Nvidia & OpenNebula deepen integration for AI factories
Nvidia and OpenNebula Systems have expanded their technology integration around what they call "AI Factories", linking OpenNebula's cloud management and virtualisation software with Nvidia's accelerated computing, networking, and server provisioning tools.
The expanded work focuses on automated deployment and tenant isolation for organisations running AI infrastructure across enterprise IT, high-performance computing environments, and so-called sovereign AI programmes. OpenNebula is also targeting neocloud providers that sell GPU capacity and AI services to third parties.
OpenNebula develops an open-source cloud management and virtualisation platform for private, hybrid, and edge infrastructure. It has also positioned itself as an alternative to VMware for organisations that want to keep existing operational practices and hardware estates.
GPU provisioning
A key update is support for Nvidia GB200 NVL4 GPUs through PCI passthrough. This gives a virtual machine direct access to a physical GPU, rather than relying on a shared virtual device layer. OpenNebula says the approach preserves native performance for workloads that need low overhead.
The integration also includes Nvidia's Multi-Instance GPU (MIG) feature, which splits a GPU into hardware-level partitions that can be assigned to different users or tenants. This structure is designed to reduce the risks of multiple tenants sharing accelerators while keeping resource allocation predictable.
GPU partitioning and passthrough have become more important as organisations try to balance utilisation with governance. Many want to share expensive accelerator infrastructure across internal teams or external customers, while ensuring clear separation for sensitive data or regulated projects.
Network offload
Another element is management of Nvidia BlueField data processing units (DPUs) within OpenNebula's control plane. BlueField DPUs can offload some network and security functions from host CPUs. The integration covers per-tenant traffic enforcement, containerised network functions, and programmable switching policies, according to OpenNebula.
Offloading to DPUs can change how operators size and run clusters by keeping CPU headroom available for virtual machines, storage services, or control-plane processes. Vendors also position DPUs as a way to tighten security boundaries between infrastructure and tenant workloads, particularly in multi-tenant environments.
The companies also point to Spectrum-X Ethernet networking integration validated on Nvidia BlueField DSX Air, which OpenNebula plans to demonstrate. Spectrum-X is Nvidia's Ethernet networking portfolio for AI data centres. BlueField DSX Air is a platform used for validation and development work around data centre networking.
Automation stack
The expanded work also brings in Nvidia NCX Infra Controller for provisioning and lifecycle operations. The software automates server provisioning, operating system deployment, and environment configuration for GPU servers. OpenNebula says this enables isolated AI Factory instances on demand through a unified workflow.
Automation and governance have become focal points as more organisations move AI workloads into production. Many early AI projects relied on manually configured clusters or specialised lab environments. Production use requires repeatable processes for hardware provisioning, isolation, and upgrades.
For cloud and service providers, the pitch is consistent operations across tenants and sites. For enterprise and public-sector buyers, it is tighter control over who can access compute and network resources, and better tracking of infrastructure changes.
"AI factories require tight integration across compute, networking, and infrastructure automation," said Warren Barkely, VP of DGX Cloud at Nvidia. "OpenNebula's work with NVIDIA technologies, including NCX Infra Controller, helps organizations build scalable, multi-tenant AI environments with greater automation and operational control."
OpenNebula frames the goal as a single control plane spanning compute, networking, and bare-metal lifecycle management. The theme reflects a broader shift towards integrated AI infrastructure stacks, where procurement decisions cover GPUs, networking, storage, and management software as a single design.
"Our focus is to provide a unified control plane for AI infrastructure that preserves hardware performance while introducing strong governance and lifecycle automation," said Ignacio M. Llorente, managing director at OpenNebula Systems. "Through deep integration with NVIDIA technologies, we enable organizations to deploy AI Factories that meet enterprise, HPC, and sovereign requirements with predictable performance and operational control."
GTC demonstrations
OpenNebula will present demonstrations at Nvidia GTC, including multi-tenancy for GPU and AI-as-a-Service workloads, DPU-based network offload, Spectrum-X Ethernet networking integration, and end-to-end deployment using NCX Infra Controller. The demonstrations are expected to show deployment from bare metal through to operational multi-tenant AI cloud services.