The procedural picture in April: Parliament voted on 26 March 2026 to merge Sections A and B of Annex I. The AI Omnibus, delivered in November 2025, was on a timetable of a second political trilogue on 28 April, an IMCO and LIBE committee vote in May, a plenary vote in June, and publication in the Official Journal targeted for July. Meanwhile the high-risk obligations were still scheduled to take effect on 2 August 2026, before the postponements later agreed.
Forty people for the frontier
Unit A3 of the EU AI Office is responsible for general-purpose AI models presenting systemic risk. It was planned to reach roughly 40 staff. A civil-society coalition has asked instead for 100 million euro a year and a minimum of 100 staff by 2030. Whichever figure one prefers, the point is that the most technically demanding supervisory task in the regime is assigned to a team of a size a mid-sized law firm would recognise.
Compute is not the constraint
The contrast with infrastructure spending is sharp. Nineteen AI Factories have been deployed across European supercomputers. The AI Gigafactories concept attracted 76 expressions of interest across 16 member states, with around one billion euro mobilised for strategic sectors. The portfolio sits with Henna Virkkunen, Executive Vice-President for Tech Sovereignty, Security and Democracy. Europe is not short of ambition or of hardware. It is short of regulators.
A concrete illustration
Anthropic limited the release of its Mythos model to twelve US-headquartered firms. Among European national agencies, only Germany’s cyber agency was contacted; no other obtained access. A supervisor that cannot get access to a model cannot evaluate it, whatever powers the statute confers. This is the practical form the capacity problem takes.
For anyone teaching the general-purpose AI provisions, a useful new resource appeared in the same period: the Cambridge Commentary on the GPAI provisions, launched jointly by the Leverhulme Centre for the Future of Intelligence at the University of Cambridge and the Institute for Law and AI.
The lesson generalises beyond Brussels. In a jurisdiction building its own institutions, the binding constraint on any technology regulation is rarely the drafting. It is whether a body exists with the people, the access and the standing to apply it.
Drawn from The EU AI Act Newsletter #100, 20 April 2026.