Proxmox vs OpenStack: An Honest 2026 Comparison of Proxmox VE 9 and OpenStack Hibiscus

Proxmox VE is a virtualization platform you install and run. OpenStack is a cloud you build and then operate. That is most of the Proxmox vs OpenStack question: if the people using the virtual machines are your own colleagues and you run up to a few dozen hosts in one site, Proxmox VE 9 does the job with a fraction of the staff and effort. OpenStack earns its complexity when strangers need self-service, with their own projects, networks, quotas and API keys, across hundreds of hosts or several sites. Most teams that start out asking about OpenStack vs Proxmox turn out to need Proxmox.

Where we come down, before the detail:

• Current versions in October 2026: Proxmox VE 9.2 (May 2026, Debian 13) and OpenStack 2026.2 Hibiscus (30 September 2026). Proxmox VE 8 reached end of life in August 2026.
• Proxmox has permissions and resource pools but no quotas; OpenStack has projects with quotas. That gap settles most hosting-provider cases.
• Both run KVM and QEMU and both do Ceph, so VM speed depends on your storage and network design more than the platform.
• Proxmox restarts a dead host's HA-managed VMs by itself in about two minutes. Plain OpenStack waits for an operator unless you add Masakari.

What are Proxmox VE and OpenStack in 2026?

Proxmox VE is a Debian-based distribution that turns a few servers into a virtualization cluster: KVM virtual machines, LXC containers, storage, networking, backups and a web interface in one install. Proxmox VE 9.2 came out on 21 May 2026 on Debian 13.5 with kernel 7.0, QEMU 11.0 and Ceph Tentacle 20.2 as the default, according to the Proxmox VE roadmap, and an ARM64 build followed on 5 August 2026, the first for anything but x86-64. For clusters still on 8.x, the in-place upgrade to 9 starts with the pve8to9 --full checklist, and hyper-converged Ceph must be on Squid 19.2 first.

OpenStack is a set of services that together behave like a public cloud you own: Nova for compute, Neutron for networking, Cinder for block storage, Glance for images, Keystone for identity and Horizon or Skyline as the dashboard, with Octavia, Designate, Heat, Ironic and more on top. A release lands every six months. 2026.2 Hibiscus, the 34th, shipped on 30 September 2026 with about 600 contributors and 11,500 changes, according to OpenInfra Foundation figures reported by Network World. The OpenStack release series page marks every ".1" release, such as 2026.1 Gazpacho, as SLURP: you can jump from one SLURP to the next and skip the ".2", so one upgrade a year is a valid plan.

Which should you pick? Proxmox vs OpenStack by team size and use case

The OpenStack vs Proxmox decision table below is how we would answer for a specific team. "Ops staff" means the people who get the call when the platform breaks.

Team size matters more than host count. An OpenStack cloud people depend on needs at least two people who can read Nova and Neutron logs at night. A Proxmox cluster can be run by one Linux generalist with a colleague who knows the basics.

OpenStack vs Proxmox architecture: one Debian install against a dozen services

Every Proxmox VE node is the same: same packages, same web interface on port 8006, same REST API. Nodes join a cluster over Corosync, and the configuration lives in a small replicated file system (pmxcfs) mounted at /etc/pve on every node. There is no controller to lose; any node manages the whole cluster.

OpenStack splits into a control plane and compute nodes. The control plane runs the API services and schedulers plus MariaDB (often Galera), RabbitMQ and memcached; Neutron dropped PostgreSQL in 2025.1. Compute nodes run nova-compute, libvirt and the OVN controller. Three control nodes is the usual minimum, installed with Kolla-Ansible (containers on Debian 13, Ubuntu 24.04 or Rocky Linux 10 hosts in 2026.1), OpenStack-Ansible, or a vendor product such as Canonical OpenStack or Red Hat OpenStack Services on OpenShift 18.0. That split is why OpenStack scales further and costs more to run. A Proxmox upgrade is apt on one node at a time; an OpenStack upgrade touches database schemas and a dozen services in a set order.

Does Proxmox support multi-tenancy?

Partly. Proxmox VE logs users in through PAM, LDAP, Active Directory, OpenID Connect or its own user database, supports TOTP and WebAuthn second factors, and grants roles on nodes, storages or resource pools, named groups of VMs, containers and storages. That is plenty for departments sharing a cluster.

What it lacks is consumption limits. The user management docs cover who may touch what, not how much anyone may use: no per-user or per-pool cap on vCPUs, RAM or disk. Tenant network separation is yours to design with VLANs, VNets and firewall rules. OpenStack starts from the other end: Keystone domains and projects, quotas for instances, cores, RAM, volumes and floating IPs, and private networks and routers per project behind security groups. If customers who don't know each other share your hardware, you want that model, and on Proxmox you would write or buy it.

Proxmox Datacenter Manager helps from above. Version 1.0 arrived on 4 December 2025 and 1.1 on 28 May 2026, managing many clusters from one interface with role-based access, central updates, Ceph monitoring and live migration between independent clusters. Its release notes list no quota feature, so the gap above stays.

Networking: Proxmox SDN against Neutron and OVN

The Proxmox SDN documentation describes cluster-wide zones of five types: Simple, VLAN, QinQ, VXLAN and EVPN (VXLAN with BGP as control plane). Core SDN is fully supported; IPAM with DHCP and complex FRRouting setups are still technology preview, and NetBox, phpIPAM or PowerDNS can serve as external IPAM. The 9.x series moved fast: 9.0 added routed spine-leaf fabrics with OpenFabric or OSPF, and 9.2 added WireGuard and BGP fabrics, route maps and prefix lists to filter what gets redistributed, and IPv6 underlays for EVPN. Capable, but admin-driven.

Neutron is built the other way round. Tenants create networks, routers and floating IPs through the API, and security groups follow the port wherever the VM lands. The recommended back end is OVN: Neutron removed its linuxbridge driver in 2025.1 and told operators to move to OVS "or, preferably, OVN". Gazpacho folded OVN BGP into the OVN driver, and Hibiscus added EVPN Type-5 advertisement of tenant prefixes into the fabric through FRR. Load balancing is Octavia (see our Octavia load balancer guide), and our earlier piece on SDN back ends for OpenStack compares OVN with the alternatives. If your network team runs EVPN, both can plug into it; for hundreds of self-service tenant networks, Neutron is the one designed for the job.

Storage: Ceph in both, run very differently

Ceph is the default shared storage on both; the difference is who installs it. Proxmox VE deploys Ceph itself from the GUI or pveceph, on the same nodes as the VMs. The recommended release is now Tentacle 20.2, end of life estimated for November 2027, while Squid 19.2 reached its estimated end of life in September 2026, so clusters still on Squid should schedule the move. ZFS, LVM, NFS and iSCSI sit alongside, and 9.0 added snapshots on thick-provisioned LVM for SANs as a technology preview. Our open source block storage comparison puts Ceph next to LINSTOR, Longhorn and ZFS.

OpenStack consumes storage through Cinder (volumes), Glance (images), Swift (objects) and Manila (file shares), each with drivers for Ceph and most commercial arrays. The deployment tools expect Ceph to exist already: Kolla-Ansible's external Ceph guide says it does not provision a Ceph cluster and points to cephadm or ceph-ansible. So you run two distributed systems with separate upgrade cycles. Our article on Ceph block and object storage covers that side.

High availability: who restarts the VM when a host dies?

Here Proxmox is the better tool for traditional workloads. The Proxmox HA manager fences a failed node with a watchdog and restarts its HA guests elsewhere, with typical detection and failover times of about two minutes, and it wants at least three nodes. Since 9.0, affinity rules say where guests should or must run and which must stay apart. 9.2 added a dynamic load balancer that moves guests by real-time load while respecting those rules.

When an OpenStack compute node dies, Nova's evacuate operation rebuilds its instances elsewhere, but an operator or script must trigger it. Automatic instance HA needs Masakari, which is part of Hibiscus but is one more service to deploy and test. The cloud-native answer is to skip it and design applications to survive losing a VM, as on a public cloud. For pets, such as one database VM or one ERP server, Proxmox's built-in HA is simpler and safer.

API, automation and Terraform

Proxmox's REST API is what its own web interface calls, so anything you can click you can script, with expiring, reduced-privilege API tokens or pvesh on a node. There is no official Terraform provider; the popular bpg/proxmox provider is a community project, not affiliated with Proxmox, that targets 9.2. OpenStack is an API first and a product second: every service has a versioned API, one openstack client covers them all, Heat orchestrates stacks, and Terraform, Ansible and Kubernetes tooling have long-standing OpenStack integrations. Teams that think in projects, security groups and volumes find it natural.

Licensing and support model

Proxmox VE comes from one company, Proxmox Server Solutions in Vienna, and every feature is free under the AGPLv3. A subscription adds support and the Enterprise repository, which Proxmox recommends for production; the free one is meant for testing and non-production use. It is counted per occupied CPU socket, not per core, and all nodes in a cluster need the same level. Above Community (repository only, no tickets) sit Basic, Standard and Premium, with a first response to critical issues within one business day, four hours or two hours; Standard and Premium also include remote help over SSH, according to the Proxmox VE subscription page. Proxmox announced in September 2026 that enterprise support goes 24/7 on 19 October 2026: Premium from that day, Standard during a fourth-quarter onboarding window, Basic unchanged. It also opened Proxmox North America Inc. in Kingston, Ontario.

OpenStack has no single vendor. Its projects must be Apache 2.0, the OpenInfra Foundation behind them joined the Linux Foundation in June 2025, and support comes from whoever packages it for you: Canonical, Red Hat, an integrator or your own team. Nobody is responsible for how the pieces fit unless you pay a distribution to be.

Operational effort: the cost nobody writes down

• Upgrades: Proxmox follows Debian, a major release about every two years, each supported at least as long as Debian's security support, roughly three years, which is why Proxmox VE 8 reached end of life in August 2026. OpenStack ships twice a year, and even with SLURP the yearly upgrade is a project that wants a test cloud.
• Skills: Proxmox needs Linux, networking and, with Ceph, storage. OpenStack adds Keystone, Nova scheduling, OVN, RabbitMQ and Galera.
• Failure surface: a Proxmox problem is usually one node or one storage. In OpenStack, a message queue backlog can make every service look broken at once.
• Footprint: a useful Proxmox cluster is three servers. In our view production OpenStack starts at six to eight once you count three controllers, compute and storage.

Our rule of thumb: two people who know OpenStack well before anything important runs on it, and months, not weeks, to the first production cloud. Without that, Proxmox is the honest answer even when OpenStack fits the feature list better.

How many nodes can a Proxmox cluster have, and how far does OpenStack go?

Many comparisons repeat a hard limit of 32 nodes. The Proxmox cluster manager documentation sets no explicit limit and mentions clusters of over 50 nodes in production as of 2021. The real limit is Corosync: it wants under 5 ms of latency between all nodes, and above about 10 ms larger clusters become unlikely to work. So a Proxmox cluster lives in one site, and growth means more clusters under Datacenter Manager.

OpenStack was designed for the other end: Nova cells split a region, regions split a cloud by site, one Keystone serves them all. The OpenInfra Foundation's 2025 annual report counts 55 million cores in production, and its Flamingo announcement named Walmart, Workday and CERN among users with a million cores or more.

How KVM fits into both (and OpenStack vs KVM)

KVM is the Linux kernel module that lets QEMU use the CPU's virtualization extensions. It is the hypervisor; Proxmox VE and OpenStack are management layers on top. So "OpenStack vs KVM" is a slightly odd comparison: OpenStack doesn't replace KVM, it orchestrates it.

Proxmox drives QEMU directly through its own qemu-server stack and the qm command, without libvirt, and its VM chapter treats QEMU and KVM as one thing. Nova uses drivers, and the Nova feature support matrix lists libvirt KVM on x86, aarch64, ppc64 and s390x next to VMware vCenter, z/VM and Ironic for bare metal; Hyper-V is not on it. Because both end at the same QEMU and KVM, performance comes down to CPU model, disk bus and cache mode, IOThreads, and local NVMe versus Ceph over the network. Both keep adding knobs: Gazpacho made an IOThread the default for QEMU instances, and Proxmox VE 9.2 manages custom CPU models in the GUI.

Can you run OpenStack on Proxmox?

Yes, and for learning it is the cheapest multi-node OpenStack lab. Enable nested virtualization on the host (the nested option of kvm-intel or kvm-amd), give the VMs the host CPU type, and run Kolla-Ansible or DevStack inside. Proxmox VE 9.1 added a vCPU flag for finer nested control. The Proxmox wiki warns that VMs with nesting active cannot be live-migrated, and that without hardware-assisted nesting, guests run ten times slower or worse. For production, keep compute nodes on bare metal; controllers as Proxmox VMs are defensible but add a layer to debug at the worst moment. The reverse exists too: since Apache CloudStack 4.21, CloudStack can orchestrate Proxmox hosts.

Proxmox vs CloudStack, OpenNebula and OpenVZ

Proxmox vs Apache CloudStack

CloudStack is the middle path: multi-tenant IaaS with domains, accounts, resource limits and self-service networks, run by one Java management server instead of a dozen services, on KVM, VMware or XCP-ng. Release 4.21.0.0 (28 August 2025) added an Extensions Framework with built-in Proxmox and Hyper-V extensions, and 4.22 LTS (November 2025) added Proxmox console access. The current LTS is 4.22.1.1, with 18 months of fixes per LTS, and 4.23.0.0 arrived on 2 September 2026. If you like Proxmox but need tenants, test CloudStack before committing to OpenStack.

OpenNebula vs Proxmox

OpenNebula courts the same VMware leavers but thinks like a cloud: users, groups, virtual data centres and quotas over KVM and LXC hosts. The current release is 7.4 "Helix", released in July 2026, with a redesigned Sunstone web interface and VLAN authorization at group level, so tenants can provision their own networks. Version 7.2, announced on 7 April 2026, had added a gRPC API, confidential computing and vTPM for KVM, and live storage migration between datastore types. The code is Apache 2.0, with an Enterprise Edition sold by subscription. It is lighter than OpenStack, more tenant-aware than Proxmox, and has a smaller ecosystem than either.

OpenVZ vs Proxmox

Proxmox VE 4.x replaced OpenVZ with LXC because OpenVZ was not available for kernels newer than 2.6.32, while LXC is in every mainline kernel. The OpenVZ wiki still lists OpenVZ 7, from July 2016, as its latest release, so "OpenVZ vs Proxmox" today is really an old kernel against a current one.

Where RS Computers fits

RS Computers runs KVM virtual servers: VPS plans on a 1 Gb/s port and VDS plans on a 10 Gb/s port, with NVMe storage, your own IPv4 and IPv6 address and free weekly backups, in Amsterdam (Netherlands), Prishtina (Kosovo) and Dublin (Ireland). The VPS and VDS plans page has the same plans and prices in all three cities. VPS plans use shared vCPUs and VDS plans dedicated vCPUs, which suits the services around a cluster better than the cluster itself:

• The third vote for a two-node Proxmox cluster, set up below.
• A deployment host for Ansible or Kolla-Ansible reaching your nodes over a VPN, or NetBox for SDN IPAM.
• Monitoring from outside the cluster's site, and an off-site copy of your backups, as in our guide to append-only off-site backups with restic.
• A simpler answer when a private cloud is overkill: a team needing six VMs may do better with a few VDS plans in two cities than a cluster it has to patch.

A VPS as the QDevice for a two-node Proxmox cluster

Two nodes can't keep a majority when one fails. The Proxmox docs support an external QDevice for clusters with an even node count, and since it talks to the cluster over TCP, it can sit outside the cluster's LAN without the 5 ms rule. The smallest plan is plenty. On a Debian or Ubuntu VPS:

# as root on the VPS
apt update
apt install -y corosync-qnetd

TCP port 5403 on the VPS must be reachable from your nodes (allow only their addresses if you run a firewall), and root needs key-based SSH login on the VPS, because setup copies the cluster's key over SSH. Then:

# as root on every Proxmox VE node
apt install -y corosync-qdevice

# as root on one Proxmox VE node
pvecm qdevice setup YOUR_VPS_IP
pvecm status

Setup ends with Done, and pvecm status should show Expected votes: 3 and the flags Quorate Qdevice. If the QDevice shows NA (not alive), check that port 5403 is reachable.

Frequently asked questions

Is Proxmox better than OpenStack?

For most teams with fewer than about ten infrastructure engineers, yes: Proxmox VE gives clustering, HA, Ceph and backups with far less to run. OpenStack is better when you need multi-tenant self-service with quotas, a cloud-style API for many users, or more scale than one low-latency cluster can hold.

Can Proxmox replace OpenStack?

For internal virtualization, often yes. Not where customers create their own networks and need enforced quotas, because Proxmox VE has permissions and resource pools but no quotas; Apache CloudStack and OpenNebula are the lighter alternatives there.

Does OpenStack use KVM?

Yes, by default. Nova manages KVM through its libvirt driver. OpenStack is the orchestration layer and KVM the hypervisor, so it is not an either-or choice, and Proxmox VE uses KVM too, through its own QEMU management instead of libvirt.

Is OpenStack still relevant in 2026?

Yes. The OpenInfra Foundation's 2025 annual report counts 55 million cores in production, an all-time high, and 2026.2 Hibiscus shipped on 30 September 2026 with about 600 contributors, 21% more than the release before.

Is Proxmox VE free and open source?

Yes. Proxmox VE is licensed under the GNU AGPLv3 and every feature works without a subscription. A subscription, counted per occupied CPU socket, adds the Enterprise repository recommended for production and, from Basic up, support tickets with set response times.

Is OpenShift an alternative to Proxmox or OpenStack?

Not directly. OpenShift is Red Hat's Kubernetes platform, running containers and, through OpenShift Virtualization, VMs. Red Hat OpenStack Services on OpenShift 18.0 runs the OpenStack control plane as containers on OpenShift, so there the two work together.

Can RS Computers help with a Proxmox or OpenStack project?

We can host the pieces around it, such as a QDevice, a deployment or monitoring host, or off-site backups, on KVM VPS and VDS plans. Message us on Telegram or email info@rscomputers-ks.com with what you are building, and we will suggest a plan and quote the setup.

If we had to choose tomorrow

Start with Proxmox VE 9.2 unless you can name the tenants who need quotas and self-service networks, and the two people who will run OpenStack for them. If you can, deploy 2026.1 Gazpacho for the yearly SLURP rhythm, or 2026.2 Hibiscus for the newest networking; both upgrade straight to 2027.1. In between, give CloudStack a week of testing. Whatever you build, keep its quorum vote, monitoring and backups outside the cluster; one small server in another city covers all three.

← All articles