The European Commission is preparing talks with leading frontier AI laboratories over how the industry’s most advanced systems are developed, while seeking deeper cooperation with the UK and Canada on model evaluation, verification, early warning, and AI security.

European Commission President Ursula von der Leyen set out the approach in her State of the Union address, arguing that risks associated with increasingly capable models and self-improving systems require greater coordination between governments and developers. She said she would invite leading frontier laboratories to discuss how public authorities could support existing industry efforts to “pace the frontier”.

The proposal does not amount to a moratorium, a new legal limit on model training, or an announced change to the EU AI Act. Von der Leyen described a planned discussion with developers alongside international work on safety and evaluation, leaving the substance of any future measures open.

At the same time, she argued that Europe should substantially increase its own computing capacity and find new ways to combine public and private capital for European AI companies. The policy direction therefore treats advanced AI both as a security issue requiring additional coordination and as a strategic technology in which Europe wants greater industrial capability.

Risk coordination moves beyond legislation

Von der Leyen linked the frontier-model discussion to cybersecurity and self-improving systems, warning that more capable models could strengthen offensive capabilities and referring to incidents involving AI agents behaving outside intended environments. Those claims formed the basis for her call to work with international partners rather than an announcement of a new EU regulatory instrument.

The proposed cooperation with the UK, Canada, and other governments would focus on model evaluation, verification, early warning, and security. Those areas sit partly outside the conventional work of writing product rules because governments need technical methods for testing systems, sharing information about emerging risks, and deciding whether evaluations performed in one jurisdiction are credible elsewhere.

International coordination could reduce duplication for developers if countries converge on testing methods, although questions remain over access to models, confidential technical information, reporting thresholds, and the extent to which governments accept one another’s assessments.

The EU already has a legal framework in the AI Act, including obligations around general-purpose AI and systemic risk, but frontier development changes more quickly than legislation can easily be rewritten. Evaluation and early-warning arrangements offer another policy layer that can develop through technical standards, regulator cooperation, and agreements with developers.

How voluntary the proposed discussions remain will therefore be important. A government-hosted process can produce shared testing practices or voluntary commitments without changing the law, while concerns raised through those discussions could later feed into guidance, standards, or legislation. Von der Leyen did not specify which route the Commission intends to take.

Europe is also expanding its AI capacity

The risk agenda sits alongside an expansionary industrial policy. Von der Leyen said Europe needs to “massively boost” its computing capacity and promised further proposals on achieving that with lower energy demands, while also signalling new mechanisms for directing public and private capital towards European startups and technology companies.

Techopia recently examined Deutsche Telekom’s planned bid for an EU AI Gigafactory, part of a wider effort to increase the amount of high-end computing available to European developers and businesses.

Von der Leyen also drew a distinction between owning the most advanced frontier model and extracting economic value from AI. She identified health, transport, agri-food, advanced manufacturing, and defence and space as sectors where European industry holds valuable data and where the Commission intends to announce further initiatives in November.

The largest general-purpose AI developers remain concentrated outside the EU, while European manufacturers, healthcare systems, industrial groups, and public bodies control substantial domain-specific information and physical infrastructure. Brussels is increasingly trying to turn those assets into leverage for specialised AI adoption rather than measuring competitiveness only through the largest training runs.

The strategy contains an inherent policy tension. More compute and greater investment can accelerate capabilities available to European developers, while concerns about cybersecurity and self-improving models create pressure to slow or constrain some forms of development. Those objectives are not necessarily incompatible, but they require policymakers to distinguish between the systems they want deployed more quickly and capabilities they believe warrant additional caution.

The UK’s role is particularly relevant because Britain sits outside the EU’s legal framework while maintaining its own AI-safety institutions and model-evaluation capability. Cooperation on evaluations or early warnings could provide practical alignment without requiring identical regulation.

No new compliance duty entered into force through the speech. The next steps will be the Commission’s invitation to frontier laboratories, the structure of international cooperation, forthcoming compute proposals, and the sector initiatives expected in November.