Compute·Atlas

Supercomputers/Analysis/Announcement to first light

Analysis 05

Announcement to first light

Across 52 systems the median gap between announcement and entering service is 29 months. The distribution matters more than the median: the slowest system in this dataset, Aurora, took 107 months: 8.9 years, and the reason is silicon that was cancelled and redesigned mid-procurement, not a building that ran late.

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: 32 of 52 systems have a readable citation that names them, and 4 are genuinely weakly sourced. Every claim carries its source and a confidence tier. Treat anything below verified as a lead, not a citation.

  • Hewlett Packard Enterprise
  • IBM
  • Atos
  • Other integrators
Announcement to first light, by year of first light Dot plot of the gap in months between a system being announced and entering service, plotted against the year it entered service. 02754811081993200020052010201520202025 median 29 months CM-5/1024 Thinking Machines 29 months, first light 1993 ASCI Red Intel 53 months, first light 1997 ASCI White IBM 53 months, first light 2000 ASCI Q Hewlett Packard Enterprise 34 months, first light 2002 Earth Simulator NEC 62 months, first light 2002 BlueGene/L IBM 46 months, first light 2004 Pleiades Hewlett Packard Enterprise 10 months, first light 2008 Intrepid IBM 24 months, first light 2008 Roadrunner IBM 29 months, first light 2008 Jaguar Hewlett Packard Enterprise 58 months, first light 2009 Tianhe-1A NUDT 22 months, first light 2010 K computer Fujitsu 65 months, first light 2011 Titan Hewlett Packard Enterprise 22 months, first light 2012 Mira IBM 41 months, first light 2012 Sequoia IBM 41 months, first light 2012 Tianhe-2A NUDT 41 months, first light 2013 Blue Waters Hewlett Packard Enterprise 74 months, first light 2013 Sunway TaihuLight NRCPC 29 months, first light 2016 Cori Hewlett Packard Enterprise 34 months, first light 2016 Stampede2 Dell Technologies 17 months, first light 2017 MareNostrum 4 Lenovo 17 months, first light 2017 TSUBAME3.0 Hewlett Packard Enterprise 19 months, first light 2017 Trinity Hewlett Packard Enterprise 41 months, first light 2017 Piz Daint Hewlett Packard Enterprise 65 months, first light 2017 ABCI Fujitsu 19 months, first light 2018 SuperMUC-NG Lenovo 20 months, first light 2018 Sierra IBM 53 months, first light 2018 Summit IBM 53 months, first light 2018 Frontera Dell Technologies 17 months, first light 2019 Selene NVIDIA 5 months, first light 2020 Hawk Hewlett Packard Enterprise 13 months, first light 2020 JUWELS Booster Atos 22 months, first light 2020 Fugaku Fujitsu 77 months, first light 2020 Vega Atos 15 months, first light 2021 Karolina Hewlett Packard Enterprise 19 months, first light 2021 ARCHER2 Hewlett Packard Enterprise 34 months, first light 2021 Perlmutter Hewlett Packard Enterprise 40 months, first light 2021 LUMI Hewlett Packard Enterprise 29 months, first light 2022 Frontier Hewlett Packard Enterprise 40 months, first light 2022 Leonardo Atos 46 months, first light 2022 Eagle Microsoft 10 months, first light 2023 Eos NVIDIA 10 months, first light 2023 Shaheen III Hewlett Packard Enterprise 22 months, first light 2023 MareNostrum 5 ACC Atos 23 months, first light 2023 Aurora Hewlett Packard Enterprise 107 months, first light 2023 Meta GenAI cluster (RoCE) Meta Platforms 0 months, first light 2024 Meta GenAI cluster (InfiniBand) Meta Platforms 0 months, first light 2024 Isambard-AI Hewlett Packard Enterprise 18 months, first light 2024 Venado Hewlett Packard Enterprise 27 months, first light 2024 Alps Hewlett Packard Enterprise 38 months, first light 2024 El Capitan Hewlett Packard Enterprise 70 months, first light 2024 JUPITER Atos 44 months, first light 2025 AuroraFugakuBlue Waters
Each dot is one system, positioned by the year it entered service and how many months earlier it was announced. The three most frequent integrators carry a colour; the rest are neutral.
Table view: Announcement to first light
Period of first lightSystemsMedian monthsLongest months
1990–199412929
1995–199915353
2000–200445062
2005–200942758
2010–201474174
2015–2019122565
2020–20242223107
2025–202914444
How this was calculated

Lag is the whole number of months from announced_date to first_operational_date, both of which are sourced fields on the system row. All 52 systems in this dataset carry both dates, so nothing is excluded here for missing data.

“Announced” is the earliest public statement of intent we hold a source for, typically a programme announcement or a contract award, whichever came first. That definition is the weak point of this chart: announcement practice is not consistent across three decades or across countries, and a laboratory that announces at contract signature will look faster than one that announces at programme approval for the same real schedule. Treat the era-over-era trend as more reliable than any individual dot.

The integrator is the supplier filling the integrator role, rolled up to today's parent, so Cray-built systems appear under HPE. Colour is capped at three groups because a dot plot compares every pair of series at once, and the categorical palette clears the colour-blindness floors for three simultaneous hues but not for eight. The remaining integrators share a neutral tone and are identified in the tooltip and the table rather than by colour.

The read

The lag has grown, and the growth is not evenly distributed. Early systems were announced close to delivery because they were built from parts that already existed. The modern exascale procurements commit to silicon that has not been fabricated yet, which converts a delivery schedule into a bet on a vendor's roadmap.

Aurora is the extreme case and worth stating plainly, because it is the data point most responsible for the shape of this chart. It was contracted in 2015 around a Xeon Phi design, that line was cancelled, the system was redesigned around Ponte Vecchio, and it entered service in 2023. That is a vendor-execution story. The site was ready long before the parts were.

The counter-example is in the same table: the systems built from parts that already shipped in volume (the academic Dell and Lenovo clusters, the DGX deployments) cluster well below the median. The trade is visible here as a number. Committing to unreleased silicon buys leadership performance and costs years of schedule risk, and which of those a buyer should want depends entirely on whether they need the top of the list or need the machine.

By integrator

Integrators with at least two systems in the dataset. Small samples: these are medians over a handful of machines each, not a vendor scorecard.

Integrator (today's parent)SystemsMedian monthsRange
Fujitsu 3 65 19–77
IBM 8 44 24–53
Hewlett Packard Enterprise 21 34 10–107
NUDT 2 32 22–41
Atos 5 23 15–46
Lenovo 2 19 17–20
Dell Technologies 2 17 17–17
NVIDIA 2 8 5–10
Meta Platforms 2 0 0–0

Longest gaps

SystemAnnouncedFirst operationalMonthsIntegrator
Aurora 2015 2023-12 107 Hewlett Packard Enterprise
Fugaku 2014 2020-06 77 Fujitsu
Blue Waters 2007 2013-03 74 Hewlett Packard Enterprise
El Capitan 2019 2024-11 70 Hewlett Packard Enterprise
K computer 2006 2011-06 65 Fujitsu