AI Sovereignty Will Not Be Funded by Good Intentions

The Missing Layer: trusted use layer bridging Europe's AI capability to public and institutional adoption

I keep coming back to a question that sounds less exciting than models, chips, or regulation, but may matter just as much:

Who pays for the AI ecosystem we say we want?

In the previous article, AI Has a Trust Problem - Europe’s Public Broadcasting Model May Hold a Clue, I explored what public broadcasting can teach us about trusted AI: not as a blueprint, but as a useful analogy for trust, public mandate, accountability, representation, and coordination.

But any conversation about trusted AI eventually reaches a harder point.

Trust infrastructure is not just a principle.

It has to be financed.

Trusted AI Has a Balance Sheet

It is easy to agree that AI should be trustworthy, sovereign, accountable, secure, representative, and usable inside critical workflows.

It is much harder to build the conditions that make this possible.

Take a hospital that wants to use an AI system to support clinical documentation.

The question is not only whether the model is good. The hospital also needs clarity on patient data boundaries, model evaluation, clinical validation, procurement rules, liability, audit trails, staff training, fallback procedures, and ongoing monitoring when the system changes.

Or take a city that wants to use AI for citizen services.

It may need multilingual support, accessibility, data protection, explainability, procurement confidence, public accountability, and a way to show residents that automated systems are not silently changing how public services are delivered.

None of this happens automatically because a good model exists.

Trusted AI is not only a technology challenge.

It is an investment challenge.

Europe Is Already Moving

Europe is not starting from zero.

The European Commission has started to frame AI as an infrastructure and competitiveness question, not only as a regulatory one. InvestAI, AI Factories, AI Gigafactories, EuroHPC, Testing and Experimentation Facilities, and the Commission’s broader European approach to artificial intelligence all point in the same direction: trusted AI is no longer only a policy aspiration. It is becoming an infrastructure agenda.

The headline number is large. InvestAI is intended to mobilise EUR 200 billion for artificial intelligence, including a EUR 20 billion fund for AI Gigafactories (European Commission).

That sounds like a decisive answer.

But it is not quite that simple.

First, the EUR 200 billion is a mobilisation ambition, not a single pot of committed public money. The actual architecture depends on EU funding, Member State participation, private capital, EIB-backed instruments, procurement demand, and the willingness of companies and institutions to co-invest.

Second, Europe is not making this bet in isolation. The United States is mobilising AI infrastructure at a different scale and through a different logic. OpenAI describes Stargate as a plan to invest USD 500 billion over four years in U.S.-based AI infrastructure, beginning with USD 100 billion immediately (OpenAI). The broader private investment gap is even sharper: global corporate AI investment reached USD 581.7 billion in 2025, private AI investment reached USD 344.7 billion, and the United States accounted for USD 285.9 billion of private AI investment. China, by comparison, recorded USD 12.4 billion in tracked private AI investment (Stanford 2026 AI Index).

That China figure should be read carefully. Stanford explicitly notes that private investment data likely understates China’s total AI capital because China uses government guidance funds and state-initiated investment vehicles. Reporting by Bloomberg, republished by The Business Times, suggests China is preparing a roughly 2 trillion yuan data-centre build-out over five years, led by key government agencies and financed through sovereign debt, strategic state funds, bank loans, and private capital (The Business Times / Bloomberg).

These figures are not directly comparable, but putting them side by side makes the capital gap much harder to miss.

AI capital is not flowing evenly, and Europe’s gap is clearest in private investment.

Region Tracked private AI investment, 2025 Tracked private generative AI investment, 2025 Strategic and public-private mobilisation signals What this means
United States USD 285.9B USD 163.6B Stargate has been announced as a USD 100B-500B AI infrastructure project through 2029. The United States is not only announcing ambition. Private capital is already moving at frontier scale.
Europe USD 20.9B USD 3.2B InvestAI aims to mobilise EUR 200B, including EUR 20B for AI Gigafactories. Europe’s headline number is large, but its actual private AI capital base is far smaller. The question is execution, not ambition.
China USD 12.4B USD 1.5B State-guided AI capital includes government guidance funds, state-backed investment vehicles, local government activity, state-owned operators, private platforms, and reported data-centre build-out plans. Conventional private investment data understates China because capital is channelled through state direction and industrial policy.

The most uncomfortable comparison is not simply Europe versus America versus China. It is three different capital systems. In the United States, AI is being pulled forward by a private-capital market that deployed almost USD 286 billion into AI in 2025 alone, including more than USD 163 billion into generative AI. In Europe, the comparable tracked private AI investment figure was around USD 21 billion, with only USD 3.2 billion going into generative AI. These private-investment figures come from the Stanford 2026 AI Index. Europe’s EUR 200 billion InvestAI headline matters, but it is a mobilisation target, not the same thing as deployed private capital (European Commission). China looks smaller in conventional private-investment datasets, but that picture is incomplete: state guidance funds, industrial-policy vehicles, local-government initiatives and infrastructure mandates channel capital in ways private-market databases do not fully capture.

So the comparison matters because it makes one thing clear:

trusted AI cannot be financed like a side project.

And there is another European complication: in the EU, announced money is not the same as absorbed money, and absorbed money is not always the same as money spent well. The European Court of Auditors has warned about delayed absorption of Recovery and Resilience Facility funds, and Coface estimated in January 2026 that only 58% of RRF funds had been disbursed, leaving roughly EUR 270 billion still to be disbursed by the end of 2026. Coface also notes that an even smaller share had actually been spent (Coface).

That matters for AI because InvestAI’s EUR 200 billion headline is a mobilisation ambition. It is not proof that trusted AI capacity will appear in hospitals, agencies, schools, SMEs, or regulated companies. Europe does not only need announcements. It needs absorption capacity, procurement capacity, implementation capacity, and accountability for whether funds become usable infrastructure.

It needs several kinds of funding at once.

Layer Example Approximate scale Funding model What it proves What it does not solve
Headline AI ambition InvestAI / AI Continent EUR 200B mobilisation target, including EUR 20B for AI Gigafactories EU-backed public-private mobilisation Europe recognises AI as strategic infrastructure Mobilisation is not the same as broad deployment capacity
Compute infrastructure AI Factories / EuroHPC At least EUR 8.2B EuroHPC budget for 2021-2027; 19 AI Factories and 13 antennas operational EU + Member State co-funding Europe is funding sovereign compute access Compute access alone does not create adoptable trusted AI systems
Large-scale model infrastructure AI Gigafactories EUR 20B fund for up to 5 AI Gigafactories Public-private investment vehicle Europe understands frontier AI as strategic infrastructure Gigafactories do not decide who gets trusted access in hospitals, cities, schools, or SMEs
Sector testing infrastructure Testing and Experimentation Facilities, including TEF-Health Over EUR 220M across four TEFs; project budgets of EUR 40-60M over five years Digital Europe + national and partner co-funding Trust infrastructure can be operationalised through real-world testing environments Testing validates solutions, but does not finance broad deployment
Model and language infrastructure OpenEuroLLM and related language-data initiatives Project-based EU and research funding Horizon Europe / research consortium model Multilingual and European-context AI can be publicly supported Project funding does not equal permanent institutional capacity
Private frontier AI capacity Mistral AI, Aleph Alpha and other European providers Significant private and strategic capital Venture capital, strategic investors, national support ecosystems Europe has credible private AI actors Private capital will not naturally fund all public-interest and high-trust use cases
Deployment and adoption layer Public sector, SMEs, regulated sectors Not clearly funded at comparable scale Fragmented procurement, pilots, grants, internal budgets This is where trusted AI becomes useful This is the gap: adoption capacity, procurement pathways, implementation support, neutrality safeguards, and trust governance are not financed like infrastructure

This is the missing distinction.

Europe is beginning to finance parts of the AI stack: compute access, supercomputing infrastructure, testing environments, research consortia, and private AI companies.

But trusted AI also depends on the less visible layer between infrastructure and adoption: procurement support, sector validation, implementation capacity, auditability, public-sector deployment pathways, and the institutional ability to maintain trust over time.

That middle layer is not just operational support.

It is where neutrality, public interest, source quality, and independence from commercial manipulation have to be designed into the system.

If that layer remains underfunded, Europe may end up with impressive AI infrastructure and too few organisations able to use it confidently.

Public Value Alone Will Not Build the Stack

This is where the public broadcasting analogy becomes useful again, but also reaches its limit.

Public broadcasting matters because it shows that some information systems are not meant to maximise attention or profit alone. They are expected to serve a public function: inform, represent, educate, include, and provide access under standards that are not set only by advertisers or commercial platforms.

In Germany, for example, public broadcasting is not one single monolith. Different broadcasters serve different regions, languages, audiences, and public purposes. The point is not that one institution says what everyone should think. The point is that there are structures, mandates, professional standards, funding models, editorial rules, and accountability mechanisms that make public-interest information possible.

That funding model matters. The German Rundfunkbeitrag is a legally prescribed contribution paid by citizens, companies, institutions, and public-interest bodies. It is paid directly to the public broadcasting service rather than to the revenue office, with the stated purpose of supporting public broadcasting’s independence from government influence. In plain terms: the system is designed so that public-interest information does not have to follow the same commercial agenda as advertising-funded media or platform businesses.

This does not make public broadcasting perfect. It does show that neutrality and public value are not only editorial questions. They are financing and governance questions too.

AI will need an equivalent debate.

Not because AI models should become broadcasters.

But because AI systems increasingly mediate knowledge, advice, decisions, search, writing, analysis, and public-service workflows. If these systems answer questions, summarise sources, recommend actions, or assist government and regulated work, then neutrality cannot be treated as a decorative value at the end.

Someone has to decide what neutrality means in a given context.

Someone has to decide which sources are accredited, which quality standards apply, how conflicts of interest are declared, where commercial incentives are excluded, when human review is required, and how manipulation is detected.

And those decisions cannot be left only to vendors.

They need human-led control institutions as well as technical assurance layers.

A national public broadcaster may help explain why public value matters.

It does not explain by itself how to finance compute, cloud capacity, chips, foundation models, evaluation infrastructure, secure deployment layers, sector-specific tooling, procurement support, and continuous monitoring at a scale that can compete globally.

This is why the European Broadcasting Union is a useful analogy for the next layer of the conversation. The European Broadcasting Union is an alliance of public service media organisations. It does not replace national broadcasters. It helps them cooperate across borders while they keep their own mandates, audiences, languages, and institutional identities.

Its funding logic is different from a national licence-fee model, but the principle is still useful. The EBU is not financed as one central European broadcaster. Its shared services are funded through membership and mandatory fees, so the coordination layer is paid for by the institutions that use and govern it. That is the important analogy for AI: a shared trust layer does not have to erase national or institutional independence. It can be funded as common infrastructure for members who could not build the same capacity alone.

For AI, the lesson is not that Europe needs one central institution to do everything.

The lesson may be that Europe needs stronger coordination layers that allow national, public, private, and institutional actors to build together at a scale none of them could reach alone.

Think of the hospital again.

A hospital should not have to become an AI audit lab, a procurement authority, a legal research institute, and a model evaluation centre before it can use AI safely. It should be able to rely on shared validation methods, approved procurement pathways, monitored deployment patterns, and sector-specific trust standards.

Or take a public agency employee who can use AI to prototype an Excel macro with dummy data, but cannot use the AI tool on the actual data where the work matters. The workaround proves the demand. The person knows AI could help. The organisation knows the workflow exists. But because there is no trusted environment, no approved data boundary, no procurement path, and no accountable deployment model, the useful work stays manual.

That is not a model problem.

It is an adoption infrastructure problem.

But coordination is not enough.

If Europe wants trusted AI to become more than a set of principles, it needs to finance the layers that make coordination operational: shared compute, shared testing environments, shared procurement pathways, shared standards, shared source-quality rules, shared monitoring, and shared deployment capacity.

What the Money Actually Buys

This is where the debate often becomes too abstract.

When people say Europe needs trusted AI infrastructure, what does that mean in practice?

It may mean funding a testing environment where a healthcare AI system can be evaluated before it touches real patients.

It may mean helping a city procure an AI assistant for citizen services without forcing every local administration to become its own AI audit lab.

It may mean giving SMEs access to European AI tools that are not only technically capable, but also documented, explainable, secure, and procurement-ready.

It may mean financing shared validation methods so that every school, hospital, municipality, or mid-sized company does not have to solve the same trust questions alone.

The Testing and Experimentation Facilities already show part of this logic. They are financed to test AI and robotics in real-world environments across sectors such as healthcare, manufacturing, agri-food, and smart cities. TEF-Health, for example, exists because healthcare AI cannot be judged only in a lab or a pitch deck. It needs clinical context, validation, safety evidence, regulatory awareness, and trust from professionals and patients.

But TEFs also show the gap.

Testing is necessary.

It is not the same as deployment.

A tested solution still needs procurement approval, implementation support, staff training, monitoring, governance, liability clarity, and a funding path for institutions that cannot afford a bespoke AI transformation team.

And even before deployment, someone has to define the trust criteria.

Who decides what counts as neutral enough for a public-service AI assistant?

Who decides which sources are acceptable for a model used in education?

Who decides whether a chatbot serving citizens should optimise for speed, legal caution, accessibility, linguistic inclusion, or institutional accountability?

Who checks whether commercial interests quietly shape the answer?

Technical assurance can help: source tracing, audit logs, benchmark tests, retrieval constraints, model cards, access controls, red-teaming, and monitoring.

But technical assurance alone is not enough.

Trusted AI needs human-led institutions that set the rules, review the evidence, challenge vendors, and update standards when models, data, and use cases change.

Europe can validate trusted AI systems and still fail to finance the pathways that let organisations actually use them.

The Financing Model Will Shape the Ecosystem

The way trusted AI is financed will influence what gets built.

If funding flows mainly into infrastructure, Europe may build impressive capacity that many organisations still cannot adopt.

If funding flows mainly into frontier model development, Europe may create powerful systems without the practical pathways needed for sensitive sectors.

If funding stays fragmented across countries and institutions, Europe may end up with many pilots and too few systems that scale.

If private capital dominates the agenda, the market may optimise for the most profitable use cases and leave behind the sectors where trust matters most but returns are slower: public administration, education, healthcare, local language contexts, civic infrastructure, and regulated industries.

That does not mean trusted AI should be funded only by public money.

It means the financing architecture matters.

Public money can de-risk early infrastructure, set standards, create demand, and protect public-interest use cases.

Private capital can bring speed, technical talent, operational discipline, and scale.

Institutional procurement can become an anchor market.

Shared infrastructure can reduce duplication.

Strategic policy can make sure investment does not only create isolated companies, but an ecosystem that is usable, resilient, neutral where neutrality is required, and trusted where trust is the condition for adoption.

The Missing Middle Layer

The most important gap may not be at the level of principles.

Europe has many of those.

It may not even be only at the level of infrastructure.

Europe is starting to fund that.

The gap is the middle layer: the financed, governed, operational layer that turns AI capability into trusted adoption.

This layer would not be one product or one institution.

It would be a shared operating system for trust.

Imagine a public agency that wants to use AI on sensitive internal processes. Today, the employee may be able to generate an Excel macro with dummy data, but not use AI with the real data where the time savings would actually matter. The result is a strange compromise: AI is present enough to reveal what could be possible, but not trusted enough to be useful.

A functioning middle layer would change that.

It could provide approved secure environments where sensitive data never leaves the permitted boundary.

It could provide pre-validated tool categories for different data classes.

It could provide procurement templates, risk tiers, logging requirements, human-review rules, and model-evaluation evidence that a public agency can use without inventing everything from scratch.

It could define when a local model is enough, when a European cloud deployment is required, when a general-purpose model is inappropriate, and when a human decision-maker must remain in the loop.

It could also define who is allowed to set these rules.

That is the early-onset governance question: not only whether an AI system behaves responsibly after launch, but who shapes the criteria before the system is bought, built, trained, connected, or deployed.

There are several possible financing models.

One model is a taxpayer-funded public-interest layer, analogous to national public broadcasting: if society needs neutral, accountable AI infrastructure for public services, education, healthcare, and civic access, then it may need stable public funding that is insulated from short-term political cycles.

Another model is a membership structure, closer to the EBU: public agencies, broadcasters, hospitals, universities, cities, and approved private actors could contribute to shared standards, shared evaluation capacity, shared procurement frameworks, and shared technical tooling.

A third model is an anchor-customer model: governments, large public institutions, and regulated industries commit procurement demand early, making it viable for trusted European providers to build the missing deployment layer.

A fourth model is a levy or contribution model: organisations that profit from large-scale AI deployment could help finance independent evaluation, public-interest datasets, source-quality infrastructure, and audit capacity.

The point is not to choose one model immediately.

The point is to recognise that this layer needs a financing concept of its own.

If Europe only funds compute, models, and pilots, the adoption layer remains nobody’s permanent responsibility.

And if nobody owns that layer, trusted AI stays stuck between policy ambition and practical use.

What This Concept Could Look Like

The concept could be called a Trusted AI Commons, a Trusted AI Adoption Layer, or a Public-Interest AI Infrastructure Layer.

Whatever the name, the visual is simple:

At the bottom, Europe is building infrastructure: compute, cloud, data spaces, models, testing facilities.

At the top, organisations are trying to use AI: hospitals, schools, municipalities, SMEs, public agencies, regulated companies, media organisations, and civic institutions.

In between sits the missing layer.

It translates capability into trusted use.

It contains five functions:

  1. Standards: What counts as neutral, source-backed, auditable, inclusive, and safe for a given use case?
  2. Validation: Which models, tools, and workflows have been tested for specific sectors and data types?
  3. Procurement: How can institutions buy or approve AI without each one writing its own trust framework from scratch?
  4. Deployment: What technical boundaries, logging, monitoring, human review, and fallback procedures are required?
  5. Governance: Who updates the rules, resolves conflicts, represents public-interest concerns, and prevents commercial capture?

This is the layer that could become the header image for the article:

A bridge between AI infrastructure and real institutional adoption, held up by five pillars: standards, validation, procurement, deployment, and governance.

That image matters because it turns the abstract funding question into something visible.

We are not only asking who pays for models.

We are asking who pays for the bridge that makes trusted AI usable.

The Question Every AI Ecosystem Should Ask

The Draghi report on European competitiveness argues that Europe needs a new investment trajectory to respond to structural competitiveness challenges.

AI is only one part of that broader picture. But it is a very visible one.

So maybe the question is not only:

Who can build the most powerful AI system?

Maybe it is also:

Who can finance the ecosystem that makes AI trustworthy enough to use?

That question applies to Europe.

But it also applies to any nation, region, or economic bloc that wants AI to serve more than the interests of the largest platforms and fastest-moving commercial actors.

Trusted AI will require models.

But it will also require neutral and accountable governance, accredited source standards, testing, procurement pathways, secure deployment environments, human-led oversight, monitoring, public legitimacy, and financing models that make all of this possible before adoption scales.

Good intentions will not be enough.

The architecture of funding may become one of the most important choices in the architecture of trusted AI.

Sources and Further Reading