Nscale increased aggregate AI processing output by nearly 50% in tests within an unchanged power budget, NVIDIA said in a blog post published on Oct. 21 about its cloud partners.
The same tests supported 37% more graphics processing units within that power limit, the chip supplier said. NVIDIA published the findings in “NVIDIA Cloud Partners Scale Across Europe — Sovereign, Open and Built for Safe AI Agents.”
Nscale used NVIDIA GB300 NVL72 systems to test software that coordinates electricity allocation with computing workloads, the company said. The output measure covered inference, the process of running a trained AI model to generate responses.
NVIDIA did not give a testing period, baseline configuration or detailed methodology. Its account covered partner deployments, development plans and selected performance results across Europe, the Nordics, Africa and Central Asia.
Electricity supply limits the expansion of AI computing facilities, the company said. Regional cloud capacity gives developers more control over where they keep data and run applications, it said.
NVIDIA also said CodeRabbit improved a model’s agreement on code-review routing decisions to about 81% from about 76%, using a specialized model developed by Baseten. Estimated inference costs fell 50% for the evaluated workload, the company said.
That qualification limits the cost comparison to the workload tested. NVIDIA said the developers used supervised training and reinforcement learning to adapt its Nemotron 3.5 Lightning model for the task.
Nscale plans to deploy NVIDIA Vera Rubin infrastructure for advanced model training and production workloads while expanding across five European markets, the company said. It gave no deployment timetable for that plan.
Mistral plans to receive Vera Rubin infrastructure in 2027, NVIDIA said. Mistral trained its Large 4 model on 3,800 NVIDIA Grace Blackwell graphics processing units in European data centres it operates, the company said.
Crusoe is developing liquid-cooled, modular computing facilities for the Vera Rubin platform to bring cloud capacity closer to users, NVIDIA said. Fireworks and Era4 are combining model services with computing capacity at Era4’s Chesterfield facility for enterprise applications, it said.
NVIDIA said developers connecting AI models to business tools also need limits on what automated agents can access and do. Its OpenShell Reference Architecture combines software and hardware intended to enforce those limits, the company said.
Partners including CoreWeave, Crusoe, DigitalOcean, Nebius and Nscale support that architecture, NVIDIA said. The company said expanded regional capacity, model services and agent controls together support the development and operation of AI applications.
