Europe's Defence AI Bill Is Bigger Than the Software

Europe's argument about sovereign defence AI is often reduced to which country's software wins the tender. That is the wrong accounting unit. The enduring leverage lies in computing capacity, secure data access, specialised personnel, model updates and the contractual right to walk away from a vendor. The international defence AI centre announced in Rennes on November 12, 2026, puts that question back on the table. France expects a multinational launch conference on November 20 and operational capability in 2027, but the participating governments and the centre's funding are not yet fully public. The proposed institution is real; a detailed price tag for its technology dependencies is not.

A centre is not a sovereign computing stack

A training workshop, standards office or applied research centre can help militaries use AI well. It does not by itself provide a national cloud, a secured training-data pipeline or software whose source code an armed force can inspect and maintain. Those are separate capital and operating decisions. If a government makes them together without revealing the lifetime cost, it risks spending public money to demonstrate autonomy while remaining commercially dependent on the same suppliers.

France's Defence Ministry has already developed domestic infrastructure for this challenge. Its Ministerial Agency for Defence Artificial Intelligence, AMIAD, was created in May 2024. A December 2024 ministry account described an initial €300 million budget for the agency and two principal hubs: a research operation at Palaiseau and a technical hub at Bruz near Rennes. The ministry has described a goal of recruiting 300 specialists by the end of 2026 and roughly 400 identified use cases. These are figures for the pre-existing French AI effort. They cannot be relabelled as the still-undisclosed cost of the new multinational Rennes centre.

That distinction is the first requirement for an honest ledger. The public needs to know whether ICAID receives its own appropriations, shares existing AMIAD resources, uses in-kind contributions from partner nations or relies on contracted services. The value of a ministry building, a laboratory used part time and dedicated annual programme funding should not be added together as though all three represented new cash outlays. Equally, shared facilities are not free: staff hours and computing cycles still have an opportunity cost.

France already has evidence of recurring demand

The ministry's internal GenIAl.intradef platform illustrates how a defence organisation's AI use can expand. France said on January 6, 2026, that GenIAl, launched in 2024, had reached 100,000 users by the end of 2025, handled 19 million annual requests and registered 84.2% overall user satisfaction. The platform supports functions including translation and is being developed toward internal search and more complex assistance. Those are ministry-reported usage and survey figures, not measures of battlefield accuracy, fiscal savings or costs.

The scale matters because each additional use case creates a recurrent workload. Even when a ministry develops software itself, computers need electricity, maintenance and hardware replacements. Secure deployments require classified infrastructure, identity controls and people cleared to manage incidents. Models need periodic evaluation and updates. Training is not a one-time cost if personnel rotate frequently. The attraction of a widely used platform is that common infrastructure may reduce duplication; the risk is that costs spread across budgets become harder to reconstruct.

One useful budget comparison is the average cost per request or per active user, but it must include more than the nominal price of an AI inference. It should include the infrastructure and staff required to deliver an acceptable answer securely. A ministry can report 19 million requests without claiming 19 million hours saved. It can also report 100,000 registered users without establishing how many use the service every week. Those distinctions matter when public officials make business cases for a second generation of defence software.

For Emmanuel Macron (opens in a new tab), the political prize is more than a European label. A credible sovereign programme could retain critical knowledge inside the French military, give domestic companies a route into allied markets and reduce the ability of external technology providers to interrupt service. The political vulnerability is the opposite: an expanding platform that delivers no independently demonstrated savings can become a protected line item because sovereignty is difficult to challenge.

The commercial model Europeans are buying into

Consider the scale of the American data-and-software business Europeans want to be less dependent on. Palantir's 2025 annual report shows total revenue of $4.475 billion, including $2.402 billion from government customers, up from $1.570 billion in 2024. That represents 53% growth in the government category. Its US government revenue was about $1.9 billion in 2025, the filing says. These are global company figures, not European defence contracts and not the price of any French project. They demonstrate that government software can be a major recurring revenue stream.

It would be misleading to infer that every dollar of government revenue is battlefield AI or that Palantir will automatically benefit from Rennes. No vendor was identified as a winner in the centre's launch announcement. Yet the filing explains the economic stakes behind procurement arguments: software sold to governments can generate significant revenue long after its initial integration. The suppliers of cloud services, data pipelines, cybersecurity tools and specialised consulting may also capture substantial portions of the total bill, even when they are not the names in the procurement headline.

A government that purchases a proprietary platform might obtain faster deployment and tested operational tools. The hidden cost can arise when formats, trained personnel and daily operating processes become inseparable from one supplier. Replacing that software can require reconstructing data integrations, retraining users and proving that the new service works at the same security classification. Such costs are real even if a procurement minister never writes them in the original press release.

A European alternative can create exactly the same lock-in under a different passport. A French company, German cloud provider or Italian integrator should not receive easier scrutiny because politicians describe it as a sovereignty champion. The economic question is whether competition, exit rights and public access to meaningful performance evidence remain intact.

NATO's own data strategy changes the argument

NATO's Data Strategy, approved in February 2025 and published later that spring, seeks a secure alliance data-sharing ecosystem in which nations retain control over their data. It sets a 2030 ambition for interoperable data resources, governance and an appropriately trained workforce. This is not a requirement to store all allied information in one giant system. The point is that systems should exchange trustworthy information while national authorities keep control over what they own.

That principle is directly relevant to the proposed Rennes centre. A multinational evaluation process may need to inspect models or metadata supplied by several governments. What may be pooled? Can partners withdraw their data? Must one contractor act as a gatekeeper for everything entering and leaving the testing environment? A sovereign architecture should allow partners to grant limited access without surrendering control of the underlying source data. Otherwise the centre could centralise precisely the power it is meant to distribute.

NATO's 2024 revised AI strategy also places emphasis on testing, evaluation, verification, validation, responsible use and interoperability. There is an existing alliance-level rationale for common assessment methods. The practical distinction is whether Rennes develops reusable evaluation techniques for each government or becomes dependent on a single commercial testing platform. Both could be called interoperability in a sales presentation, but only one would necessarily reduce future supplier leverage.

The defence-industrial production capacity (opens in a new tab) problem is not solely a matter of factory lines and missiles. For defence AI, it includes secure computing, software integration talent and an ability to deploy patches without waiting for permission from a distant vendor. A nation can own the hardware and still lack the skills to maintain the capability. The objective should be resilient operations, not ownership theatre.

How public budgets can conceal the long tail

France's 2026 defence budget proposal put the defence mission at €57.1 billion excluding pensions, up approximately €6.7 billion from the 2025 figure on the same basis. NATO's 2025 Hague declaration commits allies to 5% of GDP for defence and related security spending by 2035, with at least 3.5% for core requirements and up to 1.5% for security resilience and industry. The European Commission's Readiness 2030 discussion cites up to €800 billion of potential additional investment and a €150 billion SAFE loan instrument. None of these figures is an ICAID budget, and multiplying them into a supposed AI market size would be an abuse of accounting.

A project should instead be priced across its lifetime. Start with acquisition and integration. Add classified computing facilities, network security, data cleaning, monitoring, operators and refresh cycles. Estimate supplier support, contract changes, model revalidation and the cost of terminating the agreement. Then test the forecast against an alternative such as in-house provision, open standards or a smaller specialised tool. For classified work, detailed prices may need to be examined by protected parliamentary committees, but broad totals and the procurement method should still be accountable.

A useful public metric is the share of lifecycle expenditure that is guaranteed under contracts versus merely estimated. Another is the percentage of spend committed to foreign-owned services where source code, upgrades or user data cannot easily be transferred. A third is the number of suppliers capable of maintaining a system without an incumbent's approval. These metrics reveal dependence far better than the nationality of a brand printed on a slide.

The European taxpayer should not be asked to pay twice for the same capacity: once for an EU-backed public technology programme and again for a proprietary product that a national ministry must buy because the public system was never made usable. Nor should officials count a potential loan as a grant, or a NATO spending target as an invoice paid. The defence-expenditure (opens in a new tab) indicator tracks commitment and spending pressure; efficiency requires different evidence.

Who gains, who pays, and the real alternatives

Military users gain when AI reduces repetitive analysis and helps them identify relevant information faster. Suppliers gain from integration work, software subscriptions and service agreements. National governments gain political credit when a programme is described as sovereign. Local research institutions and skilled employees around Bruz, Rennes and Palaiseau may also gain from a sustained pipeline of projects. But these benefits are not interchangeable with a proven improvement in deterrence or a quantified return to taxpayers.

The people paying are those funding current budgets and future debt service, as well as other public services competing for the same qualified staff and capital. Allied taxpayers may pay separately if each nation insists on nearly identical infrastructure for political reasons. Yet a single shared supplier could create a different systemic risk if an outage, export restriction or contractual dispute disables many allied users simultaneously. There is no cost-free design. The answer is to expose and manage the trade-off.

A good procurement approach is to insist on exportable data, documented interfaces, independently repeatable tests and the ability to operate without uninterrupted access to a vendor's central servers. These conditions may increase the initial cost or delay delivery. They can reduce long-term dependency. A government should calculate that difference rather than claim that the cheapest upfront bid is necessarily the best value.

What the November 20 meeting should reveal

The upcoming conference should clarify membership, governance, operational tasks and whether nations will provide money, experts, facilities or data. The 2027 milestone needs a definition: is it a functioning test laboratory, a doctrine group, a network of participating laboratories or something else? A separate published architecture should distinguish AMIAD's existing French capabilities from those genuinely created for the multinational centre.

Subsequent French budget annexes, alliance policy papers and company filings can answer the financial questions that the first announcement cannot. Watch for long-duration services contracts, cloud or supercomputing arrangements, common data standards and clear termination clauses. Watch as well for evidence that operational users prefer the systems because they are better, not because they are compulsory. A single successful demonstration is insufficient.

The defensible verdict is that Europe is right to treat AI infrastructure as a security dependency, but wrong if it measures autonomy only through a national logo or new headquarters. The public-money test is whether allied militaries can control their data, replace a supplier, maintain their tools under pressure and compare the full lifetime price with alternatives. Rennes may help organise that discipline. It has not yet proved the economics.

Sources

• French Ministry of Armed Forces: AMIAD six-month assessment and €300 million agency budget, 5 December 2024 (opens in a new tab)
• French Ministry of Armed Forces: GenIAl.intradef reaches 100,000 users, 6 January 2026 (opens in a new tab)
• Palantir: 2025 Form 10-K annual report, government and total revenue tables (opens in a new tab)
• NATO: Data Strategy for the Alliance, approved February 2025 (opens in a new tab)
• NATO: Revised Artificial Intelligence Strategy, 10 July 2024 (opens in a new tab)
• Euronews: Rennes AI defence centre announcement, 12 November 2026 (opens in a new tab)
• European Commission: Future of European defence and Readiness 2030 (opens in a new tab)