Build AI as public infrastructure

Provision AI — compute, foundation models, training data, and deployment surfaces — as public infrastructure the way societies provision electricity, water, roads, broadcasting, libraries, and municipal broadband: publicly funded, democratically governed, accessible without corporate credentials or subscription, and structured so the governance layer is inseparable from the technical layer. The move is not to regulate private AI harder; it is to build an alternative provisioning stack the public owns, so the terms on which AI is available are not set by the firms that currently own it.

An actor chooses this strategy when the harm surface of AI is judged to be structural in its ownership pattern rather than fixable through regulation of private incumbents. Regulating a private stack constrains behaviour; owning a public stack constitutes an alternative. The strategy solves the leverage problem at its source: publics that co-own AI infrastructure have standing to shape it that publics regulating someone else's infrastructure do not have, and civic institutions (libraries, universities, public broadcasters, municipalities) that operate the infrastructure have a legibility with their constituencies that private platforms do not.

It trades off scale and speed for legitimacy and durability. A public stack cannot match frontier-lab compute budgets in its first decade; state-scale provisioning is politically contingent on legislative appetite; and the strategy carries the risk that "public infrastructure" degenerates into "national-security AI" in the sovereign-AI register that shares its vocabulary. The strategy is also structurally slow — public infrastructure takes decades to mature (municipal broadband from the 1990s to 2020s; public broadcasting from the 1920s to now) — and the strategy's legitimacy rests on committing to that timescale rather than trying to compete with the four-year training cycles of the private frontier.

Verdict — right strategic register for the ownership question, exposed at the scale question

The strategy earns its keep at the framing layer and at the civic-institutional integration layer. Framing AI as a public-utility question — the Public AI Network's central conceptual contribution — is doing structural work the corpus's other strategies do not: it reframes AI governance from "regulate the private firms" to "provision the public alternative," and it makes the ownership question visible as a policy choice rather than a fixed condition. That framing move is what has made "public AI" a legible policy option in AI-governance debate internationally, with more than twenty-five publications by 2025 and citation across academic and policy literature. At the civic-institutional layer, the six-library pilot program running in Utah, New Jersey, Georgia, Texas, and Massachusetts is the closest existing analogue to the municipal-broadband precedent the strategy draws on: a distributed multi-jurisdictional deployment where a civic institution (the public library) is the operator of the AI-provisioning service, giving the strategy replicable state-level infrastructure rather than a single flagship demonstration.

The strategy is exposed at the compute-scale layer and at the political-viability layer. The Public AI Inference Utility and the Swiss Public Inference Utility serve inference at real state-and-coalition scale, but neither trains frontier foundation models at the parameter counts the commercial labs deploy. Whether that gap closes — through EU consortium build-out following the "Airbus for AI" proposal, through Swiss and other state-scale public-compute investment, or through a shift in what "frontier" is even taken to mean — is the strategy's open empirical question. If the gap remains open indefinitely, the strategy functions as a values-preserving supplement to private AI rather than the structural substitute its public-utility framing implies. Political viability is the second exposure: the current civic-infrastructure work is underwritten heavily by philanthropic funding (Rockefeller Foundation, Patrick J. McGovern Foundation, Aspen Digital's Amplify Public AI project), and the strategy's transition from philanthropy-underwritten pilot layer to legislation-underwritten permanent infrastructure has not yet happened at scale in the corpus's signal. Switzerland is the exception; the US library-pilot layer is not yet.

Even with these exposures the strategic case is robust, because no other strategy in the corpus contests the ownership question at the infrastructure layer at all. Every other strategy either regulates the private stack, builds critical knowledge about it, or contests specific deployments — none of them proposes an alternative provisioning apparatus the public owns. The strategy's distinctive contribution to the corpus is that it is the only one whose theory of change is build the alternative rather than constrain the incumbent, and that theory of change is what makes it structurally load-bearing for the movement's long-horizon prospects even where its short-horizon delivery is thin.

Ecology

The strategy is closely related to but categorically distinct from parallel community research institution. Both build separate substrate outside the dominant AI infrastructure, but the parallel-research-institution form (the DAIR Institute, the Ada Lovelace Institute) produces knowledge, analysis, and critique about AI; the public-infrastructure strategy produces the AI itself — compute, models, inference services — that people actually use. Research institutes contest what AI is known to be; public infrastructure contests what AI is provided as. They can share adopters (Metagov, Berkman Klein Center, and Chatham House participate in both registers via PAINT), but the strategic form is different: analysis versus infrastructure.

The strategy overlaps with indigenous and community data sovereignty at the shared premise that data and AI can be community-owned and community-governed rather than privately extracted, but diverges on scale and locus. Data-sovereignty adopters (Te Hiku Media, FNIGC) operate at community and iwi and nation scale, with the community itself as the sovereignty-holder; public-infrastructure adopters operate at civic-institutional and state scale, with a public — a municipality, a state, a nation, an international coalition — as the ownership-holder. Both strategies build substrate the community-or-public governs, but the political-legibility register is different: sovereignty appeals to treaty-based communal rights, public infrastructure appeals to public-utility civic norms. PAINT's AI Commons initiative — reconstructing a public knowledge base for AI training data — is the strategy's clearest edge with data-sovereignty, and SJ Klein's Concordance Project piloting community-led data pools in agriculture sits in the overlap register between them.

The strategy is fed by counter-narrative framing. The "AI as public utility" framing itself — the analogy to publicly-owned water systems, public libraries, public broadcasting, municipal broadband — is a counter-narrative move against the default "AI as private product" framing, and the strategy's cognitive availability as a policy option depends on that framing move landing first. The strategy's principal risk on this register is the frame-laundering problem the counter-narrative-framing strategy's own verdict identifies at length: the "sovereign AI" and "national AI strategy" registers, which national-security policy actors are also using, share vocabulary with the public-infrastructure register while inverting its democratic-governance content. Which framing wins the register is unresolved.

The strategy contains participatory deliberation as policy input as a necessary internal component. Public infrastructure that is not democratically governed collapses into state-managed infrastructure at best and defence-tier sovereign AI at worst; the "publicly-provisioned AI that replicates private AI's governance structures is not genuinely public" test the coalition articulates is the discipline. PAINT's community-benchmarks framework (Aspen Digital's Intelligence in the Public Interest, June 2025) is the strategy's internal deliberative apparatus for keeping the governance layer inseparable from the technical layer — the corpus's most operationalised current case of a strategy binding its infrastructure delivery to a running participatory-deliberation practice.

The strategy is adjacent to coalition lobbying of binding regional regulation but occupies a different lane. Regulation-lobbying strategies (the EU AI Act, US state legislation on AI deployment) constrain the private stack; public-infrastructure strategies build the public alternative. Both are needed simultaneously — regulation of private AI without a public alternative leaves the movement fighting a defensive battle in perpetuity; a public alternative without private-AI regulation lets the private stack keep accreting harms the public alternative cannot yet substitute for — but the strategies operate on separate levers and neither is a substitute for the other. Their coordinated pursuit is the corpus's strongest theory of a two-track AI-governance strategy, and PAINT's participation in EU AI Act debates alongside its "Airbus for AI" proposal is the strategy's clearest single case of the two lanes running together.

The strategy's principal in-corpus adopter is the Public AI Network — an international coalition of more than 350 members from over 100 organisations, founded 2023 at the Internet Archive's DWeb Camp, running the coordination layer for the movement — with associated engagement from Metagov, Aspen Digital, Open Future, Mozilla, Public Knowledge, Berkman Klein Center, and Chatham House operating in and around it. The corpus's coverage of the strategy is at present coalition-heavy and civic-institution-thin: PAINT is the coordinating body, but the individual adopters at the civic-institutional layer — the six state libraries running pilots, the Library of Congress as convening host, the Swiss federal-technology apparatus operating the Swiss Public Inference Utility — are not yet in the corpus as their own entries. That is the strategy's principal gap-loop signal: the state-and-civic-institution adopters that operate public infrastructure at delivery scale are the register the corpus needs to deepen if the strategy's evaluation is to hold beyond the coalition-coordination layer.