Proxmox VE vs XCP-ng
Proxmox VE and XCP-ng compared for homelabs, small teams and VMware exits: subscription pricing per socket vs per host, free use, clustering and HA, storage (ZFS, Ceph, XOSTOR, the old 2 TiB disk limit), containers, management, backups and migration from ESXi, as of November 2026.
Quick Answer
Both are free, open-source bare-metal hypervisors that you can run in production without paying, and both sell support subscriptions. Proxmox VE 9.2 (May 2026, Debian 13, KVM plus LXC containers, AGPL-3.0) manages itself: every node serves the web UI, clusters are multi-master, and ZFS and Ceph are built in. Subscriptions are per CPU socket per year: Community €120, Basic €370, Standard €550, Premium €1,100. XCP-ng 8.3 LTS (Xen-based, supported until 30 Nov 2028) runs VMs only and is managed through Xen Orchestra, either the Vates appliance or the free build-from-source version. Vates VMS support is per host: Essential $/€2,000 a year for up to 3 hosts, Pro 1,000 per host (3+), Enterprise 1,800 per host (4+). Pick Proxmox for homelabs, mixed VM-and-container hosts and hyperconverged Ceph. Pick XCP-ng if you come from XenServer or VMware and want pool-based management, agentless backups in Xen Orchestra and per-host rather than per-socket pricing.
Last updated November 6, 2026. Covers Proxmox VE 9.2 (21 May 2026) and its arm64 build (5 Aug 2026), and XCP-ng 8.3 LTS with QCOW2 support (5 May 2026) and ISO 8.3.0-20260806, with current Proxmox per-socket and Vates VMS per-host pricing read in November 2026.
| Feature | Proxmox VE | XCP-ng |
|---|---|---|
| Licence | AGPL-3.0 | Open source (main repo GPL-2.0); Xen Orchestra AGPL-3.0 |
| Hypervisor / Base | KVM/QEMU 11.0 + LXC 7.0 on Debian 13 (Trixie), kernel 7.0 | Xen hypervisor with XAPI toolstack (XenServer lineage) |
| Current Release | Proxmox VE 9.2 (21 May 2026); arm64 build 5 Aug 2026 | XCP-ng 8.3 LTS (released 7 Oct 2024; current ISO 8.3.0-20260806) |
| Support Window | PVE 9 EOL to be announced; PVE 8 ended Aug 2026 | 8.3 LTS supported until 30 Nov 2028 |
| Free Use | Yes: full features with the no-subscription repo (Proxmox labels it testing/non-production) | Yes: XCP-ng is free; Xen Orchestra free from source (no updater, community support) or XO Lite on each host |
| Paid Support Pricing | Per CPU socket/year: Community €120, Basic €370, Standard €550, Premium €1,100 (net) | Vates VMS/year: Essential 2,000 (≤3 hosts), Essential+ 4,000 (≤3 hosts), Pro 1,000/host (3+), Enterprise 1,800/host (4+) |
| Management | Built-in web UI, CLI and REST API on every node; optional Proxmox Datacenter Manager for multi-site | Xen Orchestra (appliance or from source) web UI/API/CLI; XO Lite on 8.3 hosts; xe CLI; Terraform and Packer providers |
| Containers | LXC system containers built in; OCI images as a technology preview | |
| Clustering | Corosync multi-master cluster, no explicit node limit (50+ node clusters in production per docs) | Resource pools of up to 64 hosts, same XCP-ng version required; CPU features masked across hosts |
| High Availability | Built in; needs 3+ nodes for quorum; watchdog fencing; dynamic CRS load balancing in 9.2 | Built in; works with 2 hosts, 3+ recommended; requires shared storage (iSCSI, NFS, XOSTOR, FC) for heartbeat |
| Storage Options | ZFS, BTRFS, LVM, LVM-thin, NFS, CIFS, iSCSI, CephFS, Ceph RBD, ZFS over iSCSI; integrated Ceph | Local EXT, local LVM, NFS, iSCSI, Fibre Channel, XOSTOR, CephFS/Ceph RBD, SMB; ZFS as-is (no official support) |
| Max Virtual Disk | No platform cap stated (depends on storage backend) | VHD 2,040 GiB; QCOW2 up to 16,381 GiB (not on XOSTOR or SMB) |
| Backup | Built-in vzdump backups; Proxmox Backup Server (separate product) for incremental, deduplicated, encrypted backups | Xen Orchestra: full and incremental backups, disaster recovery, continuous replication, encryption, S3 targets |
| VMware Import | Integrated ESXi import wizard (ESXi 6.5–8.0 tested) + OVF/OVA | Xen Orchestra V2V from ESXi incl. warm migration; OVA import |
| ARM64 Hosts | Yes: official for NVIDIA Grace/Vera; other ARMv9-A UEFI best effort; no Raspberry Pi | |
| Minimum Hardware | Testing: 64-bit CPU with VT-x/AMD-V, 1 GB RAM; production: 2 GB for the OS plus guest memory | 64-bit x86 CPU with VT-x/AMD-V and SLAT, 2 GB RAM (8 GB recommended), 46 GB disk (70 GB recommended) |
Proxmox VE
Debian-based virtualization platform from Proxmox Server Solutions (Vienna) that runs KVM/QEMU virtual machines and LXC system containers side by side, with a built-in web UI on every node, multi-master clustering, integrated ZFS and Ceph, and a REST API. Licensed AGPL-3.0; enterprise repository access and support are sold per CPU socket.
Pros
- VMs and LXC system containers in one UI and API, so light services can run as containers without a nested Docker VM
- No separate management appliance: every node serves the GUI, CLI and API, and clusters are multi-master with no management single point of failure
- ZFS, BTRFS, LVM-thin, NFS, iSCSI, CephFS and Ceph RBD supported out of the box, with hyperconverged Ceph deployable from the GUI
- Free to run with the no-subscription repository; Community subscription for stable enterprise updates is only €120 per socket per year
- Integrated ESXi importer in the web UI (tested from ESXi 6.5 to 8.0), plus OVF/OVA import
- Official arm64 builds since August 2026 for NVIDIA Grace and Vera servers, with other ARMv9-A UEFI hardware on a best-effort basis
Cons
- Subscriptions are per CPU socket, so dual-socket hosts cost twice as much as single-socket ones
- The free no-subscription repository is described by Proxmox as for testing and non-production use, so production-grade updates mean paying per socket
- HA needs at least three nodes for reliable quorum, so two-node setups are a poor fit for HA
- Online VM migration needs compatible CPUs across nodes, and arm64 and x86 nodes can't live-migrate between each other
Best For
Homelabs, self-hosters and small to mid-size teams who want VMs and containers on the same hosts, built-in ZFS or hyperconverged Ceph, and a platform that runs fully without a separate management server.
XCP-ng
Open-source Xen-based hypervisor forked from Citrix XenServer, developed by Vates and incubated in the Xen Project at the Linux Foundation. Hosts are grouped into pools of up to 64 servers and managed through Xen Orchestra (web UI, API, backups), with XO Lite built into each 8.3 host. Support is sold per host as part of Vates VMS.
Pros
- Pool model familiar to XenServer and vSphere admins: up to 64 hosts per pool, live migration and live storage migration between them
- Xen Orchestra adds agentless backups (full, incremental, disaster recovery, continuous replication, encryption, S3 targets) without extra licences when built from source
- Per-host support pricing (Pro 1,000 per host per year) is cheaper than per-socket pricing on dual-socket or many-core servers
- Smooth migration paths: in-place upgrade from XenServer, and Xen Orchestra V2V streams VMs from ESXi, including warm migration
- QCOW2 support since 5 May 2026 lifts the old 2,040 GiB per-disk limit to 16,381 GiB on most storage types
- XOSTOR hyperconverged storage (DRBD/LINSTOR-based) for pooling local disks across hosts
Cons
- VMs only: no built-in container type, and the docs say an XCP-ng host is dedicated to hosting VMs, with third-party software in dom0 unsupported
- Day-to-day management at scale needs Xen Orchestra running somewhere (as a VM or from source); the free from-source build has no updater and only community support
- VHD remains the default disk format with a 2,040 GiB limit; QCOW2 isn't available on XOSTOR or SMB storage
- Smaller ecosystem than Proxmox, x86-64 only, ZFS offered as-is without official support, and per-VM limits (32 vCPUs security-supported) matter for very large VMs
Best For
Teams leaving XenServer or VMware who want pool-based management, Xen Orchestra's agentless backup and replication, and per-host support pricing on large multi-socket servers.
Pricing: per socket vs per host, and what free really means
Both platforms are free to install and use with every feature. Proxmox VE's code is AGPL-3.0, and without a subscription you point APT at the pve-no-subscription repository. Proxmox says that repository is freely accessible but intended for testing and non-production use, because packages land there before they reach the enterprise repository. Paid plans are per CPU socket per year, net of VAT: Community €120 (enterprise repository, no support tickets), Basic €370 (3 tickets, next-business-day response), Standard €550 (10 tickets, 4-hour response, 24/7 coming by end of 2026) and Premium €1,100 (unlimited tickets, 2-hour response, 24/7 since 19 October 2026). Every plan includes the enterprise repository, and Basic, Standard and Premium also list Proxmox Datacenter Manager. XCP-ng itself is free. Vates sells support for the whole stack (XCP-ng plus Xen Orchestra) as Vates VMS: Essential at 2,000 a year for up to 3 hosts (6 tickets), Essential+ at 4,000 for up to 3 hosts (unlimited tickets), Pro at 1,000 per host (minimum 3) and Enterprise at 1,800 per host (minimum 4, 24/7, 1-hour critical response). Multi-year terms save up to 10% (3 years) or 15% (5 years), and the pricing page shows the same figure in dollars and euros. A worked example: three dual-socket hosts are six sockets. Proxmox Basic costs €2,220 and Standard €3,300 a year. Vates Essential is 2,000 and Pro is 3,000 for the same three hosts. On single-socket hosts the maths flips: Proxmox Basic for three sockets is €1,110.
Architecture and management: self-managing nodes vs pools plus Xen Orchestra
Proxmox VE is Debian 13 with a KVM/QEMU and LXC stack on top. Every node runs the full web interface, CLI and REST API, and a cluster built with pvecm replicates its configuration to all nodes in real time through corosync. There is no management VM to keep alive. The docs set no explicit node limit and mention production clusters of more than 50 nodes. They also note that all nodes need UDP ports 5405–5412 and SSH between them. Proxmox Datacenter Manager adds a central view across several clusters, and it is listed with the Basic, Standard and Premium subscriptions. XCP-ng follows the XenServer model. Hosts join a resource pool of up to 64 servers, one acts as pool coordinator over XAPI on port 443, and all hosts must run the same XCP-ng version and update level. Management at scale goes through Xen Orchestra. You can deploy the XOA appliance from the vates.tech deploy page, or build XO from source for free; that version has no built-in updater and only community best-effort help. Hosts on 8.3 also serve XO Lite at https://<host-ip> for single-host tasks. The practical difference is that Proxmox needs nothing extra, while XCP-ng gives you a deliberately central console. Teams coming from vCenter tend to find XCP-ng's model familiar, and homelabbers tend to find Proxmox's simpler.
Storage and limits: ZFS and Ceph vs SMAPI, VHD and the new QCOW2 path
Proxmox's storage list is long: ZFS, BTRFS, LVM and LVM-thin, NFS, CIFS, iSCSI (kernel and libiscsi), CephFS, Ceph RBD and ZFS over iSCSI. Ceph can be deployed hyperconverged from the GUI, and the 9.2 release ships Ceph Squid 19.2.3 and Tentacle 20.2.1. The requirements page asks for about 1 GB of RAM per TB of storage on top of the base 2 GB when you use ZFS or Ceph, and warns that neither works behind a hardware RAID controller. XCP-ng's current storage stack is SMAPIv1. VHD is the default format, with a maximum of 2,040 GiB per virtual disk, which for years was the most-cited XCP-ng limitation. Since 5 May 2026, 8.3 also supports QCOW2 on the same stack, with disks up to 16,381 GiB on most storage repository types. XOSTOR and SMB are the exceptions, and disks larger than 2,040 GiB are created as QCOW2 automatically where it is supported. The docs recommend file-based SRs (local EXT, NFS) because they are thin-provisioned and easier to handle. ZFS is offered as-is without official support. XOSTOR is Vates' hyperconverged option, sold with its own support tiers (Pro 800 per host per year). Per-VM limits on XCP-ng are 32 vCPUs (the security-supported limit for untrusted VMs), 1.5 TiB of RAM and up to 8 vGPUs.
High availability, backups and leaving VMware
Proxmox HA is built into the cluster stack. The docs require at least three nodes for reliable quorum and fence failed nodes with a watchdog, falling back to the Linux softdog if there is no hardware watchdog. Release 9.2 added dynamic load balancing through the Cluster Resource Scheduler, which can live-migrate HA guests to even out real-time load, plus disarm-ha and arm-ha commands for cluster-wide maintenance without fencing. XCP-ng HA can work with two hosts, but Vates recommends three to avoid split-brain. It requires shared storage (iSCSI, NFS, XOSTOR or Fibre Channel, at least 4 GiB) for a storage heartbeat next to the network heartbeat. Backups differ in shape. Proxmox has built-in vzdump backups and a separate Proxmox Backup Server for incremental, deduplicated, encrypted backups with live restore (PBS subscriptions start at €560 per server per year). Xen Orchestra includes agentless full and incremental backups, disaster recovery and continuous replication jobs, encrypted repositories and S3-compatible targets. For VMware exits, Proxmox has an integrated importer that treats an ESXi host as a storage source (tested with ESXi 6.5 to 8.0; VMs with snapshots import more slowly). Xen Orchestra's V2V streams VMs from any reachable ESXi host and supports warm migration: copy up to the last snapshot, stop the VM, then copy the delta. Since XO 5.112, disks over 2,040 GiB come across as QCOW2.
Where each one breaks
Proxmox's weak spots are mostly policy and topology. If an auditor wants vendor-tested updates, the free repository doesn't qualify, and per-socket pricing adds up on big dual-socket fleets. Two-node clusters are awkward because the HA docs ask for at least three nodes for quorum. Live migration needs compatible CPUs, and mixed arm64 and x86 clusters can't migrate between architectures. The arm64 build officially covers only NVIDIA Grace and Vera servers, and Raspberry Pi-style device-tree boards aren't supported at all. XCP-ng's weak spots are scope and tooling. The host is a VM appliance: no containers, no extra packages in dom0 beyond the official repositories, and x86-64 only. Anything beyond single-host chores means running Xen Orchestra, and choosing between the paid appliance with updates and support and the free source build that you maintain yourself. Disk size used to be a hard wall at 2,040 GiB. QCOW2 fixes that on most storage, but not on XOSTOR or SMB, so check your SR type before planning multi-terabyte VMs. All hosts in a pool must run the same version, which makes rolling upgrades a planned exercise. Mixing Intel and AMD needs a forced join and loses live migration between them. Neither project publishes benchmark comparisons we could verify, so treat performance claims from either side as anecdotal.
Verdict
Pick Proxmox VE unless you have a specific reason to want Xen. For a homelab, a single VPS-style box or a small team, it is the shorter path: install one ISO, open the web UI, and you have VMs, LXC containers, ZFS and backups without deploying anything else. It runs fully without a subscription, and the €120-per-socket Community plan gets you the tested enterprise repository when the box starts carrying real work. It also wins if you want hyperconverged Ceph across three or more nodes, managed from the same UI. Pick XCP-ng if you are moving off XenServer, where it is an in-place upgrade, or off VMware with people who think in pools, shared storage and a central console. Xen Orchestra's agentless backup, replication and V2V tooling is the strongest part of the stack. Vates' per-host pricing also comes out ahead on big dual-socket servers once you pay for support. Skip XCP-ng if you want containers on the host or need ARM, and skip Proxmox's free repository for production if your policy demands vendor-tested updates. In that case budget at least the Basic or Community subscription.