LLNL's current system. Its own notes record a stockpile stewardship workload outright.
Showcase
What people actually run on these machines
A ranking or a spec sheet says how fast a system is. It does not say what it is for. Below are 28 systems from this dataset grouped by mission, with a note on why the work in question needs a supercomputer rather than a large ordinary cluster. Every blurb paraphrases that system's own sourced notes; click through for the citations.
These categories are ours, not a field in the data. No CSV column says "this machine does seismic imaging"; we grouped systems whose own operator has stated a purpose, in text that is already sourced on that system's page. A system with no clear, sourced statement of purpose is left out rather than guessed into a category, so this page covers 28 of the 1153 systems in the dataset, not all of them.
Weapons physics
Certifying a nuclear arsenal without detonating anything
The US and France have not conducted a live nuclear test since the 1990s, and both still have to certify that their stockpiles work and age safely. The substitute for a test is a first-principles physics simulation of the weapon's secondary, run at a resolution fine enough to stand in for the real thing. That program, in the US, is NNSA's Stockpile Stewardship Program, run through its three national laboratories; the French equivalent runs through the CEA's Military Applications Division. Several systems below state that mission directly in their own sourced notes.
Why it needs a supercomputer. These are three-dimensional, multi-physics codes (radiation transport, hydrodynamics, material equations of state) at a mesh resolution fine enough to substitute for an experiment, which is a different bar than a research code that can accept a coarser grid. The work is also air-gapped by law: it cannot run on shared or cloud capacity, so each lab needs its own dedicated machine at the frontier of what is buildable.
LLNL, under the same CORAL procurement as the unclassified Summit. Recorded workload: stockpile stewardship.
Jointly run by Los Alamos and Sandia under the Advanced Simulation and Computing programme; built in two silicon phases.
Sandia's own release calls this an application-readiness system, an on-ramp so weapons codes are validated before they run at full scale on El Capitan.
An NNSA tri-lab system inside LLNL's Restricted Zone, independently confirmed by LLNL's own hardware page.
El Capitan's hardware twin at LLNL, same MI300A node, same fabric, but LLNL's own dedication material is explicit that Tuolumne runs open science, not stockpile stewardship. It is here as the counter-example: identical silicon, different mission.
The French equivalent: phase two of CEA's EXA1 program, run by its Military Applications Division, framed by the joint CEA and Eviden release around European sovereignty over the fabric as much as the compute.
CEA's earlier defence-programme system, deployed at Bruyeres-le-Chatel and still listed on TOP500 as of June 2026.
Energy
Turning sound waves into a map of what is underground
Oil and gas operators send acoustic pulses into the earth and record what echoes back, then invert that recording into a 3D image of the rock a few kilometres down: where the traps are, and where drilling a well is worth the cost. Every system below is an operator-owned cluster whose own release, not a third party, names this as the job.
Why it needs a supercomputer. The inversion (often full waveform or full wavefield inversion) is an iterative, PDE-constrained optimisation over a 3D grid covering tens of kilometres at sub-wavelength resolution, solved from scratch for each new survey. More compute buys a finer grid and a deeper, more confident image, which is the direct input to a decision that costs tens of millions of dollars to get wrong.
Eni's current system, paired with the still-running HPC6 rather than replacing it, on a seismic and subsurface imaging, CO2 storage and battery-research workload.
Eni's Frontier-family MI250X system, the same functional class the note explicitly ties to seismic and subsurface imaging and CO2 storage.
Eni's earlier V100 system, presented in 2020, sixth on the June 2020 TOP500.
ExxonMobil's own account ties this cluster to 4D elastic full wavefield inversion for its Stabroek Block assets offshore Guyana.
Petrobras's cluster at its Cenpes R&D complex, the largest of five new systems replacing three older ones, built for pre-salt and Equatorial Margin exploration.
Saudi Aramco's system, unveiled with telecom partner stc, used for seismic imaging and deep-learning models of the subsurface.
TotalEnergies' industrial system at its Pau research centre, the most powerful commercial machine in the world when built.
Climate and earth science
Running the planet forward in time
Climate models integrate the physics of the atmosphere and ocean forward across centuries of simulated time to build the ensembles that probabilistic projections are built from; astrophysics survey work processes and models the universe at the scale modern telescopes now capture it.
Why it needs a supercomputer. A climate model is a 3D physics grid covering the whole planet, integrated over millions of timesteps; halving the grid spacing multiplies the compute needed by roughly a factor of eight once you account for all three spatial dimensions and the shorter timestep stability requires. There is no shortcut to finer resolution other than more compute.
The German Climate Computing Centre's (DKRZ) system, an EPYC cluster with an A100 partition, in service since November 2021.
NERSC's system, named for the astrophysicist Saul Perlmutter, whose supernova survey work (part of the discovery that the universe's expansion is accelerating) ran at NERSC.
Artificial intelligence
Training the models behind chat, image, video and code generation
Every system below is a dedicated cluster its operator built to train large neural networks, mostly outside any public ranking: see the shadow list for what that means for measuring this part of the industry. Sizes are the operator or a vendor partner's own disclosed figures, since almost none of these run a public benchmark.
Why it needs a supercomputer. Training a frontier model means propagating gradients across hundreds of billions to trillions of parameters, over trillions of tokens, with every accelerator in the job synchronising after each step. That needs thousands of accelerators on a fabric fast enough that the synchronisation does not dominate the run, kept busy continuously for months, which is a supercomputer-shaped problem even when the arithmetic itself does not need FP64 precision.
xAI's Memphis cluster: NVIDIA and Supermicro's own releases independently put it at 100,000 Hopper-generation GPUs on Spectrum-X Ethernet.
Built by Crusoe, operated by Oracle as OpenAI capacity: NVIDIA GB200 racks, with a stated 1.2 GW design capacity for the whole campus.
An AWS cluster built with and used by Anthropic: Amazon describes it as nearly half a million Trainium2 chips.
Tesla's own quarterly decks report roughly 50,000 H100s at Gigafactory Texas, plus 16,000 H200s added later.
Meta's research cluster: Meta and NVIDIA independently confirm 760 DGX A100 nodes, 6,080 GPUs.
The Azure system Microsoft built for OpenAI in 2020: over 285,000 CPU cores and 10,000 GPUs, by Microsoft's own account.
Google's TPU v4 pod: 4,096 chips linked by Google's own optical circuit switches, running since 2020 per Google's ISCA 2023 paper.
Cerebras' 16-system CS-2 cluster, 13.5 million AI cores, with Argonne, JasperAI and the University of Cambridge named as users.
The first Cerebras and G42 system, 64 CS-2 systems, offered through G42 Cloud.
The third Cerebras and G42 system, 64 CS-3 systems on the WSE-3, rated at 8 exaFLOPs of AI compute.
A Microsoft AI datacenter: NVIDIA Blackwell GB200 and GB300 GPUs at up to 72 per rack, linked to its Wisconsin sibling by a dedicated AI wide-area network.
Classified and undisclosed
What we can't show you
21 systems in this dataset carry a workload_class of classified. That
tag is not always an explicit label from the source: for a system like El Dorado,
it is inferred from mission context (an NNSA application-readiness system feeding codes into a stockpile-stewardship
machine), which is why that row is marked reported rather than verified. For these systems
we record what is sourced (operator, hardware generation, era, its place in a procurement programme) and nothing
about what is actually run on them day to day, because nothing about that is disclosed.
ASCI RedASCI Blue MountainASCI Blue-PacificASCI WhiteASCI QBlueGene/LTera-10RoadrunnerTera-100CieloZinSequoiaJade (LLNL)Tera-1000-2TrinitySierraMagma (LLNL)AWE VulcanCEA-HEEl CapitanEl Dorado
See confidence tiers for how a tag inferred from context differs from one a source states outright.