Supercomputers/Analysis/Sector mix
Analysis 10
Who operates compute: sector mix
In 2000, 99% of the installed FLOPS in this dataset sat in government labs and universities. Today it is 73%, and 29% of the systems in service are run by someone else, up from 9% in 2010. The operators that grew fastest are big tech, GPU clouds and AI labs, most of whose machines never submit a benchmark, so the FLOPS chart below is a floor for them, not a measurement. The machine-count chart is the fairer view of who is building, and the gap between the two is the part of the market a ranking cannot see.
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.
Read the shape, not the spikes. This is the installed base of the 1153 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.
- Government labs
- Research HPC
- Big tech
- AI labs
- Neoclouds
- Industry
- Finance
- Life sciences
Table view: Sector share of installed FLOPS
| Year | Government labs | Research HPC | Big tech | AI labs | Neoclouds | Industry | Finance | Life sciences | Installed total |
|---|---|---|---|---|---|---|---|---|---|
| 1993 | 92.0% | 8.0% | 0.0% | 0.0% | 0.0% | 0.0% | 0.0% | 0.0% | 355.5 GFlop/s |
| 1994 | 86.4% | 13.6% | 0.0% | 0.0% | 0.0% | 0.0% | 0.0% | 0.0% | 378.7 GFlop/s |
| 1995 | 79.8% | 20.2% | 0.0% | 0.0% | 0.0% | 0.0% | 0.0% | 0.0% | 693.1 GFlop/s |
| 1996 | 44.8% | 52.9% | 0.0% | 0.0% | 0.0% | 2.3% | 0.0% | 0.0% | 1.71 TFlop/s |
| 1997 | 66.0% | 32.6% | 0.0% | 0.0% | 0.0% | 1.4% | 0.0% | 0.0% | 2.78 TFlop/s |
| 1998 | 75.0% | 24.4% | 0.0% | 0.0% | 0.0% | 0.5% | 0.0% | 0.0% | 7.44 TFlop/s |
| 1999 | 68.1% | 31.5% | 0.0% | 0.0% | 0.0% | 0.5% | 0.0% | 0.0% | 8.55 TFlop/s |
| 2000 | 70.0% | 29.0% | 0.0% | 0.0% | 0.0% | 1.0% | 0.0% | 0.0% | 17.94 TFlop/s |
| 2001 | 60.6% | 38.8% | 0.0% | 0.0% | 0.0% | 0.5% | 0.0% | 0.0% | 32.15 TFlop/s |
| 2002 | 82.2% | 17.7% | 0.0% | 0.0% | 0.0% | 0.2% | 0.0% | 0.0% | 101.45 TFlop/s |
| 2003 | 74.0% | 24.5% | 0.0% | 0.0% | 0.0% | 1.5% | 0.0% | 0.0% | 155.23 TFlop/s |
| 2004 | 91.0% | 8.7% | 0.0% | 0.0% | 0.0% | 0.3% | 0.0% | 0.0% | 730.99 TFlop/s |
| 2005 | 80.6% | 10.4% | 8.4% | 0.0% | 0.0% | 0.6% | 0.0% | 0.0% | 1.09 PFlop/s |
| 2006 | 74.1% | 17.4% | 7.3% | 0.0% | 0.0% | 1.3% | 0.0% | 0.0% | 1.25 PFlop/s |
| 2007 | 35.4% | 56.0% | 3.4% | 0.0% | 0.0% | 5.2% | 0.0% | 0.0% | 2.70 PFlop/s |
| 2008 | 71.1% | 27.0% | 0.7% | 0.0% | 0.0% | 1.2% | 0.0% | 0.0% | 12.24 PFlop/s |
| 2009 | 62.5% | 36.1% | 0.5% | 0.0% | 0.0% | 0.8% | 0.0% | 0.0% | 17.23 PFlop/s |
| 2010 | 55.9% | 43.2% | 0.4% | 0.0% | 0.0% | 0.6% | 0.0% | 0.0% | 24.69 PFlop/s |
| 2011 | 37.0% | 58.1% | 0.8% | 0.0% | 0.0% | 4.2% | 0.0% | 0.0% | 43.05 PFlop/s |
| 2012 | 57.9% | 39.8% | 0.4% | 0.0% | 0.0% | 1.8% | 0.0% | 0.0% | 116.7 PFlop/s |
| 2013 | 36.9% | 61.2% | 0.5% | 0.0% | 0.0% | 1.4% | 0.0% | 0.0% | 195.1 PFlop/s |
| 2014 | 34.1% | 62.3% | 0.5% | 0.0% | 0.0% | 3.1% | 0.0% | 0.0% | 213.4 PFlop/s |
| 2015 | 31.0% | 63.3% | 0.4% | 0.0% | 0.0% | 5.3% | 0.0% | 0.0% | 272.6 PFlop/s |
| 2016 | 26.7% | 66.7% | 0.2% | 0.0% | 0.0% | 5.9% | 0.0% | 0.4% | 455.3 PFlop/s |
| 2017 | 34.5% | 60.2% | 0.2% | 0.0% | 0.0% | 4.8% | 0.0% | 0.3% | 603.7 PFlop/s |
| 2018 | 49.3% | 44.9% | 0.1% | 0.0% | 0.0% | 5.0% | 0.0% | 0.7% | 1.01 EFlop/s |
| 2019 | 45.1% | 45.3% | 0.3% | 0.0% | 0.6% | 7.4% | 0.6% | 0.7% | 1.13 EFlop/s |
| 2020 | 28.0% | 57.3% | 1.7% | 0.1% | 0.4% | 11.8% | 0.3% | 0.4% | 1.92 EFlop/s |
| 2021 | 30.1% | 51.2% | 6.3% | 0.1% | 0.3% | 11.3% | 0.3% | 0.5% | 2.44 EFlop/s |
| 2022 | 45.7% | 42.8% | 3.9% | 0.2% | 0.2% | 6.7% | 0.1% | 0.3% | 4.63 EFlop/s |
| 2023 | 46.2% | 32.2% | 12.2% | 0.5% | 1.0% | 7.7% | 0.1% | 0.2% | 6.98 EFlop/s |
| 2024 | 48.4% | 28.7% | 10.3% | 0.3% | 2.4% | 9.6% | 0.1% | 0.3% | 11.06 EFlop/s |
| 2025 | 39.2% | 34.9% | 9.9% | 0.2% | 5.2% | 10.2% | 0.2% | 0.3% | 14.01 EFlop/s |
| 2026 | 31.3% | 41.4% | 8.5% | 0.2% | 6.7% | 11.5% | 0.2% | 0.2% | 17.39 EFlop/s |
- Government labs
- Research HPC
- Big tech
- AI labs
- Neoclouds
- Industry
- Finance
- Life sciences
Table view: Sector share of systems in service
| Year | Government labs | Research HPC | Big tech | AI labs | Neoclouds | Industry | Finance | Life sciences | Systems in service |
|---|---|---|---|---|---|---|---|---|---|
| 1993 | 87.5% | 12.5% | 0.0% | 0.0% | 0.0% | 0.0% | 0.0% | 0.0% | 8 |
| 1994 | 75.0% | 25.0% | 0.0% | 0.0% | 0.0% | 0.0% | 0.0% | 0.0% | 8 |
| 1995 | 72.7% | 27.3% | 0.0% | 0.0% | 0.0% | 0.0% | 0.0% | 0.0% | 11 |
| 1996 | 56.3% | 37.5% | 0.0% | 0.0% | 0.0% | 6.3% | 0.0% | 0.0% | 16 |
| 1997 | 58.8% | 35.3% | 0.0% | 0.0% | 0.0% | 5.9% | 0.0% | 0.0% | 17 |
| 1998 | 55.0% | 40.0% | 0.0% | 0.0% | 0.0% | 5.0% | 0.0% | 0.0% | 20 |
| 1999 | 54.5% | 40.9% | 0.0% | 0.0% | 0.0% | 4.5% | 0.0% | 0.0% | 22 |
| 2000 | 53.6% | 39.3% | 0.0% | 0.0% | 0.0% | 7.1% | 0.0% | 0.0% | 28 |
| 2001 | 51.6% | 41.9% | 0.0% | 0.0% | 0.0% | 6.5% | 0.0% | 0.0% | 31 |
| 2002 | 56.4% | 38.5% | 0.0% | 0.0% | 0.0% | 5.1% | 0.0% | 0.0% | 39 |
| 2003 | 52.2% | 37.0% | 0.0% | 0.0% | 0.0% | 10.9% | 0.0% | 0.0% | 46 |
| 2004 | 52.9% | 37.3% | 0.0% | 0.0% | 0.0% | 9.8% | 0.0% | 0.0% | 51 |
| 2005 | 50.0% | 38.3% | 1.7% | 0.0% | 0.0% | 10.0% | 0.0% | 0.0% | 60 |
| 2006 | 48.4% | 39.1% | 1.6% | 0.0% | 0.0% | 10.9% | 0.0% | 0.0% | 64 |
| 2007 | 40.0% | 46.7% | 1.3% | 0.0% | 0.0% | 12.0% | 0.0% | 0.0% | 75 |
| 2008 | 37.5% | 51.1% | 1.1% | 0.0% | 0.0% | 10.2% | 0.0% | 0.0% | 88 |
| 2009 | 34.3% | 55.9% | 1.0% | 0.0% | 0.0% | 8.8% | 0.0% | 0.0% | 102 |
| 2010 | 34.2% | 56.8% | 0.9% | 0.0% | 0.0% | 8.1% | 0.0% | 0.0% | 111 |
| 2011 | 32.8% | 57.8% | 1.6% | 0.0% | 0.0% | 7.8% | 0.0% | 0.0% | 128 |
| 2012 | 30.4% | 60.7% | 1.8% | 0.0% | 0.0% | 7.1% | 0.0% | 0.0% | 168 |
| 2013 | 28.0% | 61.4% | 2.1% | 0.0% | 0.0% | 8.5% | 0.0% | 0.0% | 189 |
| 2014 | 25.2% | 63.8% | 1.9% | 0.0% | 0.0% | 8.6% | 0.0% | 0.5% | 210 |
| 2015 | 23.3% | 65.6% | 1.6% | 0.0% | 0.0% | 9.1% | 0.0% | 0.4% | 253 |
| 2016 | 24.0% | 64.0% | 1.4% | 0.0% | 0.0% | 9.9% | 0.0% | 0.7% | 292 |
| 2017 | 26.5% | 62.2% | 1.2% | 0.0% | 0.0% | 9.5% | 0.0% | 0.6% | 325 |
| 2018 | 27.5% | 61.0% | 1.1% | 0.0% | 0.0% | 9.6% | 0.0% | 0.8% | 364 |
| 2019 | 26.9% | 60.1% | 1.3% | 0.3% | 0.3% | 10.3% | 0.3% | 0.8% | 398 |
| 2020 | 25.7% | 59.6% | 1.8% | 0.5% | 0.2% | 11.3% | 0.2% | 0.7% | 443 |
| 2021 | 25.1% | 58.4% | 3.2% | 0.6% | 0.4% | 11.3% | 0.2% | 0.8% | 495 |
| 2022 | 24.2% | 57.4% | 3.9% | 0.9% | 0.6% | 12.2% | 0.2% | 0.7% | 542 |
| 2023 | 24.3% | 54.7% | 4.5% | 1.3% | 1.3% | 13.0% | 0.2% | 0.7% | 600 |
| 2024 | 23.1% | 53.0% | 6.5% | 1.3% | 2.2% | 12.6% | 0.1% | 1.0% | 675 |
| 2025 | 22.2% | 50.9% | 7.5% | 1.3% | 3.9% | 12.9% | 0.3% | 1.1% | 751 |
| 2026 | 21.8% | 49.1% | 8.4% | 1.4% | 5.0% | 12.8% | 0.3% | 1.3% | 776 |
How this was calculated
Every operator that runs a system in this dataset has exactly one sector, recorded with a one-line reason in the company_sectors table. A system inherits its operator's sector, and a sector is fixed for an operator across time: Preferred Networks is an AI lab in every year it appears, not a research centre in one and a company in another.
A system is in service from its first-operational date until its decommissioning date, or to the present. The first chart weights each in-service system by its last reported Rmax; systems with no Rmax carry no weight there and are excluded from its denominator. The second chart counts every in-service system once. Planned and under-construction systems appear in neither.
Eight sectors is the ceiling of the categorical palette this site validates for colour-vision safety, which is why the taxonomy stops at eight. Where an operator could fit two sectors, the call follows how it operates the systems listed, and the reason is in the table: NHN Cloud is a GPU-rental provider running a national allocation, so it is a GPU cloud, not big tech; Samsung's clusters serve its own chip and fab work, so it is industry; ABCI's operator AIST is a national institute that runs a shared research and industry system, so it sits with research.
The sectors, side by side
| Sector | Operators | Systems | In service | Installed Rmax | Countries | Planned or building | First system | Largest by Rmax |
|---|---|---|---|---|---|---|---|---|
| Government & national labs | 65 | 233 | 169 | 5.45 EFlop/s | 24 | 7 | 1976 | El Capitan |
| Universities & research centres | 285 | 568 | 381 | 7.21 EFlop/s | 62 | 21 | 1987 | LineShine |
| Big tech, internet & telecom | 38 | 99 | 65 | 1.47 EFlop/s | 19 | 20 | 2005 | Eagle |
| AI labs & model builders | 13 | 19 | 11 | 31.65 PFlop/s | 6 | 3 | 2019 | DeepL Mercury |
| GPU clouds & AI campus developers | 50 | 85 | 39 | 1.17 EFlop/s | 25 | 40 | 2019 | ISEG2 |
| Industry, energy & hardware vendors | 62 | 125 | 99 | 2.00 EFlop/s | 25 | 11 | 1996 | HPC7 |
| Finance | 6 | 7 | 2 | 28.13 PFlop/s | 4 | 0 | 2019 | Bank of America HPC-AI |
| Life sciences & health | 15 | 17 | 10 | 35.75 PFlop/s | 5 | 1 | 2014 | BioHive-2 |
Where each sector is
Systems by operator sector for the 14 countries with the most systems in this dataset, every status included.
| Country | Government labs | Research HPC | Big tech | AI labs | Neoclouds | Industry | Finance | Life sciences | Total |
|---|---|---|---|---|---|---|---|---|---|
| USA | 145 | 121 | 49 | 6 | 31 | 39 | 3 | 12 | 406 |
| Japan | 16 | 69 | 9 | 2 | 3 | 6 | · | 1 | 106 |
| Germany | 6 | 76 | 2 | 1 | · | 2 | · | · | 87 |
| China | 4 | 21 | 10 | 6 | 4 | 12 | 1 | · | 58 |
| United Kingdom | 8 | 42 | 1 | · | 3 | 1 | · | 2 | 57 |
| France | 15 | 8 | 1 | 1 | 2 | 13 | · | 1 | 41 |
| Canada | 6 | 13 | 3 | · | 2 | 4 | · | · | 28 |
| South Korea | 5 | 5 | 9 | · | 4 | 5 | · | · | 28 |
| Australia | 4 | 15 | · | · | 3 | 2 | · | · | 24 |
| India | 7 | 11 | 1 | · | 4 | 1 | · | · | 24 |
| Italy | 3 | 11 | 1 | · | · | 7 | · | · | 22 |
| Saudi Arabia | 1 | 7 | · | · | 4 | 10 | · | · | 22 |
| Brazil | 1 | 8 | · | · | 2 | 9 | · | · | 20 |
| Russia | 1 | 10 | 4 | · | · | · | 2 | · | 17 |
The read
Weighted by FLOPS, the story is the one every supercomputing list tells: the top of the installed base belongs to the public sector. Government labs held 70% of installed FLOPS in 2000, 56% in 2010 and 28% in 2020, as university and shared research centres took the middle of the market, then 31% today as the exascale machines arrived. Weighted by machines, 29% of systems in service are run by industry, big tech, GPU clouds, AI labs, finance and life sciences together, against 27% of the FLOPS. The totals are close; the composition is not, as the next paragraph shows.
The composition gap is a measurement artefact and a substantive fact at once. AI labs, GPU clouds, finance and life sciences together hold 8% of the systems in service and 7% of the FLOPS this chart can see, because most of their machines never submit a benchmark run. The FLOPS chart is a floor for those sectors, not a measurement, and the shadow list is the attempt to put a bounded estimate under it.
Big tech is the sector to watch: 99 systems in 19 countries and 65 in service, with 20 more planned or under construction. The GPU-cloud sector, 85 systems from 50 operators, did not exist in this dataset before 2019. Its systems are counted here as operated by the cloud, not by the customer that rents them, which is why one AI lab's training cluster can appear under a neocloud.
Sector is assigned per operator and is a judgement, not a fact the operator states. The reason for each call is one line in the downloadable table, and the ambiguous ones are recorded as such. If a call looks wrong, it is a one-row correction.