 9 years.
Public-market exposure by ticker
For each listed company, which systems it appears in and whether that exposure is one contract or many.
Silicon generation decay
Vendors ship a generation every 18 months. The installed base turns over on a seven-to-ten-year cycle.
Installed power by cooling architecture
Air stops being physically capable around 20 kW a rack. Current cabinets are rated to twenty times that.
Supply-chain concentration
Every component role in this dataset sits above the antitrust threshold for a highly concentrated market.
Who operates compute
By FLOPS the public sector still leads. By machines it does not. The gap is the operators that never submit a benchmark.
The build pipeline
Gigawatts of announced AI campuses, by operator and country, each figure tied to the page that states it.
How far to trust this data
Verified, reported and estimated by sector and country, and the citations that need a second look.
Performance per watt
Measured efficiency has doubled about every two and a half years since 2000, and accelerators set the frontier.
How long machines last
Retired systems ran a median of about six years, whatever the decade and however fast the silicon moved.
Every analysis carries a chart, a written read, a note stating exactly how it was calculated, a table of the same numbers, and a CSV of our derived output. The shadow list is the only one that publishes an estimate rather than a count, and it does so only with its estimator, its measured basis and every input shown.
Three things the spec strings hide
Who a 2002 supplier is today. ASCI Q ran on Alpha processors designed by Digital Equipment Corporation. Compaq bought DEC in 1998, HP bought Compaq in 2002, and the enterprise business became HPE in 2015. Three hops from the silicon to the ticker. Quadrics, which supplied its fabric, wound down in 2009 with no acquirer at all, so that row simply terminates. See it.
Where the substitution line actually moved. Tianhe-1A was American compute on a Chinese fabric in 2010. Sunway TaihuLight was fully domestic silicon in 2016. Tianhe-2 was rebuilt with domestic accelerators after an export control blocked the upgrade it had planned, and it got faster. See the index.
What is missing entirely. The largest AI clusters built in the last three years mostly never submitted results, so they appear on no public ranking. We carry ranked and unranked systems in one taxonomy with explicit confidence tiers, rather than pretending the ranked world is the whole world. See a system with no rank.
How supercomputers work
A technical guide in 10 chapters, cross-linked into the dataset so a claim about how something works can be checked against machines that did it.
- What actually makes a supercomputer different 12 min
- Inside a node 14 min
- The interconnect 15 min
- How supercomputer performance is measured 13 min
- Storage and I/O 12 min
Latest writing
- Google Willow: what it has and has not shown 2026-10-07
- Microsoft Fairwater: two sites, few numbers 2026-10-07
- Meta Hyperion and Prometheus: tents and gas 2026-10-07
Recently operational
All 1540 systemsFran
USA · integrated by Hewlett Packard Enterprise
Nexus (Georgia Tech)
USA · integrator not recorded
Firebird Armenian AI Factory (DC-1)
Armenia · integrated by Dell Technologies
IREN Childress Horizons 1-4 (Microsoft)
USA · integrated by Dell Technologies
AI datacenters
All 324 AI datacentersBeyond the GPU count
Power, PUE, water use, investment and utilisation for 284 AI datacenters, each figure labelled by who stated it: the operator, a filing, or an estimate.
Who buys compute from whom
76 announced cloud, lease, chip-supply and joint-venture deals, with capacity and headline value, and why the values do not add up.
How AI campuses are powered
Grid, on-site gas, nuclear, renewables and storage behind 51 campuses.
Is there a TOP500 for AI datacenters?
No. What MLPerf, InferenceX and the independent datasets measure, with 37 published results and what each leaves out.
Quantum computers
All 150 machines150 quantum machines
Named quantum computers by qubit technology, with the operators and any supercomputer they sit beside. A qubit is not a FLOP, so they are kept apart from the systems above.
Is there a TOP500 for quantum?
No. What quantum benchmarks exist, who measures them, and the 125 published figures and milestone claims recorded here, each marked vendor reported, peer reviewed or independent.
Common questions
- What is Compute Atlas?
- Compute Atlas is an open dataset of 1540 supercomputers and AI clusters in 74 countries, mapped as a graph of which company's parts (accelerators, CPUs, interconnect, storage and cooling) are in which machine. Every row carries a source link and a confidence tier. Methodology →
- How is Compute Atlas different from the TOP500?
- The TOP500 ranks systems by one benchmark result. Compute Atlas tracks the supply chain behind the machines, including AI clusters that never submit a benchmark and planned builds, and it stores only pointers to ranking entries, never a copy of any list. It is not affiliated with TOP500.org. The shadow list →
- Which company supplies the most supercomputer accelerators?
- NVIDIA supplies 42% of the installed FLOPS in this dataset as of 2026, weighted by capability rather than machine count. The share is computed for every year from 1993, and it moves by several points when one very large system is added or retired. Accelerator share over time →
- How much power are the largest planned AI data centers expected to use?
- 105 planned or under-construction systems in this dataset state a power capacity, about 117.0 GW in total. Fermi America states the most at 17.0 GW, and the five largest operators account for 44% of the total. These are announced figures, not delivered power. The build pipeline →
- How reliable is the data?
- 286 of 1540 systems (19%) are marked verified, meaning two independent organisations confirm the same fact. The rest are reported from one cited source. When every cited page was re-read, 1350 systems had at least one source that names them. How far to trust this data →
- Can I download the data?
- Yes. 22 compiled tables are available as CSV files, and every analysis page offers its derived data as a CSV. Each row keeps its source link and confidence tier. Download the dataset →