Supercomputers/Analysis/Access
Analysis 18
Who can get time on a supercomputer
Of the 439 systems in service in this dataset that record how people get time on them, 37% take applications from outside researchers through an open call and another 19% can be rented by anyone who pays: 56% in all. Weighted by measured performance, across the 222 of them with an HPL result, the share an outsider could apply for or rent is 65%. Open calls are a public-sector route and rental a private-sector one, and the restricted remainder is concentrated in a few very large machines: 43 restricted systems hold 26% of that measured compute, led by El Capitan.
Most systems have no access record yet, including the largest. Only 573 of the 1581 systems in the dataset (36%) carry an access-route row, and 439 of the 1093 in service. Those rows cover 55% of the measured performance in service, so the rest is unknown, not zero. The largest measured system in service, LineShine, has no access record, and the gap is widest for AI labs and Industry operators, where 14% and 15% of systems are covered. A system also counts in the most open class its text supports, so a machine that gives only part of its time to a national call counts as open in full: the open share is an upper bound on the capacity outsiders actually receive.
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.
- Open allocation or rentable
- Other routes (own users, restricted, lease, other)
Table view: Share of systems in service an outsider could apply for or rent
| Operator sector | Systems in service | Open allocation | Rentable | Apply or rent | Measured systems | Apply or rent, by Rmax |
|---|---|---|---|---|---|---|
| All sectors | 439 | 36.9% | 18.7% | 55.6% | 222 | 65.5% |
| Government labs | 73 | 23.3% | 4.1% | 27.4% | 47 | 31.8% |
| Research HPC | 264 | 52.7% | 6.4% | 59.1% | 134 | 89.5% |
| Big tech | 39 | 2.6% | 71.8% | 74.4% | 15 | 97.0% |
| Neoclouds | 30 | 6.7% | 80.0% | 86.7% | 11 | 100.0% |
| Industry | 20 | 5.0% | 35.0% | 40.0% | 11 | 0.0% |
| Life sciences | 8 | 25.0% | 12.5% | 37.5% | 2 | too few |
How this was calculated
Each system's access-route text, written during research from public sources (460 of the 573 rows are operator-stated; the rest come from press reports, vendors or filings), is put into one class by an ordered list of 33 keyword rules, first match wins. Explicit single-tenant and no-outside-access statements are tested first, then named allocation programmes (ACCESS and its predecessors, INCITE, EuroHPC, PRACE, HPCI, NAIRR, AIRR and national services such as NAISS, Sigma2 or the Digital Research Alliance of Canada) and open-call wording, then paid and cloud access, then testbeds and early access, then own-institution wording, then internal and mission use. A state-wide scheme open to researchers at every university in a region counts as open. A university charging its own researchers counts as own users; one selling time to outside organisations counts as rentable. The last own-users rule is deliberately broad, which is why no system is left unclassified; its errors land in own users.
The rules were tuned by printing every one of the 573 raw texts with its class and reading them all. A sample of 40 systems drawn with a fixed seed was then read by hand and judged against the class definitions: 38 of 40 agreed. The rules were fixed for the 2 that did not, and a second sample of 20 different systems agreed in 20 of 20. The hardest boundaries are open allocation against own users for regional consortia, and rentable against open for national centres that both run calls and charge.
"In service" means a recorded status of operational. Performance weighting uses Rmax, the measured HPL result, for the systems that have one; a system without one counts in the system shares and not in the weighted shares. The class describes access as the record states it now, or when the system ran, and does not track a change of model over a system's life. The downloadable file carries counts and shares only.
The access classes
| Class | What it means | Systems with a record | In service | Share of systems in service | Share of measured Rmax in service |
|---|---|---|---|---|---|
| Open allocation to outside researchers | Outside researchers can apply through a published call or allocation process (national or international programmes such as ACCESS, INCITE, EuroHPC, PRACE, HPCI, NAIRR or AIRR, or a national service open to every researcher in a country). | 248 | 162 | 36.9% | 49.0% |
| Commercial or paid rental | Anyone who pays can use it: public cloud instances, GPU clouds, reserved capacity, or a centre that sells time to outside organisations. | 94 | 82 | 18.7% | 16.4% |
| Leased or contracted to one customer | The whole system or campus is built for, leased to or contracted to a single customer. | 11 | 9 | 2.1% | none measured |
| Institutional: own users and collaborators | The operating university, laboratory or consortium's own researchers, plus sponsored collaborators and members who buy in. | 149 | 127 | 28.9% | 6.5% |
| Restricted: classified, mission or company-internal | Classified or defence work, operational mission use (weapons, weather, agency programmes) or a company running its own workloads, with no outside route described. | 48 | 43 | 9.8% | 26.3% |
| Other: testbeds, early access, mixed or not yet set | Testbeds and early-access systems for selected teams, systems split between own use and a small number of named partners, and future systems whose access model is not yet described. | 23 | 16 | 3.6% | 1.8% |
How much one machine decides
The weighted answer rests on few machines. Of the measured compute in service that has an access record, El Capitan alone holds 21% and the five largest systems 55%. Taking out each system in turn and recomputing, the largest swing comes from El Capitan: without it the share an outsider could apply for or rent rises from 65% to 82%, 17 points. The largest swing the other way is Aurora, whose removal lowers it to 61%, 4 points. So the system-count share is the steadier number, and the weighted one moves with a single machine: the restricted El Capitan is the main reason it is not higher.
The largest systems in each class
| Class | Largest measured systems in service | In service without an Rmax |
|---|---|---|
| Open allocation to outside researchers | Aurora (1.01 EFlop/s) ; JUPITER (1.00 EFlop/s) ; Fugaku (442.0 PFlop/s) ; Alps (434.9 PFlop/s) ; Isambard-AI Phase 2 (216.5 PFlop/s) | 66 |
| Commercial or paid rental | Eagle (561.2 PFlop/s) ; CHIE-4 (135.4 PFlop/s) ; MAXIMUS-384 (114.5 PFlop/s) ; CHIE-2 (89.78 PFlop/s) ; Shakti Cloud (84.31 PFlop/s) | 53 |
| Leased or contracted to one customer | none has a measured Rmax | 9 |
| Institutional: own users and collaborators | Tuolumne (208.1 PFlop/s) ; Shaheen III - GPU (122.8 PFlop/s) ; Shaheen III (35.66 PFlop/s) ; HiPerGator AI (17.49 PFlop/s) ; Kempner AI Cluster (16.29 PFlop/s) | 70 |
| Restricted: classified, mission or company-internal | El Capitan (1.81 EFlop/s) ; Eos (121.4 PFlop/s) ; Harpia (75.20 PFlop/s) ; El Dorado (68.02 PFlop/s) ; Crossroads (30.03 PFlop/s) | 13 |
| Other: testbeds, early access, mixed or not yet set | Israel-1 (41.50 PFlop/s) ; Helma (32.22 PFlop/s) ; Polaris (25.81 PFlop/s) ; Taipei-1 (22.30 PFlop/s) ; Redtail (12.85 PFlop/s) | 6 |
None of the 11 single-tenant systems with an access record has a measured Rmax: the AI campuses built for one customer have not run the benchmark, so the weighted shares on this page cannot see them at all.
- Open allocation
- Rentable
- Single-tenant lease
- Own users
- Restricted
- Other
Table view: Access class of systems in service, 2008–2026
| Year | Open allocation | Rentable | Single-tenant lease | Own users | Restricted | Other | Systems with a record |
|---|---|---|---|---|---|---|---|
| 2008 | 50.0% | 3.8% | 0.0% | 34.6% | 7.7% | 3.8% | 26 |
| 2009 | 51.9% | 3.7% | 0.0% | 33.3% | 7.4% | 3.7% | 27 |
| 2010 | 53.1% | 3.1% | 0.0% | 31.3% | 9.4% | 3.1% | 32 |
| 2011 | 54.1% | 5.4% | 0.0% | 27.0% | 10.8% | 2.7% | 37 |
| 2012 | 58.6% | 3.4% | 0.0% | 27.6% | 6.9% | 3.4% | 58 |
| 2013 | 60.0% | 4.6% | 0.0% | 26.2% | 6.2% | 3.1% | 65 |
| 2014 | 55.8% | 3.9% | 0.0% | 31.2% | 6.5% | 2.6% | 77 |
| 2015 | 56.3% | 5.2% | 0.0% | 29.2% | 7.3% | 2.1% | 96 |
| 2016 | 56.0% | 4.6% | 0.0% | 31.2% | 6.4% | 1.8% | 109 |
| 2017 | 54.5% | 4.9% | 0.0% | 31.7% | 7.3% | 1.6% | 123 |
| 2018 | 54.3% | 5.8% | 0.0% | 30.4% | 8.0% | 1.4% | 138 |
| 2019 | 52.3% | 7.8% | 0.0% | 30.7% | 7.8% | 1.3% | 153 |
| 2020 | 51.2% | 8.3% | 0.0% | 30.4% | 8.3% | 1.8% | 168 |
| 2021 | 52.8% | 9.0% | 0.5% | 28.1% | 7.0% | 2.5% | 199 |
| 2022 | 50.4% | 11.2% | 0.4% | 25.9% | 8.5% | 3.6% | 224 |
| 2023 | 49.2% | 11.4% | 0.4% | 24.4% | 11.0% | 3.5% | 254 |
| 2024 | 47.0% | 13.8% | 0.3% | 23.5% | 11.4% | 4.0% | 298 |
| 2025 | 43.6% | 17.1% | 1.4% | 22.8% | 11.3% | 3.8% | 346 |
| 2026 | 43.4% | 18.1% | 2.2% | 21.3% | 11.1% | 4.0% | 371 |
How this was calculated
A system is in service in a year if it first ran in or before that year and was not decommissioned before it. Each system keeps the one class its access record gives it for every year, so the chart shows how the mix of machines changed, not how any one machine's access policy changed. Earlier years rest on fewer records, mostly long-lived research systems, so the shape is the part to read.
Among systems with a record, rentable machines were 5% of those in service in 2015 and 18% in 2026, while open allocation went from 56% to 43%. That is the arrival of GPU clouds and hyperscaler capacity among machines that were mostly public research systems, and it also reflects which systems have had their access researched so far.
By region
| Region (systems in service) | Systems | Open | Rentable | Lease | Own users | Restricted | Other | Apply or rent | Apply or rent, by Rmax |
|---|---|---|---|---|---|---|---|---|---|
| North America | 160 | 21% | 16% | 5% | 33% | 21% | 5% | 37% | 42% |
| Europe | 151 | 53% | 13% | 1% | 29% | 1% | 3% | 66% | 96% |
| Asia-Pacific | 94 | 37% | 34% | 0% | 24% | 3% | 1% | 71% | 93% |
| Middle East | 19 | 21% | 21% | 0% | 32% | 16% | 11% | 42% | 2% |
| South America | 11 | 64% | 9% | 0% | 9% | 18% | 0% | 73% | 17% |
| Africa | 4 | 50% | 25% | 0% | 25% | 0% | 0% | 75% | too few |
By era of first operation
| First light (all systems with a record) | Systems | Open | Rentable | Lease | Own users | Restricted | Other | Apply or rent | Apply or rent, by Rmax |
|---|---|---|---|---|---|---|---|---|---|
| Before 2010 | 33 | 52% | 3% | 0% | 36% | 6% | 3% | 55% | 87% |
| 2010–2014 | 57 | 61% | 4% | 0% | 28% | 5% | 2% | 65% | 77% |
| 2015–2019 | 84 | 52% | 11% | 0% | 29% | 8% | 0% | 63% | 86% |
| 2020–2022 | 87 | 51% | 15% | 1% | 18% | 8% | 7% | 66% | 94% |
| 2023–2024 | 80 | 39% | 20% | 0% | 18% | 19% | 5% | 59% | 53% |
| 2025 onwards | 84 | 35% | 32% | 8% | 13% | 8% | 4% | 67% | 87% |
| No first-light date | 148 | 32% | 18% | 2% | 38% | 5% | 5% | 50% | 47% |
Shares are of systems; the last column weights by Rmax and reads "too few" where a group has fewer than three measured systems. The era table includes retired systems and systems not yet running, which is why its totals differ from the in-service tables.
What the records do not cover
These are the largest measured systems in service with no access record. Until they are researched, the shares above describe the systems that have been, not the dataset.
| System | Operator sector | Country | Rmax |
|---|---|---|---|
| LineShine | Research HPC | China | 2.20 EFlop/s |
| HPC7 | Industry | Italy | 571.5 PFlop/s |
| HPC6 | Industry | Italy | 477.9 PFlop/s |
| ISEG2 | Neoclouds | Netherlands | 202.4 PFlop/s |
| Discovery 6 | Industry | USA | 164.2 PFlop/s |
| Maximus-02 | Neoclouds | USA | 156.1 PFlop/s |
| SSC-24 | Industry | South Korea | 106.2 PFlop/s |
| Venado | Government labs | USA | 98.51 PFlop/s |
| Operator sector | Systems | With an access record | Coverage |
|---|---|---|---|
| Government labs | 348 | 105 | 30% |
| Research HPC | 773 | 348 | 45% |
| Big tech | 116 | 48 | 41% |
| AI labs | 21 | 3 | 14% |
| Neoclouds | 138 | 35 | 25% |
| Industry | 159 | 24 | 15% |
| Finance | 8 | 2 | 25% |
| Life sciences | 18 | 8 | 44% |
29% of ranked systems and 47% of unranked ones have an access record. 173 of the systems with an access record also link to the operator's own access page.
The read
Who can get time depends mostly on who owns the machine. Among university and research-centre systems in service, 59% take outside applications or rent time, and 39% serve only their own researchers and collaborators. Government labs look different: 27% are reachable from outside, and 37% are restricted to weapons, defence, weather or agency work. In the private sector the route is payment, not peer review: 72% of big-tech systems and 80% of GPU-cloud systems with a record can be rented, and 13% and 13% respectively are built or leased for a single customer.
By measured performance the answer looks more open than by count, because the largest public machines are allocated by national calls. But the measured share leans on few systems, and it is blind to the AI campuses that do not run the benchmark and to the 45% of measured compute in service that has no access record yet. The honest summary is that, of the compute this dataset can measure and has an access record for, most is open to an outside researcher in principle, through a competitive call or a credit card; how much of it any one researcher can get in practice is a question about oversubscription that these records do not answer.