One Platform, Two Regimes: Where the AI Act's Military Edge Sits

Analysis as of 14 August 2026.

Two dates in the summer of 2026 quietly reorganised the compliance calendar for every company selling artificial intelligence into both barracks and boardrooms. On 27 July, Regulation (EU) 2026/1744 came into force and rewrote the timetable of the AI Act while leaving its scope article untouched. On 2 August, Regulation (EU) 2024/1689 itself became generally applicable — with the heaviest obligations pushed outward: the duties attached to the high-risk uses listed in Annex III now bite from 2 December 2027, and those tied to the product legislation in Annex I from 2 August 2028. The penalty ceilings, meanwhile, are already written into law: up to €35 million or 7 per cent of worldwide turnover for prohibited practices, up to €15 million or 3 per cent for the main operator duties, up to €7.5 million or 1 per cent for supplying misleading information, and for general-purpose model providers a Commission fining power of up to 3 per cent or €15 million — a power that itself only became exercisable in August 2026.

The pivot of the whole architecture is a single word in Article 2(3): exclusively. Systems escape the regulation where, and in so far as, they are marketed, deployed or used exclusively for military, defence or national-security purposes. That carve-out belongs to the system and its use, not to the seller. A defence prime holds no sectoral immunity; a police force enjoys no automatic coverage. Recital 24 makes the line divisible: the moment a militarily developed system is also used — even temporarily — for civilian, humanitarian, policing or public-security ends, the regulation applies to that use. A single technical core can therefore live simultaneously inside and outside the statute, and the burden of proving where each deployment sits falls on the supplier's records: versions, weights, access controls, datasets, contracts, support arrangements.

The machinery that would test such claims is only half assembled. Member States had to notify their competent authorities and single points of contact by 2 August 2025; the European Parliamentary Research Service found that by March 2026 the Commission's list held eight entries out of twenty-seven. Italy has moved, designating AgID and ACN under Law No 132 of 23 September 2025, and the European Data Protection Supervisor watches over the Union's own institutions. But the Court of Justice of the European Union has yet to interpret the military exclusion, and no published enforcement decision defines what a minimum evidence file for exclusive military use looks like. The European Court of Auditors, in Special Report 08/2024, had already judged Union-national coordination in this policy field deficient — an audit of investment governance that closed before the regulation applied, yet one that maps precisely the seam along which scope disputes will run.

The upstream layer is the least settled of all. The statutory exclusion speaks of AI systems; general-purpose models are brought within scope separately, and Commission material points toward model obligations that attach regardless of what downstream customers do. Feeding a foundation model into a defence integration chain does not, on any official source, erase the provider's Chapter V duties. And since Implementing Regulation (EU) 2026/1755, adopted on 20 July 2026, the Commission holds a formal power to reach model weights, source code and hosting infrastructure — an investigatory instrument that arrived before most of the substantive rules it would serve.

Until December 2027, then, the boundary is being drawn privately, contract by contract, by suppliers and customers rather than by authorities — and two questions will decide who drew it well. The first: what degree of technical and documentary separation — distinct builds, segregated data, independent configuration authority — will a regulator or judge eventually demand before accepting that one engineering platform is genuinely two systems, one exempt and one regulated? The second: when a model or a shared component sits upstream of both a battlefield deployment and a commercial one, who in the value chain carries the compliance cost that the exclusion was assumed, wrongly, to have removed?