Dutch AI cloud company Nebius has released version 3.5 of its AI Cloud platform, adding serverless AI computing to simplify operations for developers building AI applications.
The new serverless features allow users to start workloads immediately, as the platform automatically manages infrastructure setup and runtime processes.
Go deeper with GlobalData
Access deeper industry intelligence
Experience unmatched clarity with a single platform that combines unique data, AI, and human expertise.
This change aims to reduce the time required for development teams to experiment with, train, and deploy AI models.
The update also adds the NVIDIA RTX PRO 6000 Blackwell Server Edition GPU to Nebius’ portfolio, making it available for demanding workloads such as AI inference, robotics, simulation, visual computing, and drug discovery.
Nebius has introduced a Data Transfer Service in this release, which streamlines the process of moving and replicating data between external S3-compatible storage systems and Nebius cloud regions.
Platform improvements include changes to Managed Soperator, Nebius’s Slurm-on-Kubernetes tool, giving users increased flexibility and control over cluster configuration.
Observability within Managed Kubernetes has also been enhanced to provide greater cluster oversight.
The AI application marketplace within the platform has been redesigned for improved access to tools and models relevant to development workflows.
Updates in version 3.5 include expanded user administration capabilities with more granular role-based permissions and new public APIs that enable teams to export billing data more efficiently.
All enhancements in the Aether 3.5 update are now active on Nebius’s global cloud infrastructure.
The serverless service is available in public preview, and the NVIDIA RTX PRO 6000 Blackwell Server Edition GPU is now available.
Recently, Nebius Group signed a multi-year agreement with Meta to provide AI infrastructure, with the contract valued at up to $27bn.
Under the deal, Nebius will supply $12bn in dedicated capacity at multiple sites, using Nvidia’s Vera Rubin platform as part of an early large-scale rollout. The company expects to start delivering this capacity in early 2027.
