What is the public AI movement, and who is behind it?

Public AI is the idea that a society should not have to rent all of its AI from a handful of companies. In its strong form, public bodies or public-minded groups train the models, run them at cost and answer to citizens. In its weaker form, governments fund the shared ingredients, such as computing power and open data, and let others build on them.

It differs from sovereign AI, which asks whether a country controls the AI it depends on. A sovereign model can be private and closed, so long as it is domestic. Public AI asks a different question: who does the AI answer to? Our plain-language comparison sets the two side by side.

Who is making the case

The idea has been argued in print since soon after ChatGPT's release. Bruce Schneier, Nathan Sanders and Henry Farrell say they were among the first, proposing what they called an AI public option in 2023. Since then, a recognizable set of groups has formed around it:

The Bertelsmann paper sets a concrete goal: keep at least one complete open source model in existence that comes close to what the best proprietary systems can do. Without that, it argues, public actors stay limited in what they can do. Its "gradient" is a useful corrective, since it treats publicness as a matter of degree, based on openness, governance and focus on the common good, instead of a yes-or-no label.

What has actually been built

The most visible piece of infrastructure is the Public AI Inference Utility. Inference means running a trained model to answer requests, which is a different job from training it. The Utility's argument is practical: national supercomputers can train a model, but they are not set up to serve the public around the clock. So it offers a chat site and a programming interface in front of public models, running on compute donated by partners.

Its backing is itself a lesson in how these projects survive. According to Schneier and Sanders, it sits under the nonprofit Metagov, with money from Mozilla, the Center for Cultural Innovation and the Future of Life Institute. Donors of compute include the Swiss supercomputing centre, national initiatives in Australia, Germany and Singapore, and commercial clouds such as AWS. The Utility says free access rests on donated GPU time and advertising subsidies, and it hopes states and institutions will anchor it over the long term.

The models tell a similar story of compromise. Schneier and Sanders note that SEA-LION, the Singapore family, began with models trained from scratch and then moved toward fine-tuning commercial open-weight models, which they read as the hard trade-off between cost and usefulness. Ai2's OLMo series is highly transparent but is not under public control, and it relies on corporate partners for compute. Our comparison of these models goes through them one at a time.

A worked example: how public is it?

Here is one way to apply the "gradient" idea yourself. Take any AI service your town, school or employer is considering and fill in this table. The two columns below are filled in from public documents as an illustration.

The last row is the one advocates have to answer for, and it leads to the objections.

What the critics say

The sharpest objection comes from free-market advocates. In an opinion piece for the Competitive Enterprise Institute, the author argues that government cannot keep pace with leading tech companies, that a public model would be out of date before it was finished, and that the state would be bidding for the same scarce engineers and chips as the richest companies on earth.

The cost numbers give that argument some weight, and also some limits. Schneier and Sanders estimate Apertus's computing at $36 million to $100 million (US) at commercial rental rates, a small sum next to what the leading labs spend. But they also concede that the model is about the size of the 2023 generation of commercial models. Supporters reply that this is the point: useful and affordable, not the most powerful.

A quieter criticism comes from within the field. A 2026 paper on whether governments should buy or build, whose authors include one of the Apertus leads, describes national strategies as typically mixed, with commercial, open-source and sovereign models each serving different purposes. On that view, a single public model is rarely the whole answer, and the real decision is which uses justify the extra control.

There are also unanswered questions about how such a service stays independent. A utility funded by advertising, donated commercial compute and foundation grants is not yet the stable public institution its name suggests. Reliable figures on how many people actually use these services are also hard to come by.

Canada is an interesting case. Its recent moves, including the stake in Cohere, lean toward sovereign capacity through a private company more than toward a public model. The Utility has named Canada as a place it hoped to add launch partnerships, though no such partnership appears in its announcements so far.

Could public AI ever replace commercial assistants?

Not on current evidence, and most advocates do not claim it should. The Bertelsmann paper says outright that the aim is to rebalance power, and that it does not reject private innovation. The more realistic picture is coexistence, the way public libraries and bookstores coexist.

Where a public option could matter first is in narrow, high-trust uses: public-service chatbots, schools, languages the market ignores, research that needs to see inside a model. For everyday work that demands the strongest model, commercial assistants will likely remain ahead for as long as the spending gap lasts. If you want to follow the movement, the useful thing to watch is less the next model release than who commits long-term public money to running one.