An idea older than x86
A virtual machine is a program that shows the guest system a whole computer: CPU, memory, disk, network. The guest thinks the hardware is its own, while in fact it shares it with neighbors. The idea is almost 60 years old.
1967 IBM CP-40 and CP-67 the first virtual machines: each mainframe user gets their own System/360
1972 IBM VM/370 virtual machines in a production system; the line lives on as z/VM
1974 Popek and Goldberg the conditions under which a CPU can be virtualized without loss
1999 VMware Workstation virtual machines on a regular x86 PC
2001 VMware ESX Server a hypervisor instead of an OS, right on the hardware
2003 Xen (Cambridge) paravirtualization: the guest kernel knows it's a guest
2003 QEMU (Fabrice Bellard) a whole computer emulated by a program
2005 OpenVZ containers on a patched Linux kernel: cheap VPSes
2005 Intel VT-x the CPU itself helps the hypervisor; AMD-V followed in 2006
2007 KVM in Linux 2.6.20 the Linux kernel becomes a hypervisor
2008 LXC containers on a stock kernel: namespaces and cgroups
Why x86 resisted for so long
A classic hypervisor works like this: the guest runs directly on the CPU, and every dangerous instruction — switching page tables, accessing a device — triggers a trap. Control goes to the hypervisor, which does the same thing with virtual hardware and hands control back to the guest. In 1974 Popek and Goldberg stated the condition: all such instructions must trap.
x86 didn’t meet it: some instructions in unprivileged mode silently did something else, without a trap. The guest broke, and the hypervisor never found out. There were three workarounds:
| How | Who | The price |
|---|---|---|
| Binary translation: the hypervisor rewrites the dangerous bits of guest code on the fly | VMware, 1999 | hypervisor complexity, overhead on system calls |
| Paravirtualization: the guest kernel is modified and calls the hypervisor itself | Xen, 2003 | only guests with a modified kernel; Windows can’t run that way |
| Emulation: a program imitates the whole CPU | QEMU, 2003 | speed: many times slower than hardware |
In 2005–2006 Intel (VT-x) and AMD (AMD-V) added a separate guest mode to their CPUs. Now the CPU itself hands the guest’s dangerous instructions to the hypervisor. The tricks were no longer needed, and the hypervisor shrank to a kernel module. Hence the hardware requirement: without VT-x or AMD-V Proxmox won’t start VMs (Hardware for Proxmox).
KVM: a hypervisor from the Linux kernel
KVM was written by Avi Kivity at the Israeli company Qumranet. It was announced in October 2006, and by February 2007 the module was in Linux kernel 2.6.20. With KVM the job split in two:
- the Linux kernel with the
kvmmodule runs guest code on the CPU through VT-x or AMD-V; memory and the scheduler are the ordinary kernel ones: to the host, a VM is just one of the processes; - QEMU imitates everything else: disk, network card, video, BIOS. Emulation speed doesn’t matter here: the fast devices are VirtIO, and their drivers know they’re in a virtual machine.
In 2008 Red Hat bought Qumranet, and KVM became the base of its virtualization, then of OpenStack and most clouds.
Proxmox VE runs VMs the same way: every VM on the host is a kvm process (QEMU with KVM).
OpenVZ and containers before Docker
A different line ran in parallel — not imitating a computer, but sharing one kernel. SWsoft (later Parallels, then Virtuozzo) sold Virtuozzo for hosting from the early 2000s and in 2005 opened up its core as OpenVZ. An OpenVZ container looked like a separate server with root and SSH, but ran on a shared kernel: hundreds of cheap VPSes fit on one machine.
The price was the kernel: OpenVZ needed its own heavily patched one. Mainline Linux moved ahead while OpenVZ stayed on 2.6.18 and 2.6.32 for years. Meanwhile namespaces and cgroups landed in the stock kernel, and LXC appeared on top of them. That sealed OpenVZ’s fate in Proxmox.
How Proxmox appeared
Proxmox Server Solutions was founded by the brothers Martin and Dietmar Maurer in Vienna in February 2005. Its first product wasn’t a hypervisor but a mail filter: Proxmox Mail Gateway 1.0 came out in April 2005.
Proxmox VE appeared as an assembly of ready parts: Debian, KVM for full VMs, OpenVZ for containers, and a web UI managing both. In 2008 that was rare: VMware cost money, Xen required modified kernels, and KVM and OpenVZ were sets of command-line tools.
2008 April 0.9 — the first public release: KVM and OpenVZ on Debian, web UI
2008 October 1.0 — live migration for KVM and OpenVZ, vzdump backups from the UI
2009 1.4 storage model: iSCSI, NFS, LVM, DRBD
2012 2.0 cluster on corosync and /etc/pve (pmxcfs), REST API, ExtJS UI, AGPLv3
2013 3.0 VM clones; pveproxy instead of Apache
2013 3.1 enterprise and no-subscription repositories
2014 3.3 pve-firewall, two-factor login, noVNC console
2015 3.4 ZFS in the installer with every RAID level
2015 4.0 LXC instead of OpenVZ, a new high-availability (HA) manager
2016 4.2 LVM-thin, Let's Encrypt; 4.4 — unprivileged containers in the UI
2017 5.0 storage replication on ZFS
2018 5.2 cloud-init in the UI, SMB/CIFS storage
2019 6.0 corosync 3, ZFS encryption; 6.2 — API tokens, zstd backups
2020 6.3 Proxmox Backup Server 1.0 and its support in Proxmox VE
2021 7.0 Debian 11, Btrfs (technology preview), OpenID Connect login
2023 8.0 Debian 12; 8.1 — SDN and Secure Boot
2024 8.2 ESXi import wizard, automated installation
2025 9.0 Debian 13, snapshots on plain LVM (preview), HA placement rules
2025 9.1 LXC containers from OCI images (preview)
2026 9.2 kernel 7.0, QEMU 11, cluster load balancing
Three turns from this list defined what Proxmox became.
2012: the cluster and /etc/pve
In 2.0 the configs of all VMs moved to /etc/pve — the pmxcfs file system that corosync syncs between cluster
nodes. Any node sees and edits any VM, and the UI became a REST API client. The scheme hasn’t changed since: a VM
config from 2012 and one from 2026 live in the same place.
2013: the subscription and two repositories
In 3.1 updates split. pve-enterprise is for subscribers; packages arrive there after being broken in.
pve-no-subscription is for everyone; packages arrive there earlier. The subscription pop-up at login comes from
the same time. The code stayed fully open: the subscription sells tested packages and support, not features.
2015: OpenVZ leaves, LXC arrives
By 2015 Debian 8 shipped with kernel 3.16, while the OpenVZ kernel stayed on 2.6.32. Dragging an old kernel along
for containers meant going without new drivers and new KVM features. In 4.0 OpenVZ was removed, and containers moved
to LXC — on the stock Proxmox kernel. Old containers were moved over through a backup: pct restore unpacks an
OpenVZ archive into an LXC container.
2023–2024: Broadcom and the exodus from VMware
In November 2023 Broadcom closed the VMware acquisition and moved the products to subscriptions. In February 2024 the free ESXi disappeared — the base of thousands of home labs and small offices — and corporate renewals, by Gartner’s estimate, doubled in price in places.
Proxmox turned out to be the closest free alternative and answered in April 2024: 8.2 shipped an import wizard that connects an ESXi host as a storage and pulls VMs along with their disks. In May 2024 Veeam announced Proxmox VE support. The install counter Proxmox keeps shows how the move went: over a million hosts at the 8.2 release, about 1.3 million by November 2024, over 2.3 million active servers in September 2026.
In April 2025 Broadcom quietly brought back the free ESXi 8 (8.0 Update 3e): no support, not for production, downloadable from the Broadcom portal after registration.
Where it is now
The 9 branch runs on Debian 13 “trixie”; the current version is 9.2 from May 21, 2026: kernel 7.0, QEMU 11.0, LXC 7.0, ZFS 2.4. In August 2026, 9.2 for ARM came out — the first architecture after x86-64, for now for NVIDIA Grace and Vera servers. The 8 branch on Debian 12 was supported until August 2026. A major version comes out every two years, following Debian; point releases usually in spring and autumn.
Proxmox VE now has a family around it: Proxmox Backup Server (1.0 in November 2020, now 4.2), Proxmox Mail Gateway and Proxmox Datacenter Manager — one window for several hosts and clusters (1.0 in December 2025).
The Roadmap shows where Proxmox is heading:
- Rust instead of Perl. The backend is historically Perl, new components are Rust; the UI is moving from ExtJS to its own Rust and Yew widget toolkit, as already done in the 9.0 mobile version and in Datacenter Manager.
- Application containers. OCI images as LXC containers — a preview since 9.1; next up, multi-container applications from a Compose description.
- Installation. Full LUKS disk encryption in the installer and more options for unattended installs.
- Network and cluster. SDN leaves preview, and the load balancer learns to move VMs without HA too.
If it doesn’t work
A guide from the internet doesn’t work: no vzctl, no /etc/vz, commands about OpenVZ
- Cause
- The guide is for Proxmox 3.x or earlier. OpenVZ was removed in 4.0 (October 2015); containers have been LXC ever since.
- Fix
- Look for the same thing for LXC:
pctcommands, configs in/etc/pve/lxc/. An old OpenVZ container moves over through a backup:pct restoreaccepts an OpenVZ vzdump archive.
The guide has a sources.list with bookworm, but we have .sources files and trixie
- Cause
- The guide is for 8.x on Debian 12. Since 9.0 Proxmox runs on Debian 13 and keeps repositories in the deb822 format — one
.sourcesfile per source. - Fix
- Edit
/etc/apt/sources.list.d/*.sourcesor enable the repository in the UI: node › Updates › Repositories (No-subscription repository).
There’s no ESXi import wizard in the Storage › Add menu
- Cause
- The host is older than 8.2: ESXi import appeared in April 2024.
- Fix
- Update to 9.x. If you can’t yet, move the disk by hand: export to OVF and import with
qm importovf.