Distributed AI: A European Path to the Frontier. A Manifesto for a European Network for the Distributed Development of Foundation Models

Europe is asking itself whether it can develop “frontier models”. This paper argues that the answer lies not primarily in replicating the centralised model of the great digital powers, but in methodically organising what Europe already possesses in abundance: world-class research teams, data that is rich but cannot be centralised, public computing infrastructure spread across the whole continent, and a long tradition of loosely coupled cooperation. It therefore proposes a European network for distributed AI: an open alliance of industrial actors, laboratories, computing centres and free-software communities that jointly develop the tools, the interoperability standards and the governance needed to collectively train, compose and audit state-of-the-art foundation models. We show that the scientific base is mature, at the convergence of low-communication distributed training (DiLoCo, federated pre-training), of modular composition (mixtures of experts, model merging, MoErging) and of the modular architectures cognitive science arrives at independently (Global Workspace Theory, world models, JEPA), and that the enabling technologies exist today. We then answer the objection that downloadable open-weight “mastodons” make such an effort redundant: open weights are neither open source nor a commons, they renew dependence at every generation, their concentration in fact puts them out of reach of the European reality of deployment, and their own sparse mixture-of-experts architecture vindicates the compositional thesis, only privately. We further show that the security requirements now attached to frontier models, auditable provenance, sovereign evaluation capacity, deployability in constrained environments and governed dual use, are met more directly by a distributed commons than by any downloaded artefact, and that they bind the network to specific obligations. Reading the 2026 findings of the EU AI Office Expert Forum on Frontier AI alongside the Frontier AI Initiative, the EUROPA consortium and OpenEuroLLM as an unprecedented institutional alignment, the paper closes on seven commitments: federate, pool, build the tooling, standardise, demonstrate, be accountable, secure. This record holds the reference version of the manifesto, in PDF and in editable ODT form. Shorter derivative documents, a diffusion version and a policy brief for institutional interlocutors, circulate separately. Expressions of interest in joining the network are welcome at jplorre@linagora.com.