Short answer
Public AI compute means supercomputers and shared resources paid for by governments and allocated to researchers, usually through an application rather than a bill. The US program is NAIRR, which is moving from pilot to permanent with a new operations center; the UK, the EU and Canada run or are building their own versions.
Follow-up question How can someone in Canada get access to public compute?
Compute is the computing power AI needs: mostly racks of graphics processors, or GPUs, plus the storage and networking around them. Most university labs cannot afford enough of it to train a serious model, and that leaves much of the frontier work inside a few companies. Public compute programs exist to narrow that gap. A government pays for or brokers access to the machines, and researchers apply for time.
This matters beyond universities. Apertus, the Swiss public model, exists because its team could draw on a national supercomputer. Without something like that, public or academic models at that scale would have no place to be trained.
What NAIRR is
The National Artificial Intelligence Research Resource (NAIRR) is the American program, led by the National Science Foundation (NSF). It started as a pilot in January 2024. The NSF says it has since supported more than 600 projects and 6,000 students in every state, plus Washington, D.C. and Puerto Rico, working with 13 other federal agencies and 28 non-government partners.
It is more of a broker than a single machine. The Department of Energy contributes supercomputer time, NASA and NOAA contribute datasets, and companies contribute access to their own resources. Researchers and educators request what they need through the pilot's site. A separate strand, NAIRR Secure, run jointly by the Department of Energy and the National Institutes of Health, is for sensitive data such as health records.
On September 1, 2026, the program took a step toward permanence. The NSF gave the San Diego Supercomputer Center a five-year, $35 million (US) award to run a NAIRR Operations Center, with the Texas Advanced Computing Center as its partner. Its job is to coordinate the providers and connect users with computing, data, models and training. Nextgov reports that the White House's AI Action Plan backs the program, and it also notes that the NSF has cancelled a number of grants since April 2025, including some touching on AI. Public compute depends on political support, and that support can shift.
A worked example: how far does the money go?
Dollar figures in this field are hard to judge without a yardstick, so here is one. A GPU hour is one processor running for one hour. Schneier and Sanders, writing about Apertus, use a commercial rental rate of about $6 to $10 (US) per hour.
| Item | Amount | Equivalent at $6 to $10 per GPU hour |
|---|---|---|
| Apertus, total compute (per its co-lead, as reported) | About 10 million GPU hours | $60 million to $100 million |
| Apertus, final training run only (same source) | About 6 million GPU hours | $36 million to $60 million |
| NAIRR Operations Center | $35 million over 5 years, or $7 million a year | About 700,000 to 1.2 million GPU hours a year, if all of it bought compute |
The last line is our own arithmetic and it is deliberately unfair, because the center's money pays for coordination, support and a portal, not for chips. The GPUs come from partners. That is the point of the exercise: the award is small next to a single public model's training bill, so NAIRR's reach depends on what agencies and companies contribute. The NSF describes this as bridging a gap, and it relies on those partners by design. How much of that compute is actually available is harder to see from the outside, since contributions come in kind.
How other countries do it
Other countries have made different choices, and they fall on a range from building machines to brokering access.
| Program | Who runs it | Approach | Scale, per official sources |
|---|---|---|---|
| NAIRR (US) | NSF, with SDSC and TACC | Brokers public and private resources | 600+ projects so far |
| AI Research Resource (UK) | UK government, University of Bristol and others | Owns machines | Isambard-AI has 5,448 NVIDIA GH200 chips; goal of at least 20 times more capacity by 2030 |
| AI Factories (EU) | EuroHPC Joint Undertaking | Network of national sites, with support for small firms | 19 factories, 13 antennas |
| Alps (Switzerland) | CSCS | National supercomputer with grants of time | About 11,000 GH200 chips, as reported by Schneier and Sanders |
| Sovereign AI Compute Strategy (Canada) | ISED | Mix of public machines, private projects and subsidies | $2 billion over five years |
A machine you own gives you control, and it ages fast. A brokered model is cheaper and flexible, but it leaves you dependent on the goodwill and pricing of companies that are also competitors for the same research. Nobody has shown yet which approach produces better science, and the comparison is complicated by how differently each country counts its capacity.
Try it: Pick a project you have in mind and estimate its GPU hours. Multiply the number of GPUs by the hours you would run them, then multiply by $6 to $10 (US). Do it again for ten times the size. The result will show whether you need a public allocation or just a modest cloud bill.
What a good application needs
There is no single common form, since each program runs its own process. The practical advice that follows from how these programs describe themselves is simple. Be concrete about the question you want to answer, what data you will use and why your institution's own machines are not enough. Show that you have run a small version first. Check whether the program serves your kind of organization: NAIRR describes the US research and education community, while EU AI Factories emphasize support for small firms and startups, and access rules vary by call.
How can someone in Canada get access to public compute?
It depends on who you are. If you are at a Canadian university, college or research hospital, the place to start is the Digital Research Alliance of Canada, a federally funded not-for-profit that coordinates advanced research computing, with more than 120 member institutions. Your institution's research computing office can tell you how its allocations work.
If you are a business, the federal Canadian Sovereign AI Compute Strategy is the relevant frame. Its $2 billion is divided into three parts: up to $700 million to attract private investment, up to $1 billion for public supercomputing, including a new Canadian-owned machine, and up to $300 million for the AI Compute Access Fund, which helps Canadian firms pay for compute. Check the government's strategy page for what is currently open, as programs under it have opened in stages.
Canada also has an agreement with the UK, dated January 31, 2024, to explore access to compute for researchers and industry. It is a statement of intent, and it does not by itself grant anyone time on a British machine. NAIRR, for its part, describes itself as serving the US research and education community.
Sources
- National Artificial Intelligence Research Resource — U.S. National Science Foundation, October 1, 2026
- Strengthening America's AI Ecosystem with the Launch of the NSF NAIRR Operations Center — San Diego Supercomputer Center, September 1, 2026
- Canadian Sovereign AI Compute Strategy — Innovation, Science and Economic Development Canada, June 4, 2026
From the Guide to Artificial Intelligence, by Robert Scales, published by Holistic Hybrid Inc. Share and adapt it under CC BY-NC-SA 4.0: credit the guide, link to https://guidetoartificialintelligence.com/en/learn/public-compute-nairr-researchers-get-gpus/, no commercial use without permission, and keep the same licence.