As companies increasingly turn to AI for competitive advantage, having a cloud infrastructure that can reliably support GPU-intensive workloads becomes critical. The AI Factory Deployment blueprints from OpenNebula give organizations a clear, reproducible path to build scalable, high-performance AI infrastructure, whether on-premises or in the cloud. This post walks through what the blueprints deliver, how they work, and why it’s relevant for enterprises building AI platforms.

It’s the first in a series on building an AI Factory with OpenNebula—stay tuned for more posts and practical guides in the coming weeks.

What Are the “AI-Factory Deployment Blueprints”?

• They are comprehensive guides to deploying a multi-tenant OpenNebula cloud optimized for AI workloads (training and inference).
• The blueprints include hardware and architecture recommendations, automated deployment and validation workflows to ensure performance and correctness.
• They are designed for multiple deployment scenarios: from on-premises data centers to cloud-based bare-metal servers.

As such, they enable organizations to build their own “AI Factories”: cloud environments tailored to support GPU-powered AI workloads with efficiency, control, and scalability. PCI passthrough is used so that virtual machines get direct access to physical GPUs, which is critical for ML training or inference workloads.

Core Components and Deployment: What the Blueprints Include

Deployment Automation with Tooling

• The blueprints leverage an automated toolset, OneDeploy, to streamline deployment and configuration. This reduces manual steps and guarantees consistency across environments.
• OneDeploy allows defining the cloud topology (frontend, compute nodes, GPU-enabled nodes) via an inventory file, specifying which nodes get GPU passthrough, networking settings, storage, virtual network configuration, etc.
• This makes deploying a full AI-ready cloud simpler and repeatable—whether for small proofs-of-concept or full-scale enterprise infrastructure.

Flexible Deployment Options: On-premises or Cloud/Bare-Metal

• If you have your own infrastructure, you can deploy on-premises with standard hypervisor hosts. If not, you can use public cloud providers’ bare-metal servers offering GPU capabilities. For instance, the blueprint details all the required steps to use Scaleway Elastic Metal servers for this purpose.
• This flexibility lets organizations adopt AI-ready infrastructure without committing to a fixed model, ideal for hybrid cloud, edge, or private cloud strategies.

Validation and Readiness Checks

• Once deployed, the AI-Ready Cloud can be optionally validated to confirm everything works as expected. Validation paths include: LLM inference benchmarking with vLLM, or deploying a GPU-ready Kubernetes cluster and validating with GPU workloads.
• For Kubernetes-based AI workloads, the blueprint supports integration with GPU orchestration and scheduling frameworks (NVIDIA Dynamo or NVIDIA KAI Scheduler) on top of the AI-ready cloud—enabling flexible, multi-tenant AI workloads.

Why It Matters: Benefits of Building Your Own AI-Ready Cloud

Using the AI-Ready Blueprint with OpenNebula gives organizations:

• Performance and GPU efficiency: direct GPU passthrough ensures workloads run with native GPU performance.
• Cost control and flexibility: by deploying on-premises or on bare-metal cloud, you avoid vendor lock-in and pay only for what you need.
• Sovereignty and data privacy: full control over data and infrastructure, which matters for compliance-sensitive sectors.
• Scalability and multi-tenant support: build private or hybrid clouds that support multiple teams or clients, with full isolation and GPU sharing where needed.
• Repeatability and automation: automated deployments via OneDeploy reduce manual overhead and minimize configuration drift.
• Adaptability: whether for research, production, edge AI, or hybrid workloads—multiple deployment scenarios are supported.

This makes this deployment blueprint highly attractive for enterprises, telcos, research centers, SaaS providers, or anyone building AI infrastructure.

How to Get Started: Deployment Workflow 

1. Ensure your hypervisor hosts support IOMMU (VT-d / AMD-Vi) and configure the kernel accordingly, following the described guidelines below.
2. Clone the OneDeploy tool and prepare an inventory file defining your architecture—frontend, compute nodes, GPU-enabled nodes with PCI passthrough, virtual networks, datastores, etc.
3. Run the OneDeploy playbook (make I=<inventory file>) to provision your OpenNebula cloud.
4. After deployment, validate the setup: check that GPUs are correctly listed in the cloud management UI (Sunstone), or optionally deploy a GPU-ready Kubernetes cluster and run validation tests.
5. If desired—integrate GPU orchestration frameworks (like NVIDIA Dynamo or KAI) to manage AI workloads in multi-tenant or containerized environments.

With these steps you can go from a clean server (or bare-metal cloud instance) to a fully operational, GPU-ready AI cloud. You can find more details in the guides mentioned below, including step by step deployments for easy reproducibility.

A Blueprint for Next-Gen AI Infrastructure

As AI workloads become more demanding and more ubiquitous, the days of generic VMs and standard clouds are numbered. What enterprises need now are clouds built for AI infrastructure that deliver performance, flexibility, cost-efficiency, and sovereignty.

The AI-Ready Cloud blueprint from OpenNebula offers exactly that: a proven, automation-friendly, scalable path to deploy GPU-native, multi-tenant clouds—whether on-premises, at the edge, or in bare-metal public clouds.