Nscale’s power test and a hardware agent-safety design stand out in Nvidia’s cloud roll call

Nvidia used GTC Berlin to showcase cloud partners across Europe, the Nordics, Africa and Central Asia adding regional AI capacity. Nscale, Crusoe, Nebius, Mistral, Fireworks with Era4, Verda and others announced expansions tied to Nvidia hardware, much of it aimed at the coming Vera Rubin platform; Mistral’s Vera Rubin delivery is planned for 2027. Nebius has adopted Nvidia Groq 3 LPX hardware alongside Vera Rubin NVL72 for low-latency inference (Groq 3 LPX) and runs a physical AI workbench used by robotics firms including Rhoda AI, RoboForce and Sereact.

Much of the post is a capacity roll call framed as sovereignty, though nearly every piece runs on one American company’s chips, networking and software. Choosing where data sits is real control; choosing whose stack to depend on is a separate question the post does not address. The same theme ran through Nvidia’s pitches to European governments (public sector open models).

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 within the same power budget. Europe’s AI build-out is constrained more by electricity than ambition, and extracting materially more inference from the same grid connection changes how many projects are viable. The figure comes from Nscale’s own tests and needs independent confirmation, but it is the most consequential number in the post (DSX and power).

Smaller proof points favor open models. Baseten and CodeRabbit fine-tuned Nemotron 3.5 Lightning for code-review routing, moving 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, a version of sovereignty built on local model developers rather than only local hosting.

Nvidia also announced an OpenShell Reference Architecture, backed by many cloud partners, to put hard limits on what AI agents can access and do. It 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. It builds on the Open Agent Safety Platform Nvidia launched in September (agent safety platform). Enforcing boundaries below the software layer, where a manipulated prompt cannot reach, is a sound idea that could make it easier to justify agents in production. It also ties safety to buying more Nvidia components, which will shape who can afford well-contained agents. More in our Nvidia coverage, AI agents coverage and cybersecurity coverage.