NVIDIA highlights expanding European “AI cloud” network, pushing open models and agent safety controls

NVIDIA on Wednesday mapped out how its cloud partners are scaling AI infrastructure across Europe and adjacent regions, linking that buildout to two pressures developers are facing: where computing happens, and what “agentic” systems are allowed to do in the real world. In a new company blog post, NVIDIA argues that AI-native developers increasingly need more compute capacity closer to data and enterprise workloads, plus guardrails that define boundaries for tool use and access.

The company said its NVIDIA Cloud Partners are expanding regional capacity for both model training and inference, explicitly tying the move to use cases that include agentic applications and physical AI. NVIDIA’s description emphasizes “sovereign” compute needs, noting that regional clouds can give customers more choice over where workloads and data reside, rather than forcing everything through a centralized stack.

A major constraint NVIDIA called out is power. As what it described as “AI factories” scale, the company says electricity supply and efficiency become limiting factors, pushing data centers to squeeze more useful inference or training work out of each watt. In that context, NVIDIA pointed to Nscale, describing tests on NVIDIA GB300 NVL72 systems where Nscale used NVIDIA’s DSX Platform to coordinate power allocation with workload requirements.

NVIDIA said those tests showed DSX MaxLPS enabled 37% more GPUs and nearly 50% higher aggregate inference throughput within the same provisioned power budget. The post also says Nscale plans to deploy NVIDIA Vera Rubin infrastructure for frontier-model training and production workloads across five European markets, positioning it as part of a vertically integrated full-stack platform.

NVIDIA’s broader European partner list also spans companies focused on different layers of the AI infrastructure stack. Crusoe, for example, was described as building “Spark,” liquid-cooled modular AI factories designed for the NVIDIA Vera Rubin platform, with Crusoe Cloud capacity aimed at low-latency inference. The blog lists Nebius as adopting NVIDIA Groq 3 LPX for token generation, integrating it with Vera Rubin NVL72 on Nebius’ Token Factory using Spectrum-X Ethernet networking, and also describes Nebius’ Physical AI Workbench for robotics and physical AI deployments.

Beyond capacity and performance, NVIDIA devoted significant attention to agent controls. At GTC Berlin, the company said it announced the NVIDIA OpenShell Reference Architecture, intended to give organizations full-stack control when deploying AI agents, including “clear boundaries” around what agents can access and do. NVIDIA described multiple components working together: NVIDIA OpenShell secure runtime software, Vera Rubin NVL72 for model inference, Vera CPUs with OpenShell for agent harnesses and tool use, and NVIDIA BlueField-4 DPUs running NVIDIA Sentry for in-silicon enforcement.

NVIDIA said a wide set of cloud and infrastructure partners are supporting the OpenShell reference architecture, listing companies including Aleria, Baseten, Cassava, CoreWeave, Crusoe, DigitalOcean, Firebird, Fireworks, Firmus, G42, GMI Cloud, IREN, Lambda, Lightning AI, Nebius, Nscale, Radiant, Verda and Vultr. The blog frames this as a response to the shift from static model use to production systems where AI models are connected to tools and business systems, making governance and limitation mechanisms more important.

The post also highlighted partner examples of model customization and domain-specific optimization. Baseten was cited for a specialized code-review routing approach used by CodeRabbit, using NVIDIA Nemotron 3.5 Lightning with supervised fine-tuning and reinforcement learning. NVIDIA said CodeRabbit improved routing agreement from 75.8% to 80.7% while reducing estimated inference costs by 50% for the evaluated workload. Elsewhere, Fireworks and Era4 were described as combining Fireworks’ training and inference platform with direct liquid-cooled capacity powered by NVIDIA Blackwell Ultra GPUs at Era4’s Chesterfield facility for early enterprise AI solutions in financial services, insurance, and life sciences.

NVIDIA also pointed to regional expansion timelines for major infrastructure builds. Mistral was described as expanding NVIDIA-accelerated computing in European data centers, with NVIDIA Vera Rubin infrastructure planned for delivery in 2027, and training its latest frontier multimodal model, Mistral Large 4, on 3,800 NVIDIA Grace Blackwell GPUs. In addition, Verda was described as planning Vera Rubin deployments while promoting open-weight models, including NVIDIA Nemotron and NVIDIA Cosmos in production, using NVIDIA Dynamo for inference optimization.

While NVIDIA’s blog lays out a detailed partner ecosystem and the architecture’s building blocks, it does not provide pricing, capacity volumes, or independent performance benchmarks beyond the examples cited. It also does not specify how consistently the “agent boundaries” approach will be implemented across all partners or how quickly new controls will be adopted for specific regulated industries.

For the next phase, NVIDIA’s emphasis suggests two things to watch: whether regional AI cloud providers translate the stated power-efficiency and infrastructure plans into measurable scaling for customers, and whether OpenShell-based governance becomes a standard layer for agent deployments. As more deployments move from experiments to production tool use, the practical adoption of enforcement mechanisms and reference implementations may become as important as raw compute availability.