NVIDIA Cloud Partners Scale Across Europe — Sovereign, Open and Built for Safe AI Agents

Nvidia used GTC Berlin to showcase how its cloud partners across Europe, the Nordics, Africa and Central Asia are adding regional AI capacity. Nscale, Crusoe, Nebius, Mistral, Fireworks with Era4, Verda and others all have expansions tied to Nvidia hardware, much of it pointed at the coming Vera Rubin platform.

Two items stand out from the list. Nscale reports that in tests on GB300 NVL72 systems, Nvidia's DSX MaxLPS power software allowed 37% more GPUs and nearly 50% higher aggregate inference throughput inside the same power budget. And Nvidia announced an OpenShell Reference Architecture, backed by a long list of cloud partners, meant to put hard limits on what AI agents can access and do.

Let's be clear about what this is: a vendor blog celebrating its customers. The sovereign framing is doing a lot of work, since nearly every piece of this European sovereignty runs on one American company's chips, networking and software. Choosing where data sits is real control. Choosing whose stack you depend on is a different question, and this post doesn't pretend to answer it.

That said, the power number matters if it holds up outside Nscale's tests. Europe's AI buildout is constrained by electricity more than by ambition, and squeezing materially more inference out of the same grid connection is the kind of unglamorous gain that actually changes how many projects pencil out. We'd want independent confirmation, but it's the most consequential figure in the post.

The agent-safety piece is the one I'm watching most. The OpenShell architecture splits duties across hardware: Vera Rubin for inference, Vera CPUs running OpenShell for agent tool use, and BlueField-4 DPUs running Nvidia Sentry for enforcement in silicon. The idea of enforcing agent boundaries below the software layer, where a jailbroken prompt can't reach, is genuinely good. It's also a design that ties safety to buying more Nvidia parts, and that tension will shape who can afford well-contained agents.

The smaller proof points are encouraging for open models. Baseten and CodeRabbit fine-tuned Nemotron 3.5 Lightning for code-review routing and moved routing agreement from 75.8% to 80.7% while cutting estimated inference costs in half for that workload. Mistral says its Large 4 model was trained on 3,800 Grace Blackwell GPUs in its own European data centers. That's the useful version of sovereignty: local builders shipping capable models, not just local landlords for foreign compute.

Vera Rubin is Nvidia's next platform; several partners, including Nscale, Verda and Mistral, describe plans to deploy it, with Mistral's delivery planned for 2027. Nebius has adopted Nvidia Groq 3 LPX hardware alongside Vera Rubin NVL72 for low-latency inference and runs a physical AI workbench used by robotics firms including Rhoda AI, RoboForce and Sereact.