Methodology

What counts as a cluster

One row is one identified installation of GPUs: a named supercomputer, a named AI campus build, or a named customer deployment. Fleet totals (“24,000 H100s across six data centers”), platform scaling limits (“scales to 20,000 GPUs”) and custom-silicon machines (Trainium, TPU, Ascend, wafer-scale engines) are not clusters, and they stay out.

Where the data comes from

First-party pages come first: operator user documentation, engineering blogs, lab and supercomputing-centre pages, company investor materials and SEC filings. Trade press fills gaps and is named as press in the row. For a few campuses, live and planned power comes from Epoch AI’s AI data centers dataset, used under the Creative Commons Attribution 4.0 licence. Those numbers are satellite-based estimates: the record page says “Epoch AI estimate” next to them and the source title starts with the same words. We changed nothing in Epoch AI’s figures; we pick the as-of date and set company figures ahead of them wherever a company publishes one. Every record lists its sources with the date we accessed them, and every number in the table links to the source behind it.

How GPUs are counted

The installed count is physical GPUs. Where an operator publishes nodes and GPUs per node, the count is that product (Frontier: 9,856 nodes × 4 MI250X = 39,424). An AMD MI250X board counts once, not as two dies; an MI300A APU counts as one GPU; a GH200 superchip counts as one GPU. H100-equivalents, design limits and unaudited press totals never enter the installed column. Planned and contracted counts have their own column. When a source covers one partition or one building, the note says so.

Status: announced versus operational

Operational: the operator or a primary source says the cluster runs, or the operator’s current user documentation describes it in the present tense. Partly operational: a phase is live and more is coming. Under construction: building is underway and no phase is live. Announced: disclosed or contracted with no confirmed go-live. Where an operator gave a go-live target and later sources confirm nothing, the row stays announced and the status note carries the target and the date we last looked. “Operational since” is filled only when a source gives a month or year.

Power

Live power is megawatts of IT load in service; planned power is the full-build figure. Company figures and Epoch AI estimates are marked in each row’s power note. Apparent power in MVA and facility (rather than IT) megawatts are kept in the note, not the number.

Overlap with AI Data Center List

Our sister site AI Data Center List tracks the campus: megawatts, land and tenants. This list tracks the cluster: owner, GPU model and count. Where a cluster sits on a campus both sites cover, we use the same source and the same figures. A status can differ when a campus runs one cluster and builds another: Mistral’s Bruyères-le-Châtel campus is operational there because an earlier cluster runs on it, while the new 13,800-GPU GB300 cluster is still announced here.

Corrections

Owners and readers can send corrections to hello@gpuclusterlist.com with a public source. We update the row and its verified date. Featured placement is paid and labelled; it never changes a row’s data or its position in the sorted table.