What Would It Take to Make AI Public Infrastructure?

Public models and public compute are only the beginning. The harder task is building the institutions that make them usable.

The most interesting public AI experiment may not be another government buying access to a private chatbot.

It may be an attempt to build something governments, universities, researchers, businesses, and civic institutions can actually use without renting every important layer from the same small group of companies.

The European Union is now testing that possibility. The European Commission selected the EUROPA consortium, led by the Italian company Domyn, to develop an open-source frontier model covering all 24 official EU languages. The project will receive public support and access to European supercomputing capacity through EuroHPC.

That combination matters.

A publicly supported model is one thing. A public computing system that helps institutions train, adapt, and use it is something more. Europe is trying to assemble both.

The ambition is easy to recognize. Much of frontier AI has developed through private laboratories, private cloud platforms, private application programming interfaces, and private release decisions. Governments participate as regulators, customers, funders, and national-security partners, but the ordinary arrangement remains familiar: private firms build the most capable systems, and everyone else buys access.

EUROPA points toward a different possibility.

Could AI become public infrastructure—something governments, universities, businesses, and civic institutions can use, inspect, adapt, and govern rather than merely consume?

It is a promising question. It is also a harder one than “Is the model open?”

Public funding is not the same as public capacity

Governments can spend public money on AI in several very different ways.

They can buy discounted licenses from a private provider. They can pay consultants to integrate proprietary systems. They can subsidize cloud access. They can fund research. They can build shared computing facilities. They can support open models. They can create internal technical teams, testing laboratories, procurement standards, training programs, and public-interest applications.

All of those may be useful.

They do not produce the same kind of capacity.

California’s recent agreement with Anthropic illustrates the distinction. State agencies, cities, and counties can purchase Claude at a substantial discount, with workforce training, technical assistance, and implementation support from Anthropic.

That may help public employees adopt AI more quickly. It may improve services. Training and technical support are certainly better than dropping a new tool into an agency and hoping enthusiasm will take care of the rest.

Still, the arrangement begins with access to a vendor’s system.

The model, product roadmap, pricing structure, technical architecture, and much of the implementation knowledge remain with the provider. Public employees may become more capable users while the institution itself becomes more dependent on one company’s tools and expertise.

That is not necessarily a bad bargain. It is simply a different bargain from building public infrastructure.

A discounted subscription is not a public utility. A training session is not an internal engineering team. A helpful vendor is still a vendor.

Public capacity begins to emerge when institutions retain knowledge, bargaining power, data control, operational continuity, and realistic alternatives after the initial agreement ends.

What does a public institution actually control?

That question is more useful than asking whether a project carries a public label.

Can the institution inspect the system closely enough to understand how it behaves?

Can it adapt the model or application to its own needs?

Can it move its data, workflows, and accumulated knowledge to another provider?

Can it maintain the system if the original vendor changes direction?

Can it train its own staff rather than relying indefinitely on outside specialists?

Can it compare multiple models and replace one without rebuilding the entire service?

Can the public understand who is responsible when the system fails?

These are not glamorous questions. Infrastructure rarely becomes interesting by announcing its maintenance plan. Yet maintenance, staffing, interoperability, and governance determine whether a public investment produces durable capability or a temporary demonstration.

The same is true of AI.

A model can be publicly funded and still be practically inaccessible. A platform can be inexpensive and still create lock-in. A government can “have AI” while possessing very little authority over the system it now depends upon.

The civic value lies in the surrounding institution.

Open licensing is necessary—and insufficient

EUROPA’s proposed openness is important. So is Portugal’s AMALIA model, developed by Portuguese universities and research institutions with government support and European recovery funding.

AMALIA is intended to strengthen AI capability in European Portuguese, a language variant often poorly served by larger models. The model, training data, and source code are being released openly. Early applications include public services, education, museums, research, business, and decision support.

That is a meaningful dispersion signal.

Language is itself a form of infrastructure. When important digital systems poorly represent a country’s language, law, institutions, and cultural context, that country becomes more dependent on systems built elsewhere for different users and assumptions.

Open models can reduce that dependence. They allow independent evaluation, local adaptation, specialized use, and in some cases self-hosting.

But openness comes in degrees.

A model may have downloadable weights while withholding parts of the training data or construction process. Its license may restrict commercial or institutional use. Running it may require hardware available only to large organizations. Updates may depend on the original developer. Security fixes may be irregular. Documentation may be thin. Smaller institutions may discover that the model is open in roughly the same sense that a grand public building is open when the doors are unlocked but no road leads to it.

A download link is not an implementation strategy.

For an open model to function as civic infrastructure, institutions need affordable computing, deployment tools, documentation, evaluation, cybersecurity, maintenance, technical assistance, and people who understand both the technology and the public problem being addressed.

That is why the surrounding ecosystem matters as much as the license.

Public compute changes the equation

Europe’s broader AI Factories program is potentially more important than any one model.

EuroHPC now offers computing resources and technical support to startups, small and medium-sized businesses, researchers, public-sector users, and other European institutions. Some access pathways are free for qualifying smaller firms and include assistance from the host AI Factory, not merely time on a machine.

That begins to address one of the central weaknesses of open AI.

Open models may separate users from the original model provider. They do not automatically separate them from concentrated compute, cloud hosting, specialized engineering, or implementation expertise.

Public computing infrastructure can create another route.

A university may be able to train or adapt a model without negotiating with a hyperscale cloud provider. A startup may test a product without surrendering much of its early capital to compute costs. A public agency may evaluate several models before committing to one. Researchers may study systems that would otherwise remain inaccessible.

That is real capacity.

Even so, a supercomputer is not a civic institution. High-performance computing facilities can be technically impressive and administratively forbidding. Access rules, application procedures, scheduling, software environments, staffing, and technical support will determine who can use them in practice.

A public resource that only the already sophisticated can navigate may widen the field slightly while leaving most smaller institutions exactly where they started.

The test is use.

Do municipalities, schools, hospitals, nonprofits, regional businesses, universities, and public-interest researchers gain practical access? Can they move from experiment to operation? Does knowledge spread beyond a small circle of elite technical institutions?

Public compute matters most when it becomes usable compute.

Sovereignty is not automatically civic capacity

Ukraine provides a more urgent version of the same problem.

Ukraine’s Ministry of Digital Transformation and Kyivstar are developing a national model based on Google’s open-weight Gemma family. The project was announced in late 2025, and Ukrainian officials have since emphasized the need for systems that can eventually operate without continuing provider control. The model is intended for government services, private enterprises, and military use, with a planned release in autumn 2026. (Reuters)

The security rationale is obvious. A country at war has strong reasons not to place essential government and military functions entirely inside remote systems controlled by foreign providers.

Local operation can improve resilience. It can protect sensitive data. It can create expertise that remains in the country. It may allow institutions to continue functioning when commercial access, connectivity, or geopolitical relationships change.

But “sovereign AI” can describe several different arrangements.

It may mean a nationally operated public capability. It may mean a telecom company running a system under government partnership. It may mean a domestic platform protected from foreign competition. It may mean military control. It may mean little more than national branding around technology whose crucial components still come from elsewhere.

Sovereignty answers the question, “Which country controls the system?”

Civic infrastructure asks a second question: “Who inside the country can use it, shape it, question it, and benefit from it?”

Those answers can diverge.

A national model may reduce dependence on a foreign company while concentrating authority inside the state or a favored domestic provider. That may still be justified in wartime. It should not be confused with broad public capacity.

Frontier AI may not be the most useful public AI

There is another possibility worth taking seriously.

Perhaps civic AI infrastructure will not be built mainly around frontier models.

The phrase “public AI” can summon an image of a government-funded competitor to OpenAI, Anthropic, Google, or China’s leading laboratories. That may be part of the story. EUROPA is explicitly testing whether Europe can build a public-supported frontier alternative.

But most public institutions do not need a frontier model for most tasks.

A city may need a reliable system that helps staff compare zoning provisions, summarize public comments, search ordinances, or translate service information.

A state legislature may need tools for bill comparison, legal research, fiscal analysis, and public access to legislative history.

A school district may need a shared system for evaluating vendors, protecting student data, and helping teachers adapt materials.

A nonprofit may need benefits navigation, grant research, case management, or multilingual communication.

These systems do not need to win a benchmark competition. They need to be accurate enough, secure enough, affordable enough, understandable enough, and governed well enough to earn a place in public work.

Smaller tools may be easier to audit. They may run on less expensive hardware. They can be designed around a defined institutional purpose. They may allow the people who understand the public problem to participate in development rather than receiving a general-purpose system after the important choices have already been made.

There is a certain irony in building a model capable of explaining nearly everything and then discovering that a county still needs help finding the correct version of its procurement manual.

Civic infrastructure is often unheroic. That is part of its value.

Maryland shows what the institutional layer can look like

Maryland offers a useful example of a different approach.

The state maintains a public inventory of AI systems and uses, tying that inventory to existing technology-governance processes. It has also launched an AI Innovation Lab intended to give agencies a controlled environment for testing and prototyping systems before moving them into ordinary operations.

Neither step creates an independent frontier model.

Together, however, they begin to build institutional capacity.

The inventory makes government use more visible. The lab gives agencies a place to experiment without immediately committing a system to public service. Shared tools, standards, staff, and lessons can reduce the need for every department to begin from zero—or to accept the first vendor proposal that arrives with a polished demonstration.

This may be closer to what civic infrastructure looks like in practice.

Not one giant public model.

A network of shared capabilities:

• public compute;
• open or replaceable models;
• testing environments;
• technical assistance;
• procurement expertise;
• inventories and disclosure;
• evaluation capacity;
• trained public employees;
• common security and data standards;
• institutions that preserve learning across projects.

The pieces reinforce one another. Remove enough of them, and the institution returns to dependence.

Lane 6: building the machinery to act

This is the concern at the center of Lane 6 — Civic Infrastructure and Public Capacity in The Race. Civic infrastructure is the part of public infrastructure that gives institutions and communities the practical capacity to understand, use, govern, and contest AI.

Governance is not only the power to regulate someone else’s system.

It is also the capacity to understand a technology, define public problems, build alternatives, negotiate with providers, operate useful tools, and challenge systems that do not serve the public well.

That capacity can exist in governments, universities, libraries, schools, public media, nonprofit organizations, professional associations, civic-technology groups, and community institutions.

It will not look the same everywhere.

A national government may need secure compute and a locally operated model. A city may need shared procurement and technical support. A university may need research access. A small nonprofit may need a trusted intermediary that can evaluate products and provide low-cost tools. A community may need enough knowledge and institutional support to decide whether an AI system should be used at all.

The common thread is retained capability.

People and institutions should emerge from adoption with more knowledge and more room to act—not simply with another subscription.

What would count as success?

EUROPA, AMALIA, the European AI Factories, Ukraine’s national model, California’s vendor partnership, and Maryland’s institutional approach are not versions of the same project.

That is precisely why they are useful.

They show several routes through which public institutions are trying to gain AI capability:

• build a public-supported open frontier model;
• create a national language model;
• provide shared compute and technical support;
• operate a sovereign system for resilience;
• accelerate adoption through a private vendor;
• build inventories, sandboxes, and reusable government expertise.

Each route creates different strengths and dependencies.

The evidence is still early. Announcements are plentiful. Durable public capacity is harder to demonstrate.

A convincing civic-infrastructure model would show several things over time.

Smaller institutions would actually use it. Staff would gain transferable skills. Models and vendors could be replaced without losing the service. Public data and accumulated institutional knowledge would remain under public control. Independent evaluation would be possible. The system would have maintenance, funding, security, and governance arrangements that survive the original launch.

Most important, public institutions would retain the ability to define the problems they are trying to solve.

That last point is easy to lose.

A powerful general-purpose system arrives with its own categories, workflows, measures of success, and convenient use cases. Institutions under pressure may reorganize around what the tool can readily do. Civic infrastructure should work the other way around: the public purpose comes first, and the technology is selected or built to serve it.

A public option for capacity

Private AI systems will remain important. Governments and civic institutions will continue to buy commercial products, and often should. Building every model and application internally would be expensive, slow, and occasionally a fine way to recreate software that already exists.

The alternative to dependence is not self-sufficiency.

It is capacity.

A capable institution can buy a private service without surrendering judgment. It can use an open model without mistaking openness for usability. It can share infrastructure where scale is useful and preserve local authority where context matters. It can change providers. It can learn.

EUROPA is worth watching because it combines public compute with an open model and a stated intention to make advanced capability available beyond a small group of frontier firms.

Whether it becomes civic infrastructure will depend on what surrounds the model.

Who can use it?

Who can maintain it?

Who can adapt it?

Who learns from the work?

Who remains in control when the first grant, contract, or burst of political attention ends?

Those questions are less dramatic than a frontier-model launch. They are also where public capacity is built.

AI becomes civic infrastructure when institutions can do more than gain access to it.

They must be able to make it their own.

References and Further Reading

• European Commission, “Commission Selects EUROPA Consortium as the Winner of the Frontier AI Grand Challenge”.
• EuroHPC Joint Undertaking, “AI Factories”.
• EuroHPC Joint Undertaking, “AI Factories Access Modes”.
• Afonso Simplício et al., “AMALIA Technical Report: A Fully Open Source Large Language Model for European Portuguese”.
• Reuters, “Portugal Launches First Open-Source AI Model, Joining Europe’s Sovereignty Push”.
• VEON, “Kyivstar and Ukrainian Ministry of Digital Transformation Select Google Gemma as the Foundation for Ukraine’s National LLM”.
• Reuters, “Ukraine to Pick AI Models Operated Without Provider Control, Official Says”.
• Office of Governor Gavin Newsom, “Governor Newsom Announces a First-of-Its-Kind Partnership Providing Anthropic Tools to State Agencies”.
• Maryland AI, “AI Inventory”.
• Maryland AI, “News and Updates”.