Supercomputers/Analysis
Derived views
Analysis
Eight views that cannot be produced from a ranked list as it exists, because a ranked list is system-centric and these are all questions about companies, countries, generations and time. Each will ship as a static page with a chart, a written read, a “how this was calculated” expander, and a downloadable CSV of our derived output — never of the source list.
None of these ship at Phase 1. The build order is deliberate: schema, ingest, entity resolution and per-system pages first, historical backfill second, and only then the time series that depend on it. A trend line over a 21-system seed would look like analysis and be noise. Two single-edition snapshots are live already and are linked below where they exist.
1. Accelerator share weighted by FLOPS, not machine count
Phase 2Machine-count share is the metric everyone quotes, and it systematically understates whoever supplies the largest systems. One exascale machine and one departmental cluster are one system each by count and three orders of magnitude apart by capability.
Blocked on: Historical backfill to 1993 for the time series. A single-edition version is already live on the section index.
2. Interconnect share over time
Phase 2InfiniBand, Slingshot, Omni-Path, Tofu and the proprietary Chinese fabrics, by installed FLOPS and by system count. The fabric layer consolidated hard when NVIDIA bought Mellanox and HPE bought Cray — two acquisitions that turned four independent suppliers into two.
Blocked on: Historical backfill. A current-snapshot version is on the home page.
3. Domestic substitution index
Phase 3For each country, the share of installed FLOPS whose CPU, accelerator, interconnect and OS are all domestically supplied. The China line is the single most newsworthy series this dataset can produce, and the seed already contains its inflection points: Tianhe-1A in 2010 with American compute on a Chinese fabric, Sunway TaihuLight in 2016 with neither.
Blocked on: Country-of-origin attribution on every part, plus the historical backfill.
4. The shadow list
Phase 3Total ranked FLOPS versus best-estimate FLOPS sitting in large unranked commercial AI clusters. This is the gap the incumbent structurally cannot show, because a machine that never submitted a result cannot appear on a list of results.
Blocked on: A published estimator, exposed in full, before a single number ships. See the estimates section of the methodology page.
5. Announcement-to-first-light lag
Phase 2By vendor and by country, over time. Aurora took eight years from contract to first light through a cancelled architecture and a redesign; that is a vendor-execution story, and it is only visible when you have both dates for enough systems to compare.
Blocked on: Announcement dates for the historical backfill; the seed carries both dates for most systems already.
6. Public-market exposure by ticker
Phase 4For each ticker: the systems it appears in, aggregate FLOPS attributable, and whether that exposure is concentrated in one program. A supplier whose entire presence is a single national laboratory contract is a different risk from one spread across thirty installations.
Blocked on: Role-weighted attribution, which needs a stated method before it can be published. A breadth-only version is on every company page.
7. Refresh cadence and silicon generation decay
Phase 2How long each accelerator generation stays in the installed base before displacement. Summit ran V100s for six years; the seed records its retirement, which removes the largest remaining block of POWER installed base in one step.
Blocked on: Historical backfill, so generations can be tracked in and out rather than only at their peak.
8. Installed MW by cooling architecture and vendor
Phase 3The ranking stops at the rack. Cooling architecture, utility, substation capacity and site are where the constraint on the next generation actually lives, and lead times on medium-voltage switchgear now gate schedules more often than silicon does.
Blocked on: Facility-layer coverage across sites; the seed records cooling architecture but power at only a handful of sites.