Nvidia Expands AI Facilities in Europe, Unveils Design to Limit AI Permissions
Cloud partners are adding AI compute capacity across Europe and beyond, and the use of openly available models tailored to specific work is expanding as well. A design for limiting which business tools AI can access and what actions it can take arrived alongside it.
On October 21, Nvidia announced plans by its cloud partners to expand AI facilities, particularly in Europe. The buildout covers compute capacity for both model training and running live services. Alongside it, the company unveiled a design for limiting the scope of what agentic AI, meaning AI that carries out tasks on its own, is allowed to reach. The expansion regions also include Northern Europe, Africa and Central Asia.
Why build AI facilities close to the data
The announcement speaks to a question companies are wrestling with: where should their data and their AI actually run? Nvidia said demand is growing on two fronts at once, from organizations that want to compute close to where their data already sits, and from those that want to adapt openly available models to their own work. Partners are adding facilities meant to support everything from training through to running real services.
Test results also pointed to higher throughput within a fixed power budget. According to Nvidia, Nscale used DSX MaxLPS in testing on GB300 NVL72 systems to run 37% more GPUs within the same power limit. Overall throughput for AI generating responses rose roughly 50%. Nscale plans to bring in Vera Rubin infrastructure as it expands into five European markets.
Crusoe is developing Spark, an AI facility designed for Vera Rubin. It is a modular, liquid-cooled design. The goal is to place compute closer to users and cut the time it takes AI to return an answer. Aleria, Cassava Technologies and Firebird are among the others building or expanding Nvidia-based facilities.
Models tuned for the job, at half the cost
As an example of adapting an openly available model to a specific task, Nvidia pointed to the code review service CodeRabbit. Baseten built a dedicated model on top of Nemotron 3.5 Lightning that decides where to route code review work. Nvidia said that after additional training, routing agreement on the evaluated tasks rose from 75.8% to 80.7%. The company estimated that the cost of generating responses for the same tasks was cut in half.
Image: NVIDIA Blog
Investment in services and model development continues as well. Nebius has adopted NVIDIA Groq 3 LPX for its platform, and plans to pair it with Vera Rubin NVL72 to improve response speed. Mistral AI has folded Vera Rubin infrastructure, slated for delivery in 2027, into its own expansion plans. According to Nvidia, Mistral Large 4 was trained on 3,800 Grace Blackwell GPUs in European data centers operated by Mistral AI.
Who decides what AI can touch
Nvidia also introduced a design for keeping AI connected to business systems under control. The OpenShell Reference Architecture, presented at GTC Berlin, requires organizations to define in advance what an AI system can access and which actions it can take. The aim is to ensure that when AI uses tools to get work done, it operates only within the permissions an organization has granted.
In this architecture, control is divided across several pieces of hardware and software. Vera Rubin NVL72 handles the work of models generating responses. The Vera CPU and OpenShell handle AI execution and tool use. BlueField-4 runs Sentry software to enforce limits at the chip level. Cloud partners including CoreWeave, DigitalOcean, Nebius and Nscale support the design.
For enterprise users, the criteria for choosing a service now extend beyond throughput to where data resides and how far AI is permitted to act. The open questions are when these expansion plans will reach live services, and how effective the cost savings and permission controls prove to be in each working environment.