Supercomputers/Analysis/Who pays
Analysis 17
Who pays for supercomputers
For most of the period this dataset covers, supercomputers were paid for by the state: 81% of the systems first run in the 2010s that have a funder on record were funded by a national, defence, EU or regional body, and only 1% by a company. That held into the early 2020s, at 79% public and 7% company money. It breaks with the systems first running or announced from 2025: of the 118 with a recorded funder, 45 (38%) are paid for by companies and private capital, another 13% mix public and private money, and public funders alone account for 45%. The funder records make the shift look like a step. The operator records, which exist for every system, show a steadier climb: companies run 14% of the systems first run in the 2010s, 35% of those from 2020 to 2024 and 64% of those from 2025, and the company-run machines of earlier years almost never say who paid for them.
Most systems have no funder on record, and the gap is not random. 481 of the 1581 systems in this dataset carry a funder or funding-programme row (395 a funder row, 86 a programme only). Coverage is 36% for systems run by government labs and research centres and 18% for systems run by companies, because public funders publish their awards, and companies rarely say how a cluster was financed. Of the 1100 systems with no funding row, 377 are run by companies. The shares on this page are shares of the systems that have a record, which tilts every period towards public money. As a bound, not a measurement: if every company-run system from 2025 onward without a record were also company-paid, the private share of that period would be 71% rather than 38%.
Researched with AI assistance, checked by scripts. These rows were researched by AI agents reading public operator, laboratory and vendor sources, appended by scripts and checked by automated validation; no subject-matter expert has reviewed them (how this site is made). A mechanical second-reader pass has since fetched every cited source: 1,350 of 1,581 systems have a readable citation that names them, and 64 are genuinely weakly sourced. Every claim carries its source and a confidence tier. Treat anything below verified as a lead, not a citation.
- Defence
- National civil
- EU & multinational
- Regional
- Own funds
- Philanthropy
- Mixed
- Private
- Unclassified
Table view: Funder type by period
| First light or announcement | Defence | National civil | EU & multinational | Regional | Own funds | Philanthropy | Mixed | Private | Unclassified | Systems |
|---|---|---|---|---|---|---|---|---|---|---|
| Before 2000 | 4 (50%) | 4 (50%) | · | · | · | · | · | · | · | 8 |
| 2000 to 2009 | 4 (10%) | 20 (49%) | 1 (2%) | 3 (7%) | 6 (15%) | · | 5 (12%) | 2 (5%) | · | 41 |
| 2010 to 2019 | 14 (10%) | 87 (64%) | 7 (5%) | 3 (2%) | 15 (11%) | · | 7 (5%) | 2 (1%) | 2 (1%) | 137 |
| 2020 to 2024 | 20 (15%) | 63 (47%) | 19 (14%) | 5 (4%) | 10 (7%) | 4 (3%) | 4 (3%) | 9 (7%) | 1 (1%) | 135 |
| 2025 onward | 4 (3%) | 28 (24%) | 18 (15%) | 3 (3%) | 2 (2%) | · | 15 (13%) | 45 (38%) | 3 (3%) | 118 |
How this was calculated
Every system with at least one funder or funding_program row in the
profile table is included: 481 of 1581. Those rows are short sentences, such
as “EuroHPC Joint Undertaking, half of the funding” or “$2.15 billion of senior secured notes”.
A fixed, ordered set of keyword rules reads every row and notes each kind of funder it names:
national civil government (science agencies, ministries, national programmes, sovereign funds
and majority-state-owned companies), defence and nuclear-weapons agencies, EU and multinational
bodies (EuroHPC, PRACE, EU structural funds, consortia of several states), regional, state and
city government, a university or institute's own budget (including US national-lab
institutional computing), named philanthropic foundations, and companies or private capital
(named firms, debt, equity, joint ventures).
The kinds named for a system then reduce to one type. Public and private money together, or a row that says “public-private”, is mixed. Otherwise the widest public tier wins (EU and multinational, then national, then regional), then companies, then philanthropy, then the institution's own funds. A system run by a company whose rows name only public money is also mixed, because the public part there is a subsidy, tax credit or award on top of the company's own spending; the exceptions are hardware the state bought and consigned to an operator, and operators the state owns. Vendor co-investment in a public machine, such as donated accelerators, counts as private money, so a few government-lab systems are mixed. Rows that name a delivery contract, a memorandum of understanding or a government customer are not read as funding. 6 systems name no funder the rules recognise and stay unclassified.
The period is the year of first operation, or of announcement for a system not yet running. The 2020s are split at 2025 because that is where the change is concentrated; before 2000 is pooled because only 8 systems that early have a funder on record. Coverage by period runs from 8 of 60 systems before 2000 to 118 of 305 from 2025. The operator cross-check in the opening paragraph uses the same periods over every system with a known operator sector, and counts big tech, AI labs, GPU clouds, industry, finance and life sciences as companies.
Every distinct funder and programme sentence was printed with the type its rules give and read, and wrong rules were fixed. Then 3 samples of 40 systems each were drawn at random with fixed seeds and their raw text read against the type assigned: 118 of 120 were classified correctly, and each of the 2 misses led to a rule fix. Some calls are judgements the rules make explicit: Petrobras, which the Brazilian state controls, counts as national money; Helmholtz and Max Planck budgets count as the institute's own funds.
Who has a funder on record
Systems in this dataset by operator sector, and how many have a funder or funding-programme row. Systems whose operator has no sector are left out of this table.
| Operator sector | Systems | With a funding row | Share |
|---|---|---|---|
| Research HPC | 773 | 295 | 38% |
| Government labs | 348 | 103 | 30% |
| Industry | 159 | 13 | 8% |
| Neoclouds | 138 | 40 | 29% |
| Big tech | 116 | 23 | 20% |
| AI labs | 21 | 4 | 19% |
| Life sciences | 18 | 2 | 11% |
| Finance | 8 | 1 | 13% |
- Defence
- National civil
- EU & multinational
- Regional
- Own funds
- Philanthropy
- Mixed
- Private
- Unclassified
Table view: Funder type by country
| Country | Defence | National civil | EU & multinational | Regional | Own funds | Philanthropy | Mixed | Private | Unclassified | Systems |
|---|---|---|---|---|---|---|---|---|---|---|
| USA | 48 (26%) | 69 (38%) | · | 4 (2%) | 16 (9%) | 4 (2%) | 9 (5%) | 34 (18%) | · | 184 |
| Germany | · | 36 (68%) | 5 (9%) | 3 (6%) | 6 (11%) | · | 2 (4%) | 1 (2%) | · | 53 |
| United Kingdom | · | 23 (72%) | 1 (3%) | 1 (3%) | 2 (6%) | · | 2 (6%) | 3 (9%) | · | 32 |
| Canada | · | 10 (56%) | · | 1 (6%) | · | · | 4 (22%) | 3 (17%) | · | 18 |
| Australia | · | 6 (40%) | · | 1 (7%) | 6 (40%) | · | · | 2 (13%) | · | 15 |
| India | · | 12 (80%) | · | · | · | · | · | 2 (13%) | 1 (7%) | 15 |
| Japan | · | 9 (64%) | 1 (7%) | · | 1 (7%) | · | 3 (21%) | · | · | 14 |
| Sweden | · | 5 (36%) | 2 (14%) | · | 2 (14%) | 1 (7%) | 2 (14%) | 2 (14%) | · | 14 |
| France | · | 8 (62%) | 1 (8%) | · | · | · | 2 (15%) | 2 (15%) | · | 13 |
| South Korea | · | 5 (42%) | · | · | 1 (8%) | · | 3 (25%) | 2 (17%) | 1 (8%) | 12 |
By region
| Region | Defence | National civil | EU & multinational | Regional | Own funds | Philanthropy | Mixed | Private | Unclassified | Systems |
|---|---|---|---|---|---|---|---|---|---|---|
| North America | 48 (23%) | 82 (40%) | · | 5 (2%) | 16 (8%) | 4 (2%) | 13 (6%) | 37 (18%) | · | 205 |
| Europe | · | 85 (49%) | 44 (25%) | 5 (3%) | 14 (8%) | 1 (1%) | 10 (6%) | 11 (6%) | 3 (2%) | 173 |
| Asia-Pacific | · | 39 (53%) | 2 (3%) | 4 (5%) | 10 (14%) | · | 6 (8%) | 10 (14%) | 2 (3%) | 73 |
| South America | · | 13 (68%) | 1 (5%) | · | 1 (5%) | · | 2 (11%) | 2 (11%) | · | 19 |
| Middle East | · | 3 (43%) | · | · | · | · | 1 (14%) | 2 (29%) | 1 (14%) | 7 |
| Africa | · | 2 (50%) | 1 (25%) | · | · | · | 1 (25%) | · | · | 4 |
The funders and programmes named most often
Funding programmes and agencies grouped across spellings (“ASC”, “ASCI”, “CTS-2 under NNSA ASC” are one family), counting each system once whether the name is in its funder or its programme row. Families named for at least three systems are shown. A system can appear under several families.
| Funder or programme | Kind | Systems | Countries | Most common type among them |
|---|---|---|---|---|
| US National Science Foundation awards (all programmes) | Agency | 46 | 2 | National civil |
| NNSA Advanced Simulation and Computing (ASC, formerly ASCI) | Programme | 36 | 1 | Defence |
| EuroHPC Joint Undertaking | Programme | 24 | 15 | EU & multinational |
| UK Research and Innovation councils (EPSRC, STFC, NERC, MRC) | Agency | 24 | 1 | National civil |
| EU structural and regional funds (ERDF, operational programmes) | Programme | 14 | 10 | EU & multinational |
| German Research Foundation (DFG) instrumentation grants | Agency | 14 | 1 | National civil |
| US national-lab institutional computing (incl. LLNL M&IC) | Programme | 14 | 1 | Defence |
| India National Supercomputing Mission | Programme | 11 | 1 | National civil |
| Canada Foundation for Innovation and Digital Research Alliance | Agency | 10 | 1 | National civil |
| US DOE Office of Science | Agency | 10 | 1 | National civil |
| EuroHPC AI Factories | Programme | 9 | 9 | EU & multinational |
| Gauss Centre for Supercomputing (German federal and state) | Programme | 9 | 1 | National civil |
| GENCI (France) | Agency | 9 | 1 | National civil |
| US DoD High Performance Computing Modernization Program | Programme | 9 | 1 | Defence |
| Japan MEXT (incl. HPCI) | Agency | 8 | 2 | National civil |
| Swedish SNIC and NAISS | Agency | 8 | 1 | National civil |
| Helmholtz Association | Institute | 7 | 1 | National civil |
| Korea Ministry of Science and ICT GPU programmes | Programme | 7 | 1 | National civil |
| PRACE and PRACE-LAB | Programme | 6 | 3 | EU & multinational |
| German National High Performance Computing (NHR) alliance | Programme | 5 | 1 | National civil |
| US CARES Act and Recovery Act (ARRA) | Programme | 5 | 1 | National civil |
| Max Planck Society and institutes | Institute | 4 | 1 | Own funds |
| UK DiRAC | Programme | 4 | 1 | National civil |
| Australian NCRIS and Super Science | Programme | 3 | 1 | National civil |
| National recovery and resilience plans (EU Recovery and Resilience Facility) | Programme | 3 | 2 | National civil |
| US National AI Research Resource (NAIRR) | Programme | 3 | 1 | National civil |
Stated cost by funder type
129 systems state a cost, 84 of them with a funder on record. The scope of a stated cost is not consistent: for a public machine it is usually the purchase price or a grant that may include years of operation; for a company campus it is usually the whole build, with buildings, power and cooling. The medians below compare what was stated, not like with like, and should be read as a measure of the scale of the projects each type of funder announces. Types with fewer than three stated costs are left out.
| Funder type | Systems with a cost | Median stated cost | Lowest | Highest |
|---|---|---|---|---|
| Defence and nuclear-weapons agencies | 10 | US$64 m | US$1.5 m | US$600 m |
| National government, civil (science agencies, ministries, state funds) | 39 | US$10 m | US$0.4 m | US$600 m |
| University or institute own funds | 8 | US$9.5 m | US$0.4 m | US$33 m |
| Mixed public and private | 6 | US$775 m | US$2.0 m | US$8.6 bn |
| Companies and private capital | 15 | US$6.7 bn | US$30 m | US$27.0 bn |
The read
The public-to-private shift is in the data, and in the funder records it looks like a step. Through the 2010s and into the early 2020s the systems with a funder on record were paid for almost entirely by the state, and the company share never passed 7%. The change arrives with the cohort first running or announced from 2025, and much of it has not been built: 33 of the 45 company-paid systems in that cohort are planned or under construction. Of the 62 company-paid systems across all periods, 28 are run by neoclouds, 16 by big tech and 10 by industry, and the rows of 26 of them describe debt, notes, outside equity or a joint venture rather than one company paying from its own cash.
The records are uneven in a way that shapes this picture. Public funders publish who paid; most companies do not: only 13 of 159 systems run by industry, 40 of 138 systems run by neoclouds and 23 of 116 systems run by big tech in this dataset carry a funding row. The bound in the box above shows how far that could move the latest period, and the earlier decades are worse: of the 65 company-run systems first run in the 2010s, none has a funding row. Running a machine is not paying for it, but the operator series says company compute was already a growing part of the dataset before 2025. What changed is that private money now shows up in the funding record, mostly as project debt and equity announced in public, rather than staying inside a company's general capital spending.
The public money is not one kind of money. In the United States 26% of systems with a funder record are paid for by defence and nuclear-weapons budgets: the NNSA's Advanced Simulation and Computing programme is named for 36 systems and the DoD's High Performance Computing Modernization Program for 9. Every defence-funded system on record here is American, though that partly reflects which governments publish such funding.In Europe 25% are co-funded by the EU or a group of states; the EuroHPC Joint Undertaking alone is named for 24 systems, and its rows mostly state a share of between a third and a half of a machine's cost, with the hosting country or consortium paying the rest. In Asia-Pacific 53% are paid for by national governments, through programmes such as India's National Supercomputing Mission and Japan's MEXT.
The costs differ by orders of magnitude: the median company-paid project with a stated cost is US$6.7 bn, against US$10 m for national civil funders. Most of that gap is scope, a campus against a machine, which is why this page does not put the two on one chart.
Funder type is assigned by keyword rules over short text, and the rules are a judgement about what each phrase means, not a statement from the funder. Every rule is listed in the source of this page's module; if a call looks wrong, it is a one-line fix.