senn-techsenn-tech
Infrastructure
Infrastructure2026-10-01· By Franz Senn

After leaving VMware: what the licence bill looks like in 2026

In February 2025 we wrote up why we switched hypervisor. Whether Proxmox can carry a company's infrastructure is a question we already worked through separately in September 2026. A year and a half has passed since that first post, and there are now numbers that did not exist then: documented price increases with names and dates, a Proxmox price list with clear tiers, and operating experience about where the money goes after such a move. This post is that arithmetic.

What is documented on price, and by whom

The mechanics sit in a Broadcom knowledge base article: billing per core, with at least 16 physical cores licensed per CPU. Perpetual licences have not been sold since 11 December 2023. On the size of the increase there are several statements from different groups, and those numbers should not be blended into one figure.

Who named which price increaseGartner: typical invoice400Anexia, per press500CISPE, cumulative1000AT&T, court filing1050ECCO/CISPE, upper end150001600
Five different bases, and no average of them. Gartner named 300 to 400 percent as the typical invoice (11 May 2026). Anexia was reported above 500 percent without putting its own figure on it. CISPE told the EU more than 1,000 percent cumulatively. AT&T put roughly 1,050 percent in a court filing. ECCO members reported 800 to 1,500 percent. (Quelle: The Register, Gartner, CISPE)

The Austrian proof point is a company that actually left: Anexia in Klagenfurt moved 12,000 virtual machines off VMware onto its own KVM platform, finished in May 2024. In its distributed hybrid infrastructure assessment of 22 September 2026, Gartner writes that by 2029, 55 percent of enterprises will be running proof-of-concept tests on alternatives to VMware products, against 25 percent in 2026. The dates that actually drive a migration calendar are the support endings: general support for ESXi 7.0 ended on 02 October 2025, vSphere 8 ends on 11 October 2027, and ESXi 9.1 (released 12 May 2026) on 12 August 2028. vSphere 8 was the last version with a perpetual licence and the last standalone Standard and Enterprise Plus tier.

What the Proxmox bill covers

Proxmox VE is AGPLv3, the feature set is identical in every tier, and the subscription pays for support. Prices per populated CPU socket and year, net, from the vendor page on 01 October 2026: Premium 1,100 euro with a two-hour response time on severe outages, Standard 550 euro, Basic 370 euro. These points are missing from most calculations:

  • Every node needs its own subscription, and every node has to sit at the same level.
  • The backup server is billed per machine rather than per socket: 4,480 euro Premium, 2,240 euro Standard, 1,120 euro Basic, net per year. At the top tier, the backup subscription costs four times a single hypervisor subscription.
  • Basic includes three tickets per year. On a cluster in production that is a deliberate decision with a price attached.
  • From 19 October 2026 the vendor plans 24/7 support in the Premium tier, announced on 02 September 2026 together with a new North American office and, by its own account, 2.3 million active PVE servers. That is the counterweight to the claim, common in the German trade press until summer 2025, that Proxmox had no enterprise support structure. The article carrying that claim was corrected on 23 June 2025.

Four items nobody budgets

Quorum. An HA cluster needs at least three votes, and a two-node cluster needs an additional QDevice guest. A node without quorum fences itself out after roughly one minute, and the cluster waits about two minutes before deciding again. There is no equivalent of vSphere Fault Tolerance, where two machines run in lockstep, according to vendor documentation as of 2025; work on it is under way. The corosync network and fencing details are in the HA post.

Placement. The dynamic load balancer that arrived with Proxmox VE 9.2 moves only guests under HA management. Guests without HA stay where they are, and affinity rules apply in the HA context as well. What is still missing is observability into what the scheduler does. On a cluster with many non-HA guests that means placement stays manual work.

Storage is the actual project of the switch. Ceph is the recommended line, SAN and NAS setups come with caveats, and a hand-built clustered filesystem is unsupported per the documentation and explicitly outside enterprise support. Anyone staying with ZFS has to plan memory for the ARC using the documented rule: at least 2 GiB base plus 1 GiB per terabyte of pool, so around 10 GiB on an 8 TiB pool, with the note that I/O performance depends on it. For guests on ZVOLs, mirrors beat RAID-Z groups, because RAID-Z suffers write amplification on small writes.

Alerting. After the move somebody has to own monitoring of replication and backup. Our own experience is a warning here: an evaluation of replication values produced an alert level keyed to a list holding 35 days of archive entries, so it sat in an alarm state almost permanently, and a restart of container services with roughly forty seconds without a cache produced a false alarm while replication had no problem at all. This is what happens when the old product shipped a dashboard that nobody replaced with a real check. The storage side is covered in part three of the storage series.

The move itself, state of 2026

The native ESXi importer arrived with Proxmox VE 8.2 (24 April 2024) and is tested from ESXi 6.5 through 8.0 per the wiki. The documented limits decide the schedule:

  • Disks on vSAN cannot be imported and have to be moved first. Encrypted guests and the vTPM state do not come across either, which for BitLocker means having the recovery keys ready.
  • Snapshots on a guest slow the import considerably. In bulk runs the ESXi API limits apply: four concurrent connections, a 128 MiB read-ahead buffer per disk, and a recommendation of at most four guest disks per operation.
  • If a live import fails, everything written since the start is lost. A test clone of a running guest therefore needs the usual clean-up first: address, agent identities and any access material that comes along with the clone have to be disabled before it boots.
  • Windows needs VirtIO in the initramfs or the guest will not boot after import. It is already in the 2025 post and still holds.

The current release is Proxmox VE 9.2 from 21 May 2026, with Debian Trixie, kernel 7.0, QEMU 11.0, ZFS 2.4 and Ceph Squid. On 05 August 2026 the first arm64 build appeared; it requires UEFI and ACPI, so no Raspberry Pi, and live migration stays limited to the same architecture.

Who should stay

Gartner put the counter-position on 11 May 2026: no vendor delivers a like-for-like replacement for VCF, and those who leave end up with more complex and less capable infrastructure, because dependencies force them to stack several substitute products. In the server virtualization marketplace from the same period, Proxmox is listed as a niche player, with missing certifications for top-tier enterprise applications and a small headcount as the reasons. Broadcom meanwhile looks downward again: on 02 September 2026 a new vSphere Standard was announced with a target size around 128 cores, with details promised at Explore on Tour in Frankfurt on 13 and 14 October 2026. Anyone who needs a two-site stretched cluster, mature DRS, NSX, or a vendor certification for a top-tier application is better off negotiating the price and running the clock to October 2027.

Three decisions you can make today

  • Count cores and sockets per host today and put two calculations side by side: VMware per core against Proxmox per socket, using your own socket count. The tipping point is a socket count. Philosophy does not decide it.
  • Put your own renewal date next to a support end date. On ESXi 7.x general support ended on 02 October 2025, and at that point the discussion is no longer voluntary.
  • Name the person who owns alerting for replication and backup before the first guest moves.

If you want this calculation done for your environment, including a migration plan and a test clone: contact. The start is one hour with a host list, a count of cores, and the next renewal date.

Further sources

Questions?
Does Broadcom publish an official price list for vSphere?+

No public list exists. Documented are the mechanics from the vendor knowledge base (billing per core, with a minimum of 16 physical cores per CPU) and the numbers a distributor and a systems house published for their own lists in September 2026. If you bring a number from a tender, you are comparing quotes against quotes.

How many subscriptions does a Proxmox cluster need?+

One per populated CPU socket, for every node in the cluster, all at the same support level. A cluster of four nodes with two sockets each needs eight subscriptions. Proxmox Backup Server comes on top and is billed per backup machine rather than per socket.

How big is the difference in euro?+

The clearest worked example available comes from a systems house page for three hosts with one socket and 16 cores each over five years, net prices from reseller lists in September 2026: VMware Cloud Foundation 86,938 euro, VVF 46,260 euro, Proxmox with three standard subscriptions plus a basic backup server 13,850 euro, and 5,550 euro in the cheapest case without the backup subscription. Add the migration work per guest, the Linux and ZFS knowledge, and the new runbooks, which that same worked example states openly.