NVIDIA has announced a collaboration with model builders and cloud providers across Europe and the Middle East to enhance sovereign large language models (LLMs). This initiative aims to boost enterprise AI adoption in the region.
The collaboration involves various organizations, including Barcelona Supercomputing Center (BSC), Bielik.AI, Dicta, H Company, Domyn, LightOn, the National Academic Infrastructure for Supercomputing in Sweden (NAISS) alongside KBLab at the National Library of Sweden, the Slovak Republic, Technology Innovation Institute (TII), University College London, University of Ljubljana, and UTTER. These entities will work with NVIDIA to optimize their models using NVIDIA Nemotron techniques. The goal is to improve cost efficiency and accuracy for enterprise AI workloads.
AI infrastructure in Europe from NVIDIA Cloud Partners (NCPs) participating in the NVIDIA DGX Cloud Lepton marketplace will handle model post-training and inference. The open models are designed to support an integrated regional AI ecosystem that reflects local languages and cultures. Enterprises will be able to run these models on Perplexity, an AI-powered answer engine used extensively across Europe.
Jensen Huang, founder and CEO of NVIDIA stated: “Europe’s diversity is its superpower — an engine of creativity and innovation.” He added that they are building an AI ecosystem locally developed to help Europe thrive in the age of AI.
To accelerate Europe's transformation through AI-driven technologies across multiple industries such as manufacturing, healthcare, finance, energy, and creative sectors; LLMs will support all 24 official European languages. Models specialized in national languages include contributions from H Company and LightOn in France; Dicta in Israel; Domyn in Italy; Bielik.AI in Poland; BSC in Spain; NAISS and KBLab in Sweden; TII from UAE; University College London from UK among others.
These LLMs will be enhanced using NVIDIA Nemotron model-building techniques like neural architecture search combined with reinforcement learning. Post-training optimizations aim to reduce operational costs while improving user experiences by speeding up token generation during inference. The post-training workloads will run on DGX Cloud Lepton hosted by European NCPs such as Nebius, Nscale, and Fluidstack.
Developers can deploy these sovereign models as NVIDIA NIM microservices running on-premises or via cloud service provider platforms using a new NIM microservice supporting over 100k public/private/domain-specialized LLMs hosted on Hugging Face.
Perplexity plans integration of these sovereign AI models into its answer engine used by enterprises across telecommunications/media sectors within Europe enhancing accuracy/trustworthiness through improved search query/AI output performance according Aravind Srinivas cofounder/CEO Perplexity who said: “Bringing NVIDIA-optimized sovereign AI models empowers innovation.”
The first distilled models are expected later this year according availability details provided during announcements made at GTC Paris keynote delivered by Huang at VivaTech event encouraging exploration sessions related topics discussed therein.