The 117 retired systems in this dataset with both a first-operational and a
decommissioning date ran for a median of 6.0 years, and half of them for between
5.0 and 7.1. That has barely moved: the typical machine first run in
2005 to 2009 lasted 6.0 years and in 2010 to 2014 6.1,
through four decades in which the silicon generation turned over several times faster than that.
The steadiness suggests the cadence is set by how institutions fund and write off machines more
than by how quickly the technology ages.
This measures the machines that have ended. A system still running has not
been measured yet, so the most recent groups are short-lived by construction: only the ones
that have already been retired are in them. Read the older cohorts for the shape and treat
the latest as a floor. Only 117 of the 218 decommissioned systems have both dates
on record, almost all at government labs and research centres, because those are the operators
that publish a retirement notice.
Phase 1 seed dataset, compiled by hand.
These rows were built from public operator, laboratory and vendor sources. A
mechanical second-reader pass has since fetched every
cited source: 1092 of 1153 systems have a readable citation that names
them, and 40 are genuinely weakly sourced. Every claim carries its source and a
confidence tier. Treat anything below
verified as a lead, not a citation.
Government labs
Universities & research centres
Other operators
Each dot is one retired system: the years between its first-operational and decommissioning dates, placed at the year it first ran. The line is the median across all of them.
A system is included if its status is decommissioned and it records both a first-operational
date and a decommissioning date. Dates are read to the month where the source gives one and
to the middle of the year where it gives only a year, so a life measured between two bare
years is accurate to roughly half a year either way.
Systems recorded as upgraded are not decommissioned and are not here, so a machine that was
rebuilt in place and kept its name is invisible to this page; the lives shown are of machines
that ended, not of platforms. Systems whose retirement is recorded without a date are also
left out, which is why the count here is smaller than the number of decommissioned rows.
By sector
Operator sector
Retired systems
Median years
Government labs
39
6.4
Research HPC
77
6.0
Only sectors with at least four retired systems on record are shown. Every other sector is too thinly represented to say anything about.
175 systems in this dataset are recorded as operating 2017 or more years after they
first ran. Only 23 of them have had their status confirmed in 2025 or 2026; the other
152 carry a status that was last checked earlier, or never. That is the weakest part of
this page, and of the dataset: “operational” on an old row means nobody has recorded a retirement,
not that anyone has seen it running. The retirement dates that exist were found by looking, one
operator page at a time, and most old rows have not been looked at yet.
The surprising part is how little the number depends on the era. A 1990s vector machine, a 2000s
cluster and a 2010s GPU system all tend to retire after roughly five to eight years, even
though a new accelerator generation now arrives about every two. That points to depreciation
schedules, grant cycles and procurement rhythms, which are set by finance and not by engineering,
though this dataset cannot test that directly.
The exceptions are informative. The longest life on record, Pleiades at 17.2 years, belongs to a
machine that was expanded repeatedly rather than replaced, which is also why platform names and
machine lives are different things. The shortest, Gyoukou at 0.8 years and
Tesla Dojo at 2.1, ran under three years each. Whether the AI clusters of the
2020s will follow the usual rhythm is the open question: they are bought and written off by
companies, not by funding agencies, and very few have yet been retired.
This site uses Google Analytics to see which pages get read. If you accept, it may set
cookies to recognise repeat visits. If you reject, no cookies are set, but an anonymous page-load
count is still sent. See how we use it.