Compute·Atlas

Supercomputers/Analysis/Power and cooling

Analysis 08

Installed power by cooling architecture

25 of the 52 systems here are directly liquid cooled, carrying 190 MW of the 271 MW this dataset records. The transition from air is not gradual: air cooling stops being physically capable somewhere around 20 kW a rack, and current cabinets are rated to twenty times that. This is the layer a ranking stops short of, and it is now more often the constraint on building the next machine than the silicon is.

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.

Read the shape, not the spikes. This is the installed base of the 52 systems in this dataset: a curated seed chosen for supply-chain diversity, not a census of the market. At that sample size one machine entering or leaving service moves a share line by tens of points, so year-to-year jumps are composition effects rather than market events. The multi-year trends are the part worth quoting; the individual steps usually have one system's name on them, and the table view under the chart will tell you which.

  • Direct liquid cooling, warm water
  • Direct-to-chip cold plate
  • Air cooling (CRAC/CRAH)
  • Liquid-assisted air (in-cabinet coil)
  • Not recorded
Cooling architecture share of installed FLOPS, 1993–2025 Stacked area chart showing the share of installed Rmax running under each cooling architecture by year, with a band for systems where no architecture is recorded. Direct liquid cooling, warm waterDirect-to-chip cold plateAir cooling (CRAC/CRAH)Liquid-assisted air (in-cabinet coil)Not recorded 0%25%50%75%100%19931998200320082013201820232025 Direct liquid cooling, warm waterNot recorded 1993 Direct liquid cooling, warm water 0.0% Direct-to-chip cold plate 0.0% Air cooling (CRAC/CRAH) 0.0% Liquid-assisted air (in-cabinet coil) 0.0% Not recorded 100.0%1994 Direct liquid cooling, warm water 0.0% Direct-to-chip cold plate 0.0% Air cooling (CRAC/CRAH) 0.0% Liquid-assisted air (in-cabinet coil) 0.0% Not recorded 100.0%1995 Direct liquid cooling, warm water 0.0% Direct-to-chip cold plate 0.0% Air cooling (CRAC/CRAH) 0.0% Liquid-assisted air (in-cabinet coil) 0.0% Not recorded 100.0%1996 Direct liquid cooling, warm water 0.0% Direct-to-chip cold plate 0.0% Air cooling (CRAC/CRAH) 0.0% Liquid-assisted air (in-cabinet coil) 0.0% Not recorded 100.0%1997 Direct liquid cooling, warm water 0.0% Direct-to-chip cold plate 0.0% Air cooling (CRAC/CRAH) 94.7% Liquid-assisted air (in-cabinet coil) 0.0% Not recorded 5.3%1998 Direct liquid cooling, warm water 0.0% Direct-to-chip cold plate 0.0% Air cooling (CRAC/CRAH) 94.7% Liquid-assisted air (in-cabinet coil) 0.0% Not recorded 5.3%1999 Direct liquid cooling, warm water 0.0% Direct-to-chip cold plate 0.0% Air cooling (CRAC/CRAH) 94.7% Liquid-assisted air (in-cabinet coil) 0.0% Not recorded 5.3%2000 Direct liquid cooling, warm water 0.0% Direct-to-chip cold plate 0.0% Air cooling (CRAC/CRAH) 17.6% Liquid-assisted air (in-cabinet coil) 0.0% Not recorded 82.4%2001 Direct liquid cooling, warm water 0.0% Direct-to-chip cold plate 0.0% Air cooling (CRAC/CRAH) 17.6% Liquid-assisted air (in-cabinet coil) 0.0% Not recorded 82.4%2002 Direct liquid cooling, warm water 0.0% Direct-to-chip cold plate 0.0% Air cooling (CRAC/CRAH) 91.0% Liquid-assisted air (in-cabinet coil) 0.0% Not recorded 9.0%2003 Direct liquid cooling, warm water 0.0% Direct-to-chip cold plate 0.0% Air cooling (CRAC/CRAH) 91.0% Liquid-assisted air (in-cabinet coil) 0.0% Not recorded 9.0%2004 Direct liquid cooling, warm water 0.0% Direct-to-chip cold plate 0.0% Air cooling (CRAC/CRAH) 99.1% Liquid-assisted air (in-cabinet coil) 0.0% Not recorded 0.9%2005 Direct liquid cooling, warm water 0.0% Direct-to-chip cold plate 0.0% Air cooling (CRAC/CRAH) 99.1% Liquid-assisted air (in-cabinet coil) 0.0% Not recorded 0.9%2006 Direct liquid cooling, warm water 0.0% Direct-to-chip cold plate 0.0% Air cooling (CRAC/CRAH) 99.1% Liquid-assisted air (in-cabinet coil) 0.0% Not recorded 0.9%2007 Direct liquid cooling, warm water 0.0% Direct-to-chip cold plate 0.0% Air cooling (CRAC/CRAH) 100.0% Liquid-assisted air (in-cabinet coil) 0.0% Not recorded 0.0%2008 Direct liquid cooling, warm water 5.8% Direct-to-chip cold plate 0.0% Air cooling (CRAC/CRAH) 81.3% Liquid-assisted air (in-cabinet coil) 0.0% Not recorded 12.9%2009 Direct liquid cooling, warm water 4.7% Direct-to-chip cold plate 0.0% Air cooling (CRAC/CRAH) 66.6% Liquid-assisted air (in-cabinet coil) 18.1% Not recorded 10.6%2010 Direct liquid cooling, warm water 3.7% Direct-to-chip cold plate 0.0% Air cooling (CRAC/CRAH) 52.5% Liquid-assisted air (in-cabinet coil) 14.4% Not recorded 29.4%2011 Direct liquid cooling, warm water 48.2% Direct-to-chip cold plate 0.0% Air cooling (CRAC/CRAH) 28.3% Liquid-assisted air (in-cabinet coil) 7.7% Not recorded 15.8%2012 Direct liquid cooling, warm water 55.6% Direct-to-chip cold plate 0.0% Air cooling (CRAC/CRAH) 9.7% Liquid-assisted air (in-cabinet coil) 29.3% Not recorded 5.4%2013 Direct liquid cooling, warm water 28.8% Direct-to-chip cold plate 0.0% Air cooling (CRAC/CRAH) 5.0% Liquid-assisted air (in-cabinet coil) 15.2% Not recorded 51.0%2014 Direct liquid cooling, warm water 29.0% Direct-to-chip cold plate 0.0% Air cooling (CRAC/CRAH) 5.1% Liquid-assisted air (in-cabinet coil) 15.3% Not recorded 50.6%2015 Direct liquid cooling, warm water 28.8% Direct-to-chip cold plate 0.0% Air cooling (CRAC/CRAH) 5.1% Liquid-assisted air (in-cabinet coil) 15.3% Not recorded 50.8%2016 Direct liquid cooling, warm water 55.5% Direct-to-chip cold plate 0.0% Air cooling (CRAC/CRAH) 2.8% Liquid-assisted air (in-cabinet coil) 14.3% Not recorded 27.5%2017 Direct liquid cooling, warm water 45.8% Direct-to-chip cold plate 0.0% Air cooling (CRAC/CRAH) 4.3% Liquid-assisted air (in-cabinet coil) 24.9% Not recorded 24.9%2018 Direct liquid cooling, warm water 30.4% Direct-to-chip cold plate 41.8% Air cooling (CRAC/CRAH) 2.2% Liquid-assisted air (in-cabinet coil) 12.8% Not recorded 12.8%2019 Direct liquid cooling, warm water 29.2% Direct-to-chip cold plate 40.1% Air cooling (CRAC/CRAH) 2.1% Liquid-assisted air (in-cabinet coil) 12.3% Not recorded 16.2%2020 Direct liquid cooling, warm water 58.3% Direct-to-chip cold plate 21.4% Air cooling (CRAC/CRAH) 6.7% Liquid-assisted air (in-cabinet coil) 5.0% Not recorded 8.6%2021 Direct liquid cooling, warm water 60.9% Direct-to-chip cold plate 19.7% Air cooling (CRAC/CRAH) 6.2% Liquid-assisted air (in-cabinet coil) 4.6% Not recorded 8.5%2022 Direct liquid cooling, warm water 85.0% Direct-to-chip cold plate 7.6% Air cooling (CRAC/CRAH) 2.4% Liquid-assisted air (in-cabinet coil) 1.8% Not recorded 3.3%2023 Direct liquid cooling, warm water 77.1% Direct-to-chip cold plate 4.8% Air cooling (CRAC/CRAH) 3.9% Liquid-assisted air (in-cabinet coil) 1.1% Not recorded 13.0%2024 Direct liquid cooling, warm water 84.5% Direct-to-chip cold plate 3.4% Air cooling (CRAC/CRAH) 2.7% Liquid-assisted air (in-cabinet coil) 0.3% Not recorded 9.1%2025 Direct liquid cooling, warm water 86.3% Direct-to-chip cold plate 1.4% Air cooling (CRAC/CRAH) 2.7% Liquid-assisted air (in-cabinet coil) 0.3% Not recorded 9.4%
Share of installed Rmax by cooling architecture. The grey band is systems where we have no recorded architecture: a gap in the dataset, not a category of machine.
Download this chart's data (CSV) ↓
Table view: Cooling architecture share of installed FLOPS
YearDirect liquid cooling, warm waterDirect-to-chip cold plateAir cooling (CRAC/CRAH)Liquid-assisted air (in-cabinet coil)Not recordedInstalled total
19930.0%0.0%0.0%0.0%100.0%59.7 GFlop/s
19940.0%0.0%0.0%0.0%100.0%59.7 GFlop/s
19950.0%0.0%0.0%0.0%100.0%59.7 GFlop/s
19960.0%0.0%0.0%0.0%100.0%59.7 GFlop/s
19970.0%0.0%94.7%0.0%5.3%1.13 TFlop/s
19980.0%0.0%94.7%0.0%5.3%1.13 TFlop/s
19990.0%0.0%94.7%0.0%5.3%1.13 TFlop/s
20000.0%0.0%17.6%0.0%82.4%6.07 TFlop/s
20010.0%0.0%17.6%0.0%82.4%6.07 TFlop/s
20020.0%0.0%91.0%0.0%9.0%55.81 TFlop/s
20030.0%0.0%91.0%0.0%9.0%55.81 TFlop/s
20040.0%0.0%99.1%0.0%0.9%534.01 TFlop/s
20050.0%0.0%99.1%0.0%0.9%534.01 TFlop/s
20060.0%0.0%99.1%0.0%0.9%534.01 TFlop/s
20070.0%0.0%100.0%0.0%0.0%528.00 TFlop/s
20085.8%0.0%81.3%0.0%12.9%7.95 PFlop/s
20094.7%0.0%66.6%18.1%10.6%9.71 PFlop/s
20103.7%0.0%52.5%14.4%29.4%12.24 PFlop/s
201148.2%0.0%28.3%7.7%15.8%22.75 PFlop/s
201255.6%0.0%9.7%29.3%5.4%66.10 PFlop/s
201328.8%0.0%5.0%15.2%51.0%127.5 PFlop/s
201429.0%0.0%5.1%15.3%50.6%126.5 PFlop/s
201528.8%0.0%5.1%15.3%50.8%126.1 PFlop/s
201655.5%0.0%2.8%14.3%27.5%233.1 PFlop/s
201745.8%0.0%4.3%24.9%24.9%299.7 PFlop/s
201830.4%41.8%2.2%12.8%12.8%582.3 PFlop/s
201929.2%40.1%2.1%12.3%16.2%605.9 PFlop/s
202058.3%21.4%6.7%5.0%8.6%1.14 EFlop/s
202160.9%19.7%6.2%4.6%8.5%1.23 EFlop/s
202285.0%7.6%2.4%1.8%3.3%3.21 EFlop/s
202377.1%4.8%3.9%1.1%13.0%5.03 EFlop/s
202484.5%3.4%2.7%0.3%9.1%7.16 EFlop/s
202586.3%1.4%2.7%0.3%9.4%7.01 EFlop/s
How this was calculated, and the inference behind most of these rows

Same construction as the other share charts: each in-service system contributes its last reported Rmax to every year it was live, split evenly where more than one architecture is recorded.

Most cooling edges here are inferred from the integrator's cabinet platform, not stated per system. That inference is worth being explicit about. Operators publish core counts and interconnect families; they publish cooling far less consistently. But the cabinet determines the architecture: every HPE Cray EX cabinet is 100% direct liquid cooled, every BullSequana XH2000 and XH3000 is a direct liquid cooled platform, Lenovo's SD650 is Neptune warm water, the Cray XT/XE/XK/XC generations moved air across in-cabinet water coils.

So where the platform is unambiguous we record the architecture with the integrator as supplier, confidence: reported, and the basis written into the edge's raw_string (you can read "HPE Cray EX cabinet) 100% direct liquid cooled" on the system page and judge the inference yourself. Where the platform does not settle it, we record nothing. Nine systems have no cooling edge for that reason and appear in the “Not recorded” band rather than being assigned a plausible value.

Installed megawatts use power_kw where an operator has published it. Only 29 of 52 systems do. A blank power figure means we found no source, not zero, so the megawatt totals below are a floor across the systems that disclose, never a census.

By architecture

ArchitectureSystemsDisclosing power InstalledMedian GF/WSpan
Direct liquid cooling, warm water 25 15 189.8 MW 14.8 2008–2025
Liquid-assisted air (in-cabinet coil) 6 5 29.1 MW 2.7 2009–2017
Not recorded 11 3 24.9 MW 0.6 1993–2024
Direct-to-chip cold plate 2 2 17.5 MW 13.7 2018–2018
Air cooling (CRAC/CRAH) 8 4 10.1 MW 2.7 1997–2023

Do not read the GF/W column as a cooling effect. Warm-water systems show roughly seven times the median efficiency of the liquid-assisted-air generation, and almost all of that is confounded with time: the warm-water systems span 2008–2025 and the in-cabinet-coil ones 2009–2017. What actually changed over that window is the silicon. Cooling architecture followed rack density rather than causing the efficiency gain, and this dataset cannot separate the two.

By cooling supplier

The company that supplied the cabinet, rolled up to today's parent. On direct-liquid platforms the integrator supplies the cooling, which is why this table looks like the integrator table rather than like a list of thermal vendors.

Supplier (today's parent)SystemsDisclosing powerInstalled
Hewlett Packard Enterprise 20 12 141.4 MW
Fujitsu 3 3 44.2 MW
IBM 6 6 32.1 MW
NRCPC 1 1 15.4 MW
Atos 5 2 9.3 MW
NVIDIA 2 1 2.6 MW
Lenovo 2 1 1.6 MW
Intel 1 0 -
NEC 1 0 -

The read

The chart has one clean transition and one honest hole.

The transition is physical rather than commercial. Air carries roughly 3,500 times less heat per unit volume than water, and the fan power needed to move it scales with about the cube of velocity. A hall handles 5 to 10 kW a rack comfortably and 20 kW with effort; a current accelerator cabinet is 40 to 400 kW. There was no decision point at which the industry preferred liquid: there was a density at which air stopped being an option. The Cray XT/XE/XC generation is the visible intermediate step: air, but moved across a water coil inside the cabinet.

The hole is the “Not recorded” band, and it is 11 systems wide. It sits mostly in the early years and among the machines whose operators published least, which is the same shape as the wider problem. This is the thinnest layer of the dataset precisely because it is the layer nobody discloses, and that is the argument for recording it at all.

What the megawatt column cannot show is the part that now binds. Of the 31 sites here, 2 publishes a PUE and almost none publish substation capacity or utility. Yet lead times on medium-voltage switchgear and large transformers have run to years, and in several markets the utility interconnection queue is measured in multi-year increments. The constraint on the next generation of machines is increasingly a piece of electrical infrastructure that no ranking has ever had a column for.

The power and cooling chapter covers the engineering in detail: how warm-water cooling removes the chillers, what PUE conceals, and why heat reuse works in Kajaani and not in Arizona.