Mistral says ML4 was trained using 3,800 Nvidia Grace Blackwell GPUs in its own data centers in Europe and much of its training data was multilingual (Mistral Blog)

Mistral’s ML4 launch signals that frontier AI models can now be trained and operated on large, purpose-built GPU fleets in European data centers, which strengthens regional data sovereignty and may appeal to enterprises with strict privacy, residency, and regulatory requirements. For CIOs, the strategic takeaway is that AI sourcing is becoming a platform and geography decision as much as a model decision: IT teams should evaluate vendors not only on capability and cost, but also on multilingual performance, compliance posture, infrastructure location, and long-term resilience.