A new open-source model and report published as EuroMesh asks whether Europe can train a sovereign frontier AI without waiting for its planned gigawatt datacentres to come online. The answer it gives is yes, conditionally.
Europe already operates tens of exaflops of public AI compute across the EuroHPC supercomputer fleet and 19 national AI Factories. The problem is those machines are shared, batch-scheduled, and heterogeneous, so the addressable fraction for a single training run is a political decision, not a hardware one. The EuroMesh report models federated training using a DiLoCo-style low-communication approach and finds the federation can deliver a frontier-class model around 2028. A new 1 GW campus, by contrast, faces a mean grid-connection wait of 7.6 years, putting first training around 2033.
The catch the authors are careful to flag: frontier-scale distributed training is unproven above about 10 billion parameters, so the 2028 target is a credible stopgap, not a guarantee. The model is fully reproducible, with 52 pytest tests and sourced parameters. Whether European governments can align on actually running a coordinated training job across their existing machines remains the harder open question.