It's in the announcement: "ML4 was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral’s own datacenters in Europe".

https://mistral.ai/news/mistral-large-4/