Reports/AI compute versus public compute
Sample report · Data as of 8 November 2026
AI compute versus public compute
How the build-out of company-run AI clusters compares with the publicly funded supercomputing world: who is building, who pays, who can get time, what measured performance costs, what the machines are built for, and what the data cannot see. Drawn from the 1,581 systems in the Compute Atlas dataset as of 8 November 2026.
This is a sample. It shows the form of a commissioned Compute Atlas report: a defined question, findings with the number behind each one, the method and its limits, and the derived data. A commissioned report is scoped to the buyer's question and delivered as a document with its data files; see how commissioned reports work. Every figure below is computed from the dataset when the site is built, so the same page rebuilt on later data will give later numbers.
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.
The page prints as a plain document without the site navigation.
Summary
- Company-run operators account for 77% of the 159 systems planned or under construction in this dataset and 99.9% of the 117 GW of power capacity those projects state; public labs and universities state 87 MW across 5 projects.
- In service the balance is still the other way: public-sector machines hold 72% of the measured performance (Rmax) installed in 2026, while the company-run share has gone from 2% in 2010 to 28%.
- Companies pay for 38% of the systems first running or announced from 2025 that have a recorded funder, against 1% of those first run in the 2010s, and the share would be 71% if every company-run system without a funder record were company-paid too.
- Company-run machines with an access record are more often open to outsiders than public ones (67% against 52% of those in service), but almost entirely by paying: 61% are rentable and 6% take applications, where 46% of public machines allocate time through an open call.
- Measured performance keeps getting cheaper, but more slowly: the inflation-adjusted price of a GFlop/s fell tenfold every 3.7 years before 2010 and every 10.3 years since, to a median of $3.8 in 2025 dollars for machines first run from 2020.
- AI machines are built and quoted differently: AI training systems first run since 2020 carry a median 0.52 PB of storage per PFlop/s against 2.5 for scientific machines, and on the 33 systems where both can be computed, the vendor's dense FP8 peak is a median 47 times the measured FP64 Rmax.
- The data sees the public side better: 82% of public machines in service report a measured Rmax against 62% of company-run ones, and the unranked clusters whose disclosed accelerators can be bracketed against measured systems are estimated at 17.93 EFlop/s to 73.72 EFlop/s, against 18.70 EFlop/s for every ranked system here combined.
The two sides
This report splits every system by who operates it, because operator is the one attribute recorded for every system in the dataset. Public sector means government and national labs, universities and research centres. Company-run means big tech, ai labs, gpu clouds and ai campus developers, industry, finance and life sciences. The labels are about operators, not money: a university machine can be paid for by a company, and a company data centre can carry a state subsidy. Who actually pays is section 2.
Company-run is also not a synonym for AI. Of the 460 company-run systems, 263 (57%) are classed as AI training or inference and 89 as mixed HPC and AI; the rest are classed as HPC. On the public side 64 systems (6%) are AI machines and 240 are mixed. Where a finding is about AI machines specifically, it says so.
| At a glance | Public sector | Company-run |
|---|---|---|
| Systems in the dataset | 1121 | 460 |
| In service in 2026 | 753 | 272 |
| Installed measured performance (Rmax) in service | 12.88 EFlop/s | 4.96 EFlop/s |
| AI training or inference systems | 64 | 263 |
| Planned or under construction | 36 | 123 |
| Stated capacity of that pipeline | 87 MW | 117 GW |
| Recorded power draw of machines in service (systems with a figure) | 526 MW (360) | 84 MW (69) |
1. Who is building
Companies run a growing share of the machines in service: 7% of the systems in service in 2010 and 27% in 2026. Their share of installed measured performance moved up with it, from 2% to 28%, so public machines still hold most of the measured compute, at 72%. The Rmax line is a floor for the company side, because most company-run AI clusters never run the benchmark (section 5).
- Share of systems in service
- Share of installed Rmax
Table view: Company-run share of systems in service and of installed Rmax
| Year | Company-run, systems | Company-run, Rmax | Systems in service | Installed Rmax |
|---|---|---|---|---|
| 2000 | 3.9% | 0.8% | 51 | 21.03 TFlop/s |
| 2001 | 3.6% | 0.4% | 55 | 39.15 TFlop/s |
| 2002 | 2.8% | 0.2% | 71 | 116.73 TFlop/s |
| 2003 | 7.3% | 2.3% | 82 | 181.27 TFlop/s |
| 2004 | 6.4% | 0.5% | 94 | 803.53 TFlop/s |
| 2005 | 7.5% | 8.2% | 107 | 1.21 PFlop/s |
| 2006 | 7.5% | 7.6% | 120 | 1.43 PFlop/s |
| 2007 | 7.9% | 7.5% | 139 | 3.12 PFlop/s |
| 2008 | 7.6% | 3.2% | 170 | 13.57 PFlop/s |
| 2009 | 7.3% | 2.4% | 192 | 19.64 PFlop/s |
| 2010 | 7.4% | 2.3% | 215 | 30.16 PFlop/s |
| 2011 | 8.4% | 6.0% | 239 | 50.16 PFlop/s |
| 2012 | 8.2% | 3.3% | 304 | 131.0 PFlop/s |
| 2013 | 10.0% | 3.0% | 339 | 218.1 PFlop/s |
| 2014 | 10.4% | 4.7% | 376 | 248.4 PFlop/s |
| 2015 | 10.3% | 6.1% | 435 | 323.1 PFlop/s |
| 2016 | 11.1% | 6.9% | 477 | 509.4 PFlop/s |
| 2017 | 11.3% | 6.4% | 523 | 671.7 PFlop/s |
| 2018 | 11.8% | 7.0% | 570 | 1.09 EFlop/s |
| 2019 | 13.1% | 10.8% | 602 | 1.22 EFlop/s |
| 2020 | 14.1% | 15.1% | 651 | 2.02 EFlop/s |
| 2021 | 15.5% | 18.8% | 702 | 2.55 EFlop/s |
| 2022 | 16.9% | 11.7% | 750 | 4.74 EFlop/s |
| 2023 | 19.0% | 21.6% | 810 | 7.10 EFlop/s |
| 2024 | 21.7% | 23.1% | 904 | 11.23 EFlop/s |
| 2025 | 24.6% | 26.1% | 985 | 14.29 EFlop/s |
| 2026 | 26.5% | 27.8% | 1025 | 17.84 EFlop/s |
The future looks different. Of the 159 systems recorded as planned or under construction, 123 (77%) are company-run, and 102 of them state a power capacity, together 117 GW, of which 48% is described as under construction. The largest share by far belongs to neoclouds, at 84 GW. On the public side, 36 projects are in the pipeline and 5 state a capacity, together 87 MW.
The two sides do not describe their projects in the same terms, and that is part of the finding. A public machine is announced as a system with a performance target; a company campus is announced as a site with a power figure. For scale only: the company-run pipeline states about 222 times the 526 MW of power recorded for the 360 public machines in service that have a power figure. The two numbers are not like for like (one is mostly facility or unspecified capacity that has not been built, the other is measured system draw for fewer than all machines), but the gap is wide enough that the comparison survives the mismatch. 77% of the stated pipeline capacity does not say whether it is IT load or facility power.
- Under construction
- Planned
Table view: Stated capacity of the build pipeline by operator sector
| Sector | Projects stating capacity | Under construction | Planned | Total | Share |
|---|---|---|---|---|---|
| Neoclouds | 70 | 37 GW | 47 GW | 84 GW | 72% |
| Big tech | 19 | 15 GW | 2.6 GW | 18 GW | 15% |
| AI labs | 4 | 600 MW | 5.6 GW | 6.2 GW | 5% |
| Finance | 1 | 0 MW | 5.0 GW | 5.0 GW | 4% |
| Industry | 8 | 3.3 GW | 108 MW | 3.4 GW | 3% |
| Research HPC | 4 | 0 MW | 72 MW | 72 MW | under 1% |
| Government labs | 1 | 0 MW | 15 MW | 15 MW | under 1% |
2. Who pays
Of the systems first run in the 2010s that have a funder on record, 81% were paid for by national, defence, EU or regional public bodies and 1% by companies. The early 2020s were similar, at 79% public and 7% company money. For systems first running or announced from 2025, 45 of the 118 with a funder record (38%) are paid for by companies and private capital, 13% mix public and private money, and 45% are public alone.
Funder records are thinnest where the change is happening: 18% of company-run systems have a funder record, against 36% of public ones. Operator records, which exist for every system, show the same move more steadily: companies run 14% of the systems first run in the 2010s, 35% of those from 2020 to 2024 and 64% of those from 2025. If every company-run system from 2025 without a funder record were company-paid, the company share for that period would be 71% rather than 38%. The true figure is somewhere between.
The two sides also buy different things. The median stated cost of a company-paid system is US$6.7 bn (15 systems), against US$10 m for one paid for by a national civil funder (39 systems). That is less a price difference than a unit difference: companies announce campuses with buildings and power, governments announce machines, and the stated figures cover whatever was announced.
- Defence
- National civil
- EU & multinational
- Regional
- Own funds
- Philanthropy
- Mixed
- Private
- Unclassified
Table view: Funder type by period of first operation
| Period | 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 |
3. Who can use it
For an outside researcher or company, there are two ways onto a machine: apply through an open allocation call, or pay to rent it. Among public-sector systems in service that record how time is given out, 46% allocate it through open calls, 37% serve their own users and collaborators, and 52% in all can be applied for or rented. Among company-run systems, 61% can be rented, 6% take applications, and 67% in all are reachable from outside.
So the company side is more open by this measure, but it is open to whoever pays, while the public side is open to whoever wins an allocation. Weighted by measured performance, 81% of company-run and 61% of public measured compute with an access record can be applied for or rented, the same ordering. The company-run Rmax figure rests on only 41 measured machines.
These weighted shares are fragile. Across all systems in service, the share of measured compute an outsider could reach is 65%; removing El Capitan alone would raise it to 82%, and removing Aurora would lower it to 61%. A single machine moves the answer by more than most trends do.
- Apply or rent
- Other routes
Table view: Share of compute in service an outsider could apply for or rent
| Group | Systems | Open allocation | Rentable | Apply or rent |
|---|---|---|---|---|
| Public, by systems | 337 | 46.3% | 5.9% | 52.2% |
| Public, by Rmax | 181 | 59.9% | 1.1% | 61.1% |
| Company-run, by systems | 102 | 5.9% | 60.8% | 66.7% |
| Company-run, by Rmax | 41 | 10.6% | 70.2% | 80.8% |
| Access class, systems in service | Public, systems | Company-run, systems | Public, Rmax | Company-run, Rmax |
|---|---|---|---|---|
| Open allocation | 156 (46%) | 6 (6%) | 60% | 11% |
| Rentable | 20 (6%) | 62 (61%) | 1% | 70% |
| Single-tenant lease | 0 (0%) | 9 (9%) | 0% | 0% |
| Own users | 124 (37%) | 3 (3%) | 8% | 0% |
| Restricted | 29 (9%) | 14 (14%) | 29% | 15% |
| Other | 8 (2%) | 8 (8%) | 1% | 4% |
4. What performance costs and what the machines are built for
57 systems have both a stated cost and a measured Rmax. Restated in 2025 US dollars, a GFlop/s of measured performance cost a median of $3.8 on the 14 machines first run from 2020. The price fell tenfold every 3.7 years before 2010 and every 10.3 years since: still falling, but at a fraction of the old pace. Since 2020, the 4 AI training machines with both figures come to a median of $3.1 per GFlop/s and the 6 scientific machines to $3.2; samples this small cannot separate the two.
That comparison only reaches the machines that publish an Rmax. Of the 129 systems with a stated cost, 69 have no usable Rmax, and 55 of those are AI or mixed systems. AI campuses are priced in power instead: across 41 AI systems with a stated investment and a power figure, the median is US$17 m per megawatt, and half fall between US$10 m and US$38 m. The megawatts are facility power for 21, IT load for 17 and a stated capacity claim for 3, and the investment figures range from hardware alone to whole campuses, so treat this as an order of magnitude.
The machines are also built for different work. AI training systems are designed around accelerators doing low-precision arithmetic, and the question is whether that shows in how much memory and storage is bought with each unit of compute. The data bears that out: AI training systems first run since 2020 carry a median 0.52 PB of parallel storage per PFlop/s (9 systems), against 2.5 for scientific machines (40 systems). Memory points the same way, 5.6 GB per TFlop/s against 52, though only 3 AI training systems record a memory total.
The precision caveat. Rmax is a measured FP64 result. AI capacity is usually quoted as a peak in FP8, FP4 or BF16. On the 33 systems in this dataset that have a measured Rmax and an accelerator inventory with a recorded FP8 peak for every part (15 of them AI machines), the vendor's dense FP8 peak is a median 47 times the measured Rmax, from 14 to 140 times depending on the accelerator generation. An AI exaflop and an HPL exaflop are not the same unit, and a comparison of the two sides in one unit has to say which.
5. What the data cannot see
The public side is better documented on 6 of the 7 measures below, and the gaps are not random: they sit where the company build-out is. The exception is pipeline projects stating a power capacity, because company campuses are announced by their power figure.
| Share of systems that have… | Public sector | Company-run |
|---|---|---|
| In service and reporting a measured Rmax | 82% (618 of 753) | 62% (168 of 272) |
| In service with a recorded power draw | 48% (360 of 753) | 25% (69 of 272) |
| AI systems reporting a measured Rmax | 45% (29 of 64) | 22% (57 of 263) |
| With a funder or funding-programme record | 36% (398 of 1121) | 18% (83 of 460) |
| With an access-route record | 40% (453 of 1121) | 26% (120 of 460) |
| Pipeline projects stating a power capacity | 14% (5 of 36) | 83% (102 of 123) |
| Rated verified (two independent sources) | 20% (223 of 1121) | 16% (72 of 460) |
Unranked compute. 602 systems in the dataset appear on no public ranking. 120 of them disclose an accelerator count for a part that some measured system also uses, which allows a bracket: 17.93 EFlop/s to 73.72 EFlop/s of HPL-equivalent compute, or 96% to 394% of the 18.70 EFlop/s held by every ranked system here. Another 423 unranked systems disclose no accelerator count, and 59 disclose one only for parts that no measured system uses; they are not in the estimate at all. Of the 120 clusters in the estimate, 35 are company-run, and they carry 74% to 78% of the estimated compute: the compute the rankings miss is mostly on the company side, in fewer and larger machines.
A few machines decide the totals. The 5 largest measured systems in service hold 41% of all measured performance in service, and all of them are public. The largest, LineShine, holds 12% on its own. Without the single largest public machine (LineShine), the public share of installed Rmax would be 68% rather than 72%. Any Rmax-weighted share, here or elsewhere, is mostly a statement about a handful of national machines.
Stated is not built. Pipeline capacity is what operators and developers say they intend to build. Some projects will shrink, slip or stop, and 62 of the 159 planned or under-construction records have not had their status checked for six months or more, or carry no status date. The figures are a ceiling of intent, not a forecast.
Reading the findings together
On the evidence here, the public supercomputing world is still where most measured, documented and openly allocated compute sits, and the company build-out is where most new capacity is being planned. The two are hard to compare directly because they are described in different units: performance targets and measured FP64 on one side, megawatts, accelerator counts and low-precision peaks on the other. That mismatch is itself the most practical finding. A comparison that does not state its unit, its coverage and its largest single machine is not yet a comparison.
Three questions follow, and this dataset can track all of them as it grows: how much of the stated company pipeline reaches operation, and on what timetable; whether rental becomes the main way researchers reach large-scale compute, as open allocation has been; and whether company-run AI machines start publishing measured results, which would let the two sides be put on one scale.
Method and limits
- A curated dataset, not a census. The 1,581 systems were chosen for coverage of the supply chain and the largest builds, not sampled at random. Shares describe this dataset.
- Sides by operator. Public sector and company-run are defined by the operator's sector. Funding is classified separately from free-text funder records by keyword rules; a hand check of 120 systems found 2 misses, since fixed. Access classes come from access-route records by ordered rules; hand checks agreed on 38 of 40 and then 20 of 20.
- In service. A system counts as in service in 2026 if its first-operational date is on or before that year and it was not decommissioned before that year. The access figures use systems whose status is operational. Both follow the analysis pages they come from.
- Rmax is FP64 and partial. Measured performance exists only for systems that ran HPL. Most company-run AI clusters have not, so every Rmax-weighted figure undercounts the company side.
- Mixed cost scopes. Stated costs mix contract prices, grant awards (which may include operations) and estimates; AI investment figures range from hardware to whole campuses. Costs are deflated with the US consumer price index to 2025 dollars.
- Stated capacity. One headline figure per project; scopes are mixed between IT load, facility power and unstated, and they are added together only so that a total exists.
- Concentration. Rmax-weighted shares are dominated by a few machines; the leave-one-out figures in section 3 and the concentration figures in section 5 show by how much.
- How it was made. The rows were researched by AI agents from public sources and checked by scripts; this report's text was drafted with AI assistance around numbers computed by code. No subject-matter expert has reviewed it. Nothing here is investment advice.
Sources and data
Each section is built on a published analysis page, which carries the full method, tables and named systems. The derived data behind each figure is a CSV on that page; the underlying tables, with a source on every row, are at /data. Method, confidence tiers and the source check are on /methodology.
- Sector mix · CSV
- Build pipeline · CSV
- Who pays · CSV
- Who can get time · CSV
- Cost per unit of performance · CSV
- Machine balance · CSV
- The shadow list · CSV
- Power and cooling
Compute Atlas, “AI compute versus public compute”, sample report, data as of 8 November 2026. computeatlas.net/reports/ai-vs-public-compute