AI Office: EU has two-year window to catch frontier AI

Brussels does not usually put dates on strategic anxiety, so the sentence that stands out from the AI Office's new report is a specific one: the next one to two years could be decisive for Europe on frontier AI. That framing came out of the Expert Forum on Frontier AI, whose findings the AI Office published on 15 July 2026, drawing on over 100 experts from AI developers, industry, investors, academia, Member States and Commission representatives.

The diagnosis is blunt. Frontier AI development is 'currently concentrated overseas', and the report is explicit that 'non-frontier models are an imperfect substitute', which is another way of saying Europe cannot simply lean on smaller open models and call it a strategy. Talent and research the EU has; what it lacks, in the experts' account, is comparable computing infrastructure and growth-stage capital, alongside legal clarity on training data that would let European labs train at frontier scale. Computing infrastructure, and the energy to power it, are described as the most urgent two-year priorities.

Why that timing matters is in the report's own pace claim: capabilities have advanced 'within roughly three years from struggling with basic tasks to approaching the limits of what current tests can measure.' A two-year window on that curve is roughly the gap between one generation of frontier system and the next, not a comfortable policy horizon. The Commission's vehicle is the European Frontier AI Initiative, announced under the Apply AI Strategy, with 2030 set as the year by which the ambitious targets should land.

The sovereignty framing in the report is careful rather than autarkic. The stated goal is to strengthen the EU's sovereign ability to 'access, select, control, and benefit from frontier AI models' through collaboration with trusted partners, which is a long way from 'build our own OpenAI'. It concedes that in the near term Europe will still consume some of its frontier capability from outside.

The honest caveat is that this is a set of directional findings from a forum, not a budget. The report names no individual experts, no euro figures, and no specific compute commitments, and the reporting doesn't tell you which 'trusted partners' the sovereignty text is pointing at, or how much public compute the EU is actually prepared to underwrite. If the political capital does follow the timeline the experts have put on the table, operators building on European infrastructure could see procurement, capital and legal reform arriving together, which would be the first coordinated push in a while that treats frontier AI as a compute problem as much as a governance one.