NVIDIA Expands European Sovereign AI Clouds With Vera Rubin and OpenShell
NVIDIA partners expand sovereign AI clouds across Europe, adding Vera Rubin capacity, faster inference, open models and OpenShell safety controls for AI agents.
Summary
On October 21, 2026, NVIDIA detailed cloud expansion across Europe, the Nordics, Africa and Central Asia for sovereign training, inference, agentic AI and physical AI. Nscale tests on GB300 NVL72 found DSX MaxLPS ran 37% more GPUs and delivered nearly 50% more aggregate inference throughput within one power budget; it plans Vera Rubin across five European markets. Crusoe is developing Spark, liquid-cooled modular Vera Rubin AI factories for low-latency Crusoe Cloud capacity serving Base44, Gradium, LTX and Wonderful. Aleria, Cassava Technologies, Firebird, G42, Radiant, SWI Compute and Vultr are expanding DSX AI Factory, Vera Rubin, BlueField and Spectrum-X Ethernet infrastructure.
Nebius adopted NVIDIA Groq 3 LPX with Spectrum-X and Vera Rubin NVL72 for managed serverless and dedicated Token Factory endpoints and APIs, while its Physical AI Workbench supports Rhoda AI, RoboForce and Sereact. It launched three early-access retail agent blueprints for inventory, store operations and fine-tuning, is expanding Physical AI Living Lab, and will supply computers for SPRIND’s German Next Frontier Challenge. Baseten fine-tuned Nemotron 3.5 Lightning for CodeRabbit, raising routing agreement from 75.8% to 80.7% and cutting estimated inference cost 50% on the evaluated workload. Fireworks and Era4 pair open-model software with Blackwell Ultra capacity in Chesterfield for UK and European finance, insurance and life sciences. Mistral Large 4, trained on 3,800 Grace Blackwell GPUs in Mistral-operated European data centers, targets coding, agents and multimodal tasks; Mistral plans Vera Rubin delivery in 2027. Verda plans Vera Rubin, with Nemotron, Cosmos and Dynamo already supporting open-weight production and inference optimization.
Announced at GTC Berlin, OpenShell Reference Architecture combines Vera Rubin NVL72, Vera CPUs running OpenShell, and BlueField-4 DPUs running Sentry to restrict agent access and actions. Support comes from Aleria, Baseten, Cassava, CoreWeave, Crusoe, DigitalOcean, Firebird, Fireworks, Firmus, G42, GMI Cloud, IREN, Lambda, Lightning AI, Nebius, Nscale, Radiant, Verda and Vultr.
Positives
- Nscale’s DSX MaxLPS tests supported 37% more GPUs and nearly 50% higher aggregate inference throughput within the same power budget.
- Baseten raised CodeRabbit routing agreement from 75.8% to 80.7% while reducing estimated inference costs by 50% for the evaluated workload.
- Mistral Large 4 was trained on 3,800 Grace Blackwell GPUs in European data centers for coding, agentic and multimodal workloads.
- Nebius launched three retail agent blueprints and provides a Physical AI Workbench used by Rhoda AI, RoboForce and Sereact.
- Nineteen cloud partners support OpenShell architecture for controlling agent access and actions from silicon through software.
Risks & concerns
- Power is already constraining AI factory growth, making the efficiency of each provisioned watt critical to further capacity expansion.
- Nscale’s 37% GPU gain and nearly 50% throughput improvement came from tests, not disclosed production deployments.
- Vera Rubin capacity from Nscale, Crusoe, Mistral and Verda remains planned, under development or scheduled for future delivery.
- Nebius’s three retail agent blueprints remain in early access rather than general availability.
- Agents connected to tools and business systems create access and action risks that OpenShell is designed to contain.