# EU consortium trains 1‑trillion‑parameter AI model on public supercomputer

> Using PyTorch and TorchTitan, the EUROPA team achieved ~1,300 tokens per second per GPU on 384 NVIDIA GH200 GPUs, proving frontier‑scale training on shared EU infrastructure.

Oossa · 2026-11-12 · https://oossa.com/en/eu-consortium-trains-1-trillion-parameter-ai-model-on-public-supercomputer

The EUROPA consortium – Domyn, Fraunhofer IAIS and IIS – trained a mixture‑of‑experts model with about 1 trillion parameters on JUPITER, the EuroHPC exascale supercomputer at Forschungszentrum Jülich. By tuning PyTorch and TorchTitan they raised throughput from 90 to roughly 1,300 tokens per second per GPU across 384 GH200 GPUs (96 nodes). The work shows that public, sovereign compute can support AI at the same scale as private clusters.

## The facts

- Throughput: ~1,300 tokens/s per GPU on 384 NVIDIA GH200 GPUs
- Model size: 951.9 B total parameters (32.8 B active per token)

## Why it matters

It demonstrates that European public compute resources can run frontier‑scale AI, reducing reliance on private cloud rentals.

## Sources & references

1. [Training a 1T-Parameter MoE on EU public infrastructure with PyTorch and TorchTitan](https://pytorch.org/blog/training-a-1t-parameter-moe-on-eu-public-infrastructure-with-pytorch-and-torchtitan/) – PyTorch

Last updated: 2026-11-12