OpenNebula Systems and Axelera AI are collaborating to develop infrastructure for European sovereign AI, combining open-source cloud orchestration with specialized AI inference hardware. The initiative targets enterprise AI environments, telecommunications infrastructure and large-scale AI Factories, where organizations need control over how AI workloads are deployed, managed and scaled. Building on an existing integration of Axelera’s Metis AI Processing Unit (AIPU), the companies plan to validate additional hardware, evaluate Europa accelerators for generative AI and large language models, and explore integration with vLLM and enterprise model-serving frameworks.
OpenNebula and Axelera AI Target European Sovereign AI
European organizations are looking for ways to expand artificial intelligence infrastructure without becoming dependent on a single cloud provider, hardware vendor or software ecosystem. That demand is putting renewed attention on sovereign AI: infrastructure designed to give organizations greater control over where AI workloads run, how data is handled and which technologies underpin their systems.
OpenNebula Systems and Axelera AI are pursuing that opportunity through a collaboration that combines cloud and edge orchestration with dedicated AI inference accelerators. The companies plan to develop and validate infrastructure configurations for enterprise AI, telecommunications, AI Factories and larger AI Gigafactory deployments.
Rather than announcing a finished product, the partnership builds on technical integration work already completed and outlines further testing, certification and software compatibility efforts.
Bringing AI Accelerators Into Virtualized Infrastructure
OpenNebula develops an open-source, vendor-neutral cloud and virtualization platform used to manage infrastructure across data centers, edge environments and hybrid deployments. Axelera AI develops specialized hardware and software for accelerating AI inference, the process through which trained models generate predictions or responses.
The companies have already integrated Axelera’s Metis AI Processing Unit (AIPU) into OpenNebula’s orchestration platform using virtualization and PCI passthrough. This allows AI accelerators to be assigned to virtualized workloads rather than requiring every application to run directly on a dedicated physical server.
PCI passthrough enables a virtual machine to access a physical hardware device more directly. In AI infrastructure, this can make specialized accelerators available to workloads running in virtualized environments, although performance and isolation depend on the hardware, drivers, hypervisor configuration and deployment architecture.
The integration is intended to let infrastructure operators provision accelerator resources on demand. For organizations running multiple AI applications, that could help bring AI inference into existing cloud management and resource-allocation workflows.
The approach is particularly relevant to enterprises that want to combine AI acceleration with the flexibility of virtualized infrastructure. It may also appeal to service providers seeking to offer accelerator-backed compute resources to multiple customers.
Europa Brings Generative AI Into Focus
The next phase of the collaboration will extend beyond the existing Metis integration. OpenNebula Systems will participate in Axelera AI’s Europa Early Access Program, with the companies planning further technical validation of Axelera hardware and software within the OpenNebula environment.
The work includes certifying Metis accelerators with PCI passthrough and the Kernel-based Virtual Machine (KVM) hypervisor. The partners will also evaluate Axelera’s newer Europa products for generative AI and large language model workloads.
This matters because inference infrastructure is becoming more varied. AI deployments increasingly span different model sizes, performance requirements, power constraints and locations. Specialized accelerators can provide an alternative to relying exclusively on general-purpose processors or GPUs, depending on the workload and supported software.
However, accelerator suitability cannot be determined from hardware specifications alone. LLM serving also depends on model compatibility, memory capacity, batching, latency, throughput and software optimization. The announced evaluation should help establish how Europa products fit into OpenNebula-managed environments, but the partnership has not yet published comparative benchmarks or detailed performance results.
The companies also plan to explore integration with vLLM, an open-source framework for serving large language models, and other enterprise LLM serving systems. Compatibility with established serving software could make it easier for infrastructure operators to incorporate supported accelerators into existing AI application architectures.
Sovereignty, Vendor Choice and AI Factories
The partnership’s broader objective is to support European sovereign AI infrastructure. Sovereignty in this context concerns more than the physical location of a data center. It can also involve control over infrastructure administration, data access, software dependencies, deployment choices and the ability to move workloads between environments.
OpenNebula’s open-source, vendor-neutral approach is intended to give operators flexibility in how they manage infrastructure. Axelera contributes an AI acceleration stack designed for inference workloads. Together, the companies aim to create configurations that combine those capabilities without requiring customers to adopt a single vendor’s entire infrastructure ecosystem.
The companies will also explore joint reference architectures for AI Factories and AI Gigafactories. AI Factories generally refer to integrated environments designed to support the development, deployment and operation of AI workloads at scale. Gigafactory deployments extend that ambition toward larger infrastructure footprints and greater compute capacity.
Potential deployment settings include enterprise data centers, telecommunications infrastructure and neocloud providers that offer specialized compute services. These environments have different requirements for workload scheduling, network connectivity, hardware utilization and operational management, so reference architectures could help define repeatable deployment patterns.
What Enterprises Should Watch
The partnership comes as European organizations assess how to scale AI while managing infrastructure costs, energy consumption and reliance on external providers. Specialized inference hardware may offer advantages for selected workloads, but those advantages need to be measured against application requirements and the operational cost of deploying a different accelerator ecosystem.
The key milestones will be practical rather than promotional: hardware certification, stable virtualization support, model-serving compatibility and documented performance across representative workloads. The companies have not announced a joint commercial product launch date or released benchmark data demonstrating how their combined infrastructure compares with alternative AI systems.
For enterprise technology teams, the collaboration offers a potential route to integrating specialized AI inference hardware into an open-source cloud environment. For European infrastructure providers, it represents an effort to broaden the range of hardware and orchestration choices available for sovereign AI deployments.
The outcome will depend on how well the components work together in production. If the planned validation delivers reliable integrations and repeatable architectures, OpenNebula and Axelera AI could make it easier for organizations to deploy AI inference across virtualized infrastructure while retaining greater control over their technology choices.
Market Landscape
The AI infrastructure market is diversifying beyond general-purpose GPU clusters toward heterogeneous compute environments that combine CPUs, GPUs and specialized inference accelerators.
- Sovereign AI infrastructure: European organizations are exploring deployment models that offer greater control over data, infrastructure administration and technology dependencies.
- AI accelerators: Specialized inference chips compete on workload suitability, energy efficiency, throughput and cost, rather than replacing GPUs across every AI application.
- Open-source cloud orchestration: Platforms such as OpenNebula give operators options for managing virtualized infrastructure across private clouds, edge environments and data centers.
- LLM serving frameworks: Compatibility with tools such as vLLM can help connect model-serving software with different hardware backends.
- AI Factories: Integrated infrastructure architectures are emerging as a way to coordinate compute, software and operations for enterprise AI deployments.
The partnership’s differentiator is its combination of open-source orchestration and European AI acceleration technology. Its commercial relevance will depend on certification results, supported workloads, software maturity and independently measured performance.
Top Insights
- OpenNebula Systems and Axelera AI are collaborating on European sovereign AI infrastructure, combining cloud orchestration with specialized AI inference accelerators.
- Their existing Metis AIPU integration uses PCI passthrough to make accelerator resources available to workloads running in virtualized infrastructure.
- Planned work includes Metis certification with KVM, evaluation of Europa hardware for generative AI and LLM workloads, and exploration of vLLM integration.
- The companies aim to develop reference architectures for AI Factories, AI Gigafactories, enterprise data centers and telecommunications infrastructure.
- No comparative performance benchmarks or commercial launch timeline have been announced, making technical validation a key milestone for prospective customers.
Power Tomorrow’s Intelligence — Build It with TechEdgeAI
