Ten use cases that show why the infrastructure operating model matters.

Bare-metal servers, GPU virtual machines, Kubernetes clusters, Slurm environments and inference endpoints now appear in many AI infrastructure service catalogues. At first glance, the platforms behind them can look much the same.

The differences become clearer when customers start using the AI Cloud. A platform team wants to create several environments within its allocation. A provider needs to take hosts down for maintenance. A customer requires dedicated infrastructure for production but shared resources for development. Another wants to operate across two countries.

These are ordinary cloud requirements. They concern how infrastructure is allocated, governed and operated throughout its life.

Kubernetes-based platforms bring valuable capabilities to service delivery and workload management. But for HPC centers, Neoclouds, telcos and cloud service providers building AI Clouds on AI Factory infrastructure, the question is how those capabilities fit into the wider infrastructure architecture.

OpenNebula starts with the cloud infrastructure itself. The following ten use cases explain the value of that approach, using illustrative customer scenarios.

1. Delegate a Cloud Allocation to the Tenant

Some customers want ready-to-use services. Others want their own platform team to build and operate those services within an agreed infrastructure allocation.

OpenNebula’s Virtual Data Center model defines which resources a tenant group can access. The group administrator manages permissions, and authorized users operate the cloud environments. People consuming the resulting services do not necessarily need OpenNebula accounts.

Example: A pharmaceutical company’s platform engineers create Kubernetes clusters for inference and Slurm environments for research. Scientists access notebooks, submit jobs and consume inference endpoints. The Neocloud manages the infrastructure, the company’s engineers manage the services, and researchers use them.

2. Combine Dedicated and Shared Infrastructure

A tenant’s workloads rarely have identical isolation requirements. Sensitive production systems may need dedicated hosts, while development environments can use shared capacity.

The provider can reflect those requirements in the resources available to the tenant and the policies governing their use. The customer’s team retains a common administration model across both.

Example: A bank uses dedicated GPU hosts for confidential model training and shared CPU infrastructure for development VMs. Its engineers manage both environments, while the provider determines which hosts and resources each environment can use. The bank pays for physical isolation where it needs it, without applying that cost to every workload.

3. Extend the AI Cloud across Data Centers

As providers add locations, customers need a consistent way to access them. Maintaining unrelated accounts and permissions at every site quickly becomes an operational burden.

OpenNebula’s federation model shares users, groups and permissions across cloud zones, while each zone retains its own infrastructure control plane.

Example: A Neocloud expands from Amsterdam to San Francisco. A customer’s platform team deploys services in either location according to capacity and data-location requirements, using coordinated identities and permissions. Resources remain local to each zone; cross-site migration and disaster recovery require their own design.

4. Retain Control over the Platform’s Future

Infrastructure investments can last longer than supplier contracts. For some organizations, the ability to inspect and maintain the software is part of the purchasing decision.

OpenNebula’s public source repository gives operators access to the implementation, subject to the applicable component licenses. That provides a basis for auditing, adapting and maintaining the platform. Making use of that freedom still requires technical expertise.

Example: A public-sector organization plans to operate its AI Cloud for ten years. Its engineers retain the code, documentation and operational knowledge needed to support the open source platform if its supplier arrangements change. The organization can plan for continuity beyond its next procurement cycle.

5. Develop a Distinct AI Cloud Product

For commercial operators, competitive advantage may depend on something more substantial than portal design. Reservation rules, scheduling policies, service tiers and commercial integrations can all shape the product customers buy.

OpenNebula’s source access and resource APIs allow providers to develop that behavior at the infrastructure layer. New capabilities need engineering and validation, but the provider has a route to implementing them.

Example: A Neocloud wants to offer guaranteed GPU reservations alongside cheaper, interruptible capacity. Its team develops the admission, scheduling and billing integrations required for those tiers. The service can evolve around the provider’s business model rather than relying entirely on a standard vendor catalogue.

6. Introduce Kubernetes Where It Is Needed

An AI Factory may need to support several workload environments. Some use containers. Others need conventional VMs, dedicated servers or Slurm.

OpenNebula operates the underlying cloud independently of Kubernetes. Kubernetes can therefore be introduced as one environment among several, alongside other consumption models.

Example: An HPC center already offers Slurm and GPU VMs to researchers. A new team requests Kubernetes-based inference. The center adds that environment on the same cloud foundation, without making its existing HPC services dependent on Kubernetes.

7. Extend Enterprise Cloud Infrastructure into AI

Organizations investing in AI still run databases, business applications and Windows servers. Managing those systems separately from the AI estate can duplicate infrastructure, operational processes and skills.

OpenNebula provides a common cloud foundation for enterprise virtualization and accelerated computing. An organization can therefore modernize or replace its existing VMware environment and progressively extend the same platform into an AI Cloud.

Example: A telco begins by replacing VMware for its enterprise VMs. It later deploys AI Factory infrastructure and introduces GPU VMs and Kubernetes services for AI customers. The operations team reuses the tenant model, networks, storage and many of the operational practices established during the virtualization migration. VMware replacement becomes a path to AI Cloud expansion rather than a separate infrastructure silo.

8. Manage Placement throughout the VM Lifecycle

Provisioning finds somewhere to run a VM. Operations must also deal with changing demand, resource contention and maintenance.

OpenNebula’s Distributed Resource Scheduler evaluates allocation within a cluster and produces migration recommendations that administrators can review and execute.

Example: A provider needs to service a CPU host. Its operations team moves compatible VMs to other hosts, taking capacity, affinity and storage requirements into account. VMs with GPU passthrough receive a separate maintenance plan because their device dependencies constrain live migration. The placement decision continues after deployment.

9. Allocate GPUs across Different Services

GPU workload scheduling and infrastructure allocation address different questions. A scheduler distributes resources among jobs inside an environment. The AI Cloud decides which environments receive the capacity in the first place.

Keeping those decisions separate gives operators a way to manage GPU demand across Kubernetes, VMs, Slurm and other services.

Example: A pharmaceutical tenant needs inference during working hours and training overnight. Its platform team reduces the inference environment and releases eligible GPU capacity for training VMs or Slurm workers. Run:ai and Slurm continue scheduling workloads within their respective environments. Automating this process requires suitable hardware, lifecycle policies and capacity-management integration. The objective is to reuse released capacity across services, rather than keep every environment permanently sized for its peak.

10. Reuse and Protect the Infrastructure Already in Place

An AI Cloud depends on more than GPUs. Networks, persistent storage, databases and management services all need to be operated and protected.

OpenNebula integrates supported infrastructure into cloud provisioning. Its NetApp driver, for example, manages VM disk provisioning, mapping, cloning, snapshots and backup, with documented limits on advanced array functions. It also provides VM backup operations and configurable host-failure recovery.

Example: A provider uses its existing VLAN networks and enterprise SAN for a tenant’s database VMs, Kubernetes workers and Slurm controller. With shared storage, fencing and spare capacity configured, affected VMs can restart after a host failure. Scheduled backups provide a separate recovery path for deleted or corrupted data. Applications still need their own recovery procedures. Restarting a controller does not automatically resume an interrupted training job.

Look beyond the Provisioning Screen

The useful evaluation is a complete operating cycle. Allocate resources to a tenant. Let its team create several services. Change capacity. Maintain a host. Restore data. Add another location.

At each step, establish what the platform manages directly, what depends on an integration and what requires additional work.

For OpenNebula, the value is in that infrastructure operating model: operators control the AI Cloud, tenant teams build and operate services within their allocations, and end users consume those services.

Whether the operator is an HPC center maximizing scientific and AI productivity, a Neocloud monetizing GPU capacity, a telco delivering distributed AI services, or a cloud service provider expanding its existing portfolio, the underlying requirement is similar: turn accelerated infrastructure into a cloud that can be operated reliably at scale.

That is what turns AI Factory infrastructure and a portfolio of AI services into an AI Cloud customers can depend on.

Explore how OpenNebula enables flexible, scalable AI Factories for demanding AI workloads here.