The European Union’s public supercomputing infrastructure has demonstrated its capacity to rival proprietary Big Tech clusters. On November 12, 2026, the EUROPA Consortium (comprising Domyn and Fraunhofer institutes) published a technical report on the PyTorch blog detailing the training of a 951.9-billion-parameter Mixture-of-Experts model (32.8 billion active per token) on the European JUPITER supercomputer.
Through deep memory and inter-GPU communication optimizations on NVIDIA GH200 accelerators, engineers scaled training throughput from 90 to 1,300 tokens per second per GPU. For European Sovereign AI, this marks a strategic milestone, proving that open-source PyTorch and torchtitan stacks can train trillion-scale foundation models on public compute infrastructure.
Sources: PyTorch Blog
Europe’s Trillion-Parameter Milestone: EUROPA Consortium Trains MoE Model on JUPITER
