Nvidia, Perplexity to partner with EU and Middle East AI firms to build sovereign LLMs

Nvidia and AI search firm Perplexity said they are joining hands with model builders and cloud providers across Europe and the Middle East to refine sovereign large-language models (LLMs) and accelerate enterprise AI uptake in local industries.

Under the deal, model builders and AI consortia from the region will fine-tune their language models with Nvidia’s Nemotron techniques, a move the company says will slash costs and boost accuracy for enterprise tasks, including emerging agentic AI.

Some of the model builders and AI consortia that Nvidia and Perplexity will be working with include Barcelona Supercomputing Center (BSC), Bielik.AI, Dicta, H Company, Domyn, LightOn, the National Academic Infrastructure for Supercomputing in Sweden (NAISS) together with KBLab at the National Library of Sweden, the Slovak Republic, the Technology Innovation Institute (TII), the University College of London, the University of Ljubljana and UTTER.The models will be trained and run on servers within Europe supplied by Nvidia Cloud Partners through the DGX Cloud Lepton marketplace, Nvidia said in a statement.

“The open, sovereign models will provide a foundation for an integrated regional AI ecosystem that reflects local languages and culture,” Nvidia added. “Europe’s enterprises will be able to run the models on Perplexity, an AI-powered answer engine used to answer over 150 million questions per week.”

Strengthening EU presence

The deal cements Nvidia’s role in Europe’s sovereign-AI drive and expands Perplexity’s regional distribution reach.

“This not only increases model choice for enterprises but also disrupts legacy procurement patterns for GPUs, cloud capacity, and AI infrastructure,” said Sanchit Vir Gogia, chief analyst and CEO at Greyhound Research. “As Nvidia rolls out its AI Factories across Europe, including large-scale H100 and Blackwell GPU clusters, supply will rise, but so will competition for priority access.”In a separate statement, Nvidia has said it will help build an AI factory in Germany to power industrial AI workloads for European manufacturers.

“For CIOs, the implication is clear: GPU reservation and pricing are no longer background infrastructure but board-level variables,” Gogia added. “Enterprises must now negotiate burst entitlements, reserve tiers, and failover pathways to hedge against partnership-induced resource contention. This alliance may strengthen Europe’s AI sovereignty, but it also intensifies infrastructure nationalism, and buyers must recalibrate accordingly.”

Other analysts said the partnership could further strain Europe’s supply of high-end GPUs.

“This move will likely intensify demand for high-performance compute (HPC) resources across Europe,” said Himanshu Mhatre, senior analyst at Everest Group. “As more organizations train or fine-tune models locally, enterprise access to GPUs, especially H100s and upcoming Blackwell chips, may tighten, causing spot-market pricing to spike and longer lead times for new deployments.”

Managing data compliance concerns  

The agreement >comes amid growing scrutiny of AI firms’ data handling, with the EU’s AI Act set to require general-purpose and high-impact models to disclose how they are trained, fine-tuned, and updated.

Almost 67% of CIOs in France, Germany, and the Nordics carry out quarterly audits of their AI systems to check compliance with the EU AI Act, and 54% of those executives cite “local hosting without upstream transparency” as a regulatory blind spot, according to Greyhound Research.

“Hosting such models in European data centers supports compliance with GDPR and national data sovereignty ambitions but does not absolve vendors of obligations under Articles 13 and 53–55 of the AI Act,” Gogia said. “These require a ‘sufficiently detailed summary ‘ of training data, technical documentation on model risks, and structured record-keeping for public audits.”

CIOs should make vendors disclose where their training data comes from, including synthetic or non-EU sets, explain how the models produce results, and ensure back-end services do not send data outside the bloc, Mhatre said.