For a research-heritage site like this one, OpenNebula is not a neutral subject. The platform’s lineage runs back through the same grid-computing and resource-scheduling research that produced GridWay and the early distributed-systems work at the Distributed Systems Architecture group — the intellectual ancestry of much of what later became open cloud management. So it is worth being explicit that this is a topic we have reason to follow closely, and reading the latest release through that lens rather than as a vendor announcement.
That release, landing in March 2026, is significant beyond the usual incremental-feature cadence. It deepens OpenNebula’s integration with NVIDIA GPU infrastructure and adds confidential-computing capabilities — two features that together address the specific gap that has kept open private clouds out of serious AI deployment conversations. The question this raises is whether an open, self-hosted cloud-management platform can now be a credible home for AI workloads, rather than a place to run general-purpose VMs while the GPU-heavy work goes to a hyperscaler.
Why GPUs Were the Missing Piece
Open private-cloud platforms have been production-viable for general workloads for years. OpenNebula’s particular strength has always been operational simplicity — a lighter-weight platform than the full OpenStack service catalogue, well-suited to edge, distributed multi-site, and organizations that want cloud management without the operational weight of a large cloud-native service stack.
What kept these platforms on the sidelines of the AI buildout was GPU handling. Running AI workloads well requires more than attaching a GPU to a virtual machine. It requires scheduling that understands GPU topology, the ability to partition a physical accelerator among multiple workloads (NVIDIA’s Multi-Instance GPU capability), passthrough and virtualization modes that match the workload, and an operational model for the firmware, drivers, and monitoring that GPU fleets demand. Hyperscalers built all of this into their managed services. Open private clouds had partial, often manual, GPU support — enough for experiments, not enough for a production AI estate.
The 2026 release is OpenNebula’s move to close that gap: first-class GPU scheduling, support for GPU partitioning and the relevant virtualization modes, and the management plumbing to operate accelerators as a proper resource class rather than a special-case device bolted onto a VM. The practical effect is that an organization can build a GPU cluster on its own hardware and manage it through OpenNebula with the kind of scheduling and operational support that previously required either a hyperscaler or a substantial in-house engineering investment.
The Confidential-Computing Angle
The second headline feature — confidential computing — matters precisely because of why organizations want private AI infrastructure in the first place. Confidential computing uses hardware-based trusted execution environments to keep data and code encrypted and protected even while in use, not just at rest and in transit. On modern GPUs and CPUs, this means an AI workload can run with its model weights and input data shielded from the underlying infrastructure operator.
For a private cloud this closes a subtle but important loop. The reason many organizations want AI infrastructure off the hyperscalers is to keep sensitive training data and proprietary models under their own control. But a private cloud still has operators — administrators with privileged access to the hosts. Confidential computing extends the protection so that even those operators cannot read the protected workload’s memory. For regulated-sector deployments and for organizations protecting valuable proprietary models, this is the difference between “we control where the data lives” and “we control where the data lives and even our own infrastructure team cannot see it.” It is a meaningful strengthening of the sovereignty argument.
What This Means for the VMware Migration Wave
The timing is not coincidental. The infrastructure market of 2024 and 2025 was reshaped by organizations re-evaluating their private-cloud strategy after the changes to VMware licensing under Broadcom. Many of those organizations went looking for open alternatives, and OpenNebula was one of the platforms in that conversation — alongside OpenStack, Proxmox, and others.
The AI dimension adds urgency to that migration calculus. An organization rebuilding its private cloud after a VMware re-evaluation now has to ask not just “where do my VMs go?” but “where do my AI workloads go?” A platform that handles both — general workloads and GPU-accelerated AI, with confidential computing for the sensitive parts — lets that organization consolidate on a single open stack rather than splitting AI off to a separate environment or a hyperscaler. For teams already inclined toward open infrastructure for sovereignty or cost reasons, removing the GPU gap removes one of the last reasons to keep a foot in the proprietary or hyperscaler camp for AI specifically.
The Honest Caveats
A capable feature set is necessary but not sufficient, and it is worth being clear-eyed about what a release does and does not change.
Operating GPU infrastructure is hard regardless of platform. The platform can schedule and partition GPUs, but the organization still has to source the accelerators, provision the power and cooling for high-density racks, manage drivers and firmware, and build the operational competence to run a GPU fleet. None of that is delivered by a software release. The power and cooling demands that constrained cloud capacity through 2025 apply just as much to a private GPU cluster as to a hyperscaler’s.
Managed services still carry their own value. Hyperscaler AI platforms bundle not just hardware but managed model-serving, autoscaling, and integration that an open private cloud leaves to the organization to assemble. The trade is the familiar one: control and sovereignty versus convenience and offloaded operations. The 2026 release improves the control side of that trade; it does not eliminate the convenience advantage of managed services.
Ecosystem maturity takes time to prove. First-class GPU and confidential-computing support is new, and production confidence accumulates through deployment experience rather than release notes. Early adopters will surface the rough edges that broader adoption later smooths.
The right framing is not that open private clouds have caught up with hyperscalers for AI — that overstates it. It is that the specific technical gap that excluded them from the conversation has been substantially narrowed, which means the decision is now governed by the usual control-versus-convenience and cost-versus-operations trade-offs rather than by a hard capability ceiling.
Conclusion
OpenNebula’s 2026 release is a notable marker in the maturation of open infrastructure for AI. By bringing first-class GPU scheduling and confidential computing into an open, self-hostable cloud-management platform, it removes the capability gap that had kept open private clouds on the sidelines of the AI buildout. For organizations that already want infrastructure sovereignty — driven by the VMware migration wave, by data-residency requirements, or by the wish to keep proprietary models under their own control — the release means AI workloads no longer have to be the exception that sends them back to a hyperscaler. The hard parts that remain are operational and physical, not architectural, and those are the same hard parts every GPU operator faces. Given this platform’s research lineage, it is a development worth watching closely.
Further Reading
- OpenNebula Documentation — the platform’s official documentation, including GPU support, virtualization modes, and deployment guidance.
- Confidential Computing Consortium — the Linux Foundation project defining standards and reference implementations for trusted-execution-environment-based computing.