On 5 August 2026, Proxmox released an official Arm64 build of Proxmox VE 9.2. If you run Proxmox in production, the practical question is not whether the news is real — it is — but whether it should change what you buy. For almost every deployment, the answer is no, and the reasons come from Proxmox's own documentation rather than from anyone's opinion.

This article is for infrastructure engineers and IT buyers planning their next Proxmox host. It covers exactly what shipped, the six limitations the announcement did not mention, and the five specific situations where Arm is genuinely the right purchase.

What actually shipped on 5 August 2026

Proxmox VE 9.2 for arm64 is a new build of an existing release, not a new release. Proxmox VE 9.2 shipped for x86-64 on 21 May 2026; the arm64 build of that same 9.2 shipped on 5 August 2026 as a separate entry on the Proxmox VE roadmap. Proxmox describes it there as the "first Proxmox VE release for a CPU architecture besides x86-64 (amd64)."

The installer is proxmox-ve_9.2-1-arm64.iso, listed on the Proxmox downloads page. The build is based on Debian 13 "Trixie" with Linux kernel 7.0, QEMU 11.0, LXC 7.0 and ZFS 2.4 — the same component versions as the x86-64 build. It shares the code base, package repositories and release lifecycle with x86-64, and the pve-enterprise repository is available for arm64.

Hardware support comes in three tiers, in Proxmox's own words:

Tier	Hardware	Proxmox's wording
Fully supported	NVIDIA Grace and Vera platforms	"Fully supported on the NVIDIA Grace Hopper and NVIDIA Vera platforms"
Best-effort	Other UEFI-based ARMv9-A or newer	"Best-effort support on other UEFI-based ARMv9-A or newer hardware"
Best-effort	ARMv8-A	"ARMv8-A generally works as well, likewise best-effort"
Not supported	Device-tree-only SBCs, including Raspberry Pi	"Device-tree-only single-board computers, such as the Raspberry Pi, are not supported"

Every host must boot through UEFI and describe its hardware through ACPI. Proxmox developed the release in collaboration with NVIDIA and Supermicro, performing joint validation on NVIDIA Grace Hopper Superchip systems.

Sizing a Proxmox host — CPU flags, ECC RAM budgeting, storage controllers, networking — is unchanged by this release and is covered in our guide to Proxmox hardware requirements. What follows is only the architecture question.

This reverses a position Proxmox held until this year

The top search results for "Proxmox Arm" still tell you official Arm64 support does not exist, because until 5 August 2026 that was true and the sources saying so are append-only forum threads.

On 27 February 2024, Proxmox staff member Fiona wrote in the official forum thread arm64 support: "There is no such work in progress from our side. It was evaluated a few years ago, but not deemed worth pursuing further." Four weeks later, in the same thread, Fiona added: "Only running ARM guests on an x86_64 host is officially supported by us."

As recently as 14 March 2026, the Ampere community forum recorded the same conclusion: "As I know, the Proxmox team doesn't have any plan to officially support arm64."

Both statements are now obsolete: Proxmox on Arm is official as of 5 August 2026. If you are researching this and land on a 2024 forum thread, check the date before you act on it.

Six limitations the announcement left out

The press release describes "full feature parity," and its own scope is specific: Proxmox's CTO lists "KVM, networking, clustering, ZFS, and Ceph all behaving exactly as our users expect from our x86-64 builds." That list is the claim. Six things sit outside it, all documented in the staff announcement or on Proxmox's own download page.

1. Arm nodes cannot live migrate to or from your x86 cluster

This is the single most consequential planning fact in the release. Proxmox states: "Guests only run on nodes matching their architecture, and live migration is only possible between nodes of the same architecture."

An arm64 node can join an existing x86-64 cluster — Proxmox says this is "not blocked technically, but mixed-architecture clusters are not officially supported." What it cannot do is host your existing workloads. Your x86 VMs will not start on it, will not live migrate to it, and cannot fail over to it, because a guest that only runs on architecture-matched nodes has nowhere to land.

The operational consequence: an arm64 node in a mixed cluster is a management-plane member, not usable capacity. Evacuating an x86 host for maintenance still requires another x86 host. Buying one Arm node "to try it" adds no headroom to an existing cluster.

2. Arm64 requires its own subscription

Proxmox states plainly: "Subscriptions for arm64 are separate from the x86-64 ones and are currently available on request." An existing x86-64 subscription does not cover an arm64 node.

Proxmox VE subscriptions are priced per CPU socket per year, across four tiers — Community, Basic, Standard and Premium — with the enterprise repository included at every tier and support ticket allowances rising by tier. Those published prices apply to x86-64. The Proxmox pricing page handles arm64 differently: "Need ARM64 support? Request support for bleeding-edge server hardware through our sales team." No arm64 price is published.

For budgeting, that matters twice. An arm64 node is an unpriced line item until Proxmox sales responds, and a mixed-architecture estate means two separate subscription relationships covering one cluster. As an authorized Proxmox reseller, we sell Proxmox subscriptions attached to x86-64 server builds and pre-install the chosen tier before shipment, so this is the part of the release we watch most closely — the technical gate is open on arm64, but the commercial one is quote-only.

3. There is no SeaBIOS — every VM boots UEFI

Proxmox states: "Virtual machines always boot through UEFI, provided by the ARM build of OVMF (AAVMF). SeaBIOS is not available on arm64."

On x86 Proxmox, switching a stubborn guest to SeaBIOS is a standard fix for an appliance image that will not boot under UEFI. On arm64 that lever does not exist. Any guest whose boot path exists only in legacy-BIOS mode has no boot path at all, and no fallback firmware to fall back to.

4. No AMD SEV, no Intel GVT-g, no microcode packages

Proxmox: "AMD SEV memory encryption and Intel GVT-g mediated vGPUs are x86-only," and "There is no operating-system-level CPU microcode package, unlike intel-microcode and amd64-microcode on x86."

AMD SEV encrypts guest memory so the host operator cannot read it — a compliance control some organizations have already written into policy. Note the honest limit: SEV is an AMD feature, so an Intel-based host does not offer it either. Intel GVT-g splits an integrated GPU across VMs, which matters to small VDI deployments carving up an iGPU.

The microcode gap is the one most likely to surprise an operator. On x86, intel-microcode and amd64-microcode deliver CPU errata and side-channel mitigations through the OS, independent of your server vendor's BIOS cadence. On arm64 there is no equivalent, so silicon-level fixes arrive only through platform firmware updates from the hardware vendor.

5. Raspberry Pi and single-board computers are still excluded

Proxmox explicitly excludes device-tree-only single-board computers and names the Raspberry Pi. The requirement is that hosts "must boot through UEFI and describe their hardware through ACPI" — the firmware standard behind Arm SystemReady. SBCs describe their hardware with a compiled device tree instead, which puts them outside supported configurations regardless of how capable the silicon is.

If you came to this looking for official Proxmox on a Pi, the answer is still no.

6. Proxmox Backup Server has no official Arm64 build

A fully supported all-Arm Proxmox estate, backups included, is not currently possible. Proxmox VE is the only Proxmox product with an arm64 ISO on the downloads page — Proxmox Backup Server, Mail Gateway and Datacenter Manager are amd64 only.

Community-maintained arm64 builds of Backup Server exist, and in July 2026 a Proxmox staff member asked one such project to state clearly that these are "not official builds but rather community maintained." If your backup target must be vendor-supported, it stays on x86-64 even if your hypervisors do not.

An Arm64 host runs Arm64 guests — and nothing else, usefully

Because guests only run on architecture-matched nodes, an arm64 hypervisor runs arm64 guests. Your existing x86 VM templates, golden images and appliance images do not run on it. There is no conversion path: a VM disk image contains architecture-specific kernel and userland binaries.

QEMU can emulate a foreign architecture using its Tiny Code Generator instead of KVM, so x86 guests on an Arm host are technically possible. Linaro measured TCG software emulation at roughly 8 times slower than hardware acceleration for x86-64 guests. A workload that needed one core now wants eight, which spends the Arm efficiency advantage several times over to run the wrong architecture.

Windows is a harder constraint than it looks. Microsoft publishes Windows 11 Arm64 ISOs, but its documentation names exactly three virtualization paths: Hyper-V on an Arm-based Windows PC, Azure, and Arm-based Apple Silicon Macs. It also notes that "Hyper-V on x64 hardware doesn't support Arm64 VMs." Microsoft does not name KVM among its supported hosts for Windows on Arm.

Linux containers are the least painful part. An arm64 host runs arm64 container images, and coverage on the standard LXC image server is good for mainstream distributions. The real migration cost is the VM and appliance library, not the containers.

The hardware economics, with the arithmetic shown

Arm's efficiency advantage is real, and it is smaller than the headline numbers suggest.

The NVIDIA Grace CPU Superchip carries 144 Arm Neoverse V2 cores, with NVIDIA describing "the entire CPU including the memory consuming just 500 watts." Compare that against the x86 platforms most Proxmox hosts actually run on. Dell's technical guide gives the PowerEdge R750 a "Max TDP: 270 W" per socket across two 40-core 3rd Gen Xeon Scalable processors. Dell's spec sheet gives the PowerEdge R760 up to two 64-core 5th Gen Xeon Scalable processors and "2 x 350 watt processors."

Platform	Cores	Package power	Watts per core
Grace CPU Superchip	144	500 W (includes memory)	3.47 W
R760, 2× 64-core Xeon	128	700 W (CPU only)	5.47 W
R750, 2× 40-core Xeon	80	540 W (CPU only)	6.75 W

Grace is about 1.6 times more power-efficient per core than a fully populated R760 and about 1.9 times more efficient than a fully populated R750 — and that understates it, because NVIDIA's 500 W includes memory while Dell's TDP is CPU package only. These are package figures, not measured wall power. Measured draw on a real R750 — idle, typical and peak — is covered separately in our breakdown of R750 power consumption.

Now run it to a dollar figure. At an electricity price of $0.135/kWh and a facility PUE of 1.54 — the Uptime Institute's 2025 weighted average, and both numbers you should substitute with your own — an 80-core R750 costs roughly $983 a year in CPU-attributable energy. Eighty cores' worth of Grace costs roughly $506. The saving is about $477 a year.

Against that, systems built on Grace Hopper start around $41,500. Even if the refurbished R750 were free, recovering that difference on energy alone takes roughly 87 years. Change the electricity price and the payback moves; it does not move enough to matter at SMB or mid-market scale.

Two other numbers deserve care. NVIDIA's widely quoted SPECrate 2017_int_base figure of 740 is footnoted as a "pre-silicon Grace projection" from 2022 — it is an estimate, not a published benchmark result. And the "double the bandwidth of traditional DDR5" claim was made against the DDR5 designs of that era; against a current two-socket DDR5-5600 R760, Grace's 1024 GB/s is closer to 1.4 times, not 2.

You can buy used Arm servers — just not the ones Proxmox fully supports

This is where our view differs from most coverage of the release, and it is the part most relevant to anyone comparing acquisition costs.

A secondary market for Arm server silicon does exist. As of 5 August 2026, a used Ampere Altra Max M128-26 — 128 cores — was listed on eBay at $557.10, alongside Altra Q80-30 and Q32-17 parts and complete dual-Altra systems. Individual listings expire, but the supply is real and continuous. Anyone claiming there is no used Arm supply has not looked.

The problem is which tier that silicon lands in. Ampere Altra is built on Arm Neoverse N1, an ARMv8-A design — which puts it in Proxmox's "generally works as well, likewise best-effort" category, not the fully supported tier. The fully supported tier is Grace and Vera, and those are new-only, allocation-constrained platforms with no meaningful off-lease supply. You can buy used Arm servers today, but not the ones Proxmox fully supports — and you cannot buy the ones it does fully support used at all.

The x86 secondary market has the opposite shape. Three-year refresh cycles and lease expirations return large volumes of 15th- and 16th-generation PowerEdge and Gen10 ProLiant hardware while it still has years of service life, which is why a production-grade refurbished Dell PowerEdge server is available today, in quantity, with a warranty. We stock, test and warranty x86-64 hardware exclusively — we sell no Arm servers and have no first-hand Arm operational experience to offer. That is precisely why the sourcing above comes from Proxmox, NVIDIA and Dell rather than from us.

Who should actually buy Arm for Proxmox

Arm is the right purchase in five specific situations:

• AI inference co-located with the hypervisor. The GH200 links Grace and Hopper with an NVLink-C2C interface running at 900 GB/s. A PCIe Gen5 x16 link — the best any x86 server gives a GPU — is roughly an order of magnitude below that. If your bottleneck is moving data between CPU and GPU memory, no x86 server closes that gap.
• Memory-bandwidth-bound workloads. Grace's 1024 GB/s beats a two-socket DDR5-5600 host by roughly 1.4 times. In-memory analytics, HPC solvers and large-graph workloads feel that.
• A hard kW-per-rack cap on a greenfield build. When the constraint is power rather than capital, 3.47 W per core fits more cores inside the cap than 5.47 or 6.75 does.
• Greenfield with an already Arm-native software stack. Containerized services in Go, Rust or Java with no x86 template library to port.
• Standardizing on Grace or Vera anyway. If you are buying that hardware for AI regardless, PVE 9.2 arm64 means you no longer need a separate hypervisor story for it.

Outside those five, x86-64 remains the correct answer — and that is not a close call given no cross-architecture migration, a separate subscription, and no supported backup server.

Frequently Asked Questions

Does Proxmox support ARM64?

Yes, as of 5 August 2026. Proxmox VE 9.2 for arm64 is an official release with its own installer ISO and access to the pve-enterprise package repository. It reverses Proxmox's earlier position — in April 2024, Proxmox staff stated on the official forum that there was "no such work in progress from our side."

Can ARM64 nodes join my existing x86 Proxmox cluster?

Technically yes, officially no. Proxmox states: "This is not blocked technically, but mixed-architecture clusters are not officially supported." More importantly, guests only run on nodes matching their architecture, so your existing x86 VMs cannot start on an arm64 node and cannot migrate to it. An arm64 node in a mixed cluster is a management-plane member, not usable capacity.

Can I live migrate VMs between ARM and x86 Proxmox nodes?

No. Proxmox states that "live migration is only possible between nodes of the same architecture." This is architectural rather than a missing feature — a running VM's CPU state cannot be transplanted onto a different instruction set. Plan arm64 as a separate cluster, not as extra capacity in an existing one.

Does my existing Proxmox subscription cover ARM64 nodes?

No. Proxmox states that "subscriptions for arm64 are separate from the x86-64 ones and are currently available on request." Published per-socket pricing across the Community, Basic, Standard and Premium tiers applies to x86-64; the pricing page directs arm64 inquiries to Proxmox sales, and no arm64 price is published.

Can Proxmox run on a Raspberry Pi now?

No. Proxmox explicitly excludes device-tree-only single-board computers and names the Raspberry Pi. Hosts must boot through UEFI and describe their hardware through ACPI, the firmware standard behind Arm SystemReady. Single-board computers use a compiled device tree instead, placing them outside supported configurations.

Which ARM hardware does Proxmox actually support?

Three tiers. Fully supported: NVIDIA Grace and Vera platforms. Best-effort: other UEFI-based ARMv9-A or newer hardware, with ARMv8-A also described as generally working on a best-effort basis. Not supported: device-tree-only single-board computers. Note that Ampere Altra, the Arm server CPU most available on the used market, is a Neoverse N1 ARMv8-A design and therefore sits in the best-effort tier.

Can I run x86 virtual machines on an ARM64 Proxmox host?

Only through emulation, and not usefully in production. KVM requires guest and host architectures to match, so an x86-64 guest on an arm64 host falls back to QEMU's TCG software emulation. Linaro measured TCG at roughly 8 times slower than hardware acceleration for x86-64 guests.

Can I run Windows VMs on Proxmox ARM64?

Only Windows on Arm builds, and Microsoft does not list KVM among its supported paths. Microsoft publishes Windows 11 Arm64 ISOs but names only three virtualization routes: Hyper-V on an Arm-based Windows PC, Azure, and Arm-based Apple Silicon Macs. If your estate is Windows-heavy, arm64 is not the platform.

Does Proxmox Backup Server run on ARM64?

Not officially. Proxmox VE is the only Proxmox product with an arm64 ISO on the downloads page — Backup Server, Mail Gateway and Datacenter Manager are amd64 only. Community-maintained arm64 builds exist, and Proxmox staff have asked those projects to state that they are not official builds. A fully supported all-Arm Proxmox estate including backups is not currently possible.

Should I buy ARM or x86 for my next Proxmox host?

x86 for almost every deployment, with three technical exceptions: if you need CPU-to-GPU coherent bandwidth for AI inference, if you are memory-bandwidth-bound, or if you face a hard power-per-rack cap. Otherwise the economics do not work. Fully supported Arm platforms start around $41,500, while the CPU-attributable energy saving over an 80-core refurbished PowerEdge R750 is roughly $477 a year at $0.135/kWh and a PUE of 1.54.

What does "full feature parity" not cover on ARM64?

The parity claim covers KVM, networking, clustering, ZFS and Ceph. Outside that list: no SeaBIOS, since VMs boot UEFI through AAVMF only; no AMD SEV memory encryption; no Intel GVT-g mediated vGPUs; no OS-level CPU microcode package; no cross-architecture live migration; a separate subscription; and no official Proxmox Backup Server build.

Where to go from here

If you run Proxmox today, the practical next step is to do nothing to your architecture and note the date on any Arm advice you find — anything written before 5 August 2026 is describing a different product. If you are specifying a new host, size it the same way you always have and buy x86-64 unless one of the five Arm cases above describes your workload precisely.

If you are buying a host and want Proxmox already running on it, we configure and pre-install the tier you choose before the server ships. Compare Proxmox subscription tiers and the PowerEdge platforms they run on.