Microsoft's upgrade-paths page says that, as of Windows Server 2025, nonclustered systems can move up as many as four versions at a time, and it marks 2016 → 2025 as a supported route: direct, in a single pass. Three sentences further down, in that same paragraph, one line takes it away from exactly the machines you cannot stop: if it sits in a failover cluster, you move one version at a time. Your Hyper-V cluster goes 2016 → 2019 → 2022 → 2025. That is three full passes, not one. And the last cumulative update for 2016 ships on 12 January 2027: as of 18 November, 55 days away.
We have been running customer infrastructure since 2015 and we do not resell anybody's licences, so what follows is not an invitation to buy Windows Server 2025. It is a reading of four pages of Microsoft's own documentation, placed side by side, and what comes out of crossing them with a calendar.
The date you have been told, and the one in the table
Everybody repeats 12 January 2027, including the Windows Server blog. Microsoft's lifecycle table, which is the authoritative record, says something else: end of extended support at 1/13/2027 6:59:59 AM, Pacific time. Both are true and they say different things.
- 112 January 2027 is the second Tuesday of the month. It is the last cumulative update you will receive. Whatever is not patched by that package will not be.
- 2On the morning of the 13th, Pacific time, the clock runs out. In the same table, ESU Year 1 starts that very 13th at 8:00. There is no overlap to lean on: either you have ESU contracted and provisioned, or there is a gap.
The date we plan against is the 12th, because that is the last patch. The 13th matters for the invoice, and we already wrote about that invoice's fine print here: Windows Server 2016 ESU is billed from 12 January, whenever you buy it.
The promise: 2016 to 2025 in a single pass
The good news is real and worth saying before the bad. Up to Windows Server 2022, an in-place upgrade could skip two versions at most: from 2016 you reached 2022 and that was it. With 2025 the limit went up to four —which is why a 2012 R2 box now goes straight to 2025— and the supported-paths table for installation media marks Yes on 2016 → 2019, 2016 → 2022 and 2016 → 2025. The jump from 2016 is three versions, so it sits comfortably inside that limit: one pass and done.
For a standalone server —a line-of-business application, a file server, a licence server— that means one pass and one long reboot. It is the cheapest route on the table and, if that is your estate, the rest of this article is surplus. Do the inventory, book the afternoons and work through them machine by machine.
The line that cancels it if the machine is clustered
On the same page, in the paragraph that introduces that table and immediately behind the promise, Microsoft writes this: "If you're using a Cluster OS rolling upgrade, you can only upgrade one version at a time." Further up, among the factors for choosing a method, it had already said it in one line: "Cluster rolling upgrades can only advance one version at a time." And on the dedicated page for that procedure it spells it out with an example it did not pick at random: "To upgrade across multiple versions, such as from Windows Server 2016 to Windows Server 2025, run the upgrade sequentially (first to Windows Server 2019, then to Windows Server 2022, and finally to Windows Server 2025), or migrate to a new cluster."
The documentation's own example is your case. Each pass ends with Update-ClusterFunctionalLevel, which raises the cluster functional level —9 is 2016, 10 is 2019, 11 is 2022, 12 is 2025— and warns on screen: "You cannot undo this operation." Three passes are three of those points of no return, with three rounds of validation in between.
What does not fit is not the hours, it is the windows
While a pass is half done, the cluster runs two operating system versions at once. Microsoft calls that mixed-OS mode and puts a recommended limit on it: "We recommend moving through the cluster upgrade process within four weeks because some cluster features aren't optimized for clusters that run two different OS versions."
This is where the sum is easy to get wrong, and we nearly did. Those four weeks are a ceiling, not a duration: they do not add up. Microsoft in fact says a Hyper-V or Scale-out File Server cluster can be upgraded "with no downtime in less than four hours total". Three four-hour passes fit into 55 days with absurd room to spare. The hours are never the problem.
What does not fit are the windows, which is a different thing. Each pass is a change with its approval, its notice to users, its night shift and, above all, its validation afterwards: checking that virtual machines move between nodes, that backups still complete, that failover responds. You do not start the next one until that is confirmed, because otherwise you stack two changes without knowing which broke what. That is three chained cycles of change and validation inside eight weeks, two of them the Christmas ones. That is the squeeze, and it is a people-calendar problem, not an arithmetic one.
And it does not arrive through Windows Update
That same paths page has two tabs. The installation media one is where 2016 → 2025 lives. The Windows Update one has two rows and only two: 2019 → 2025 and 2022 → 2025. Windows Server 2016 does not appear in that table.
In operational terms: there is no cumulative update that lifts the machine on its own, and no remote push from the console. Every 2016 server gets upgraded with the ISO mounted, one at a time, with somebody in front of it. That is the line of the plan that almost never gets written down, and the one that actually decides how many afternoons you need.
The five things that break along the way (and are written down)
None of these five is forum hearsay. All five are in the official documentation —and all five on the same page, the one for the procedure— and all five show up mid-job if you have not looked them up first.
- 1The network adapter team. "LBFO teams are no longer supported with Hyper-V in Windows Server 2022 and newer." You have to break the team before upgrading and rebuild the virtual switch on SET afterwards. If your virtual switch hangs off an LBFO team, this is not a detail: it is the network for every virtual machine on that node.
- 2The virtual machine configuration version. "VM configuration versions older than 8.0 (corresponding to Windows Server 2016), can't run on Windows Server 2022." And raising it with
Update-VMVersionrequires the machine powered off —the guide frames it "during a scheduled maintenance window when you can take the VMs offline"— and warns that "this operation is not reversible". Mind the scope, because it is easy to over-worry here: machines born on 2016 are already at 8.0 and boot on 2022 and 2025 untouched. The downtime step is only mandatory if you are dragging guests from 2012 or 2012 R2, below 8.0. Raising the rest to 12.0 unlocks new features, but it is optional. Which is why this inventory column gets checked before you promise anybody a window. - 3Where your way back is. The process is reversible —you can add old-version nodes back and remove the new ones— until you raise the functional level. After that, the documentation says, "you can't go back to an earlier functional level and you can't add nodes running earlier versions of Windows Server". With one caveat the same page puts in its FAQ: if before raising the functional level you took a System State backup of the cluster database, you can perform an authoritative restore onto a node running the previous version. So a way back after the point of no return does exist, but you buy it beforehand, by taking that backup. If you did not take it, the on-screen warning is literal.
- 4Capacity with one node out. During the pass the cluster runs one node short. The guide asks you to check that you can hold the service that way and, for two-node clusters, suggests temporarily adding a node. If you have exactly two, the "no downtime" promise does not hold without borrowed hardware.
- 5The "no downtime" applies to two workloads only. Verbatim: "it's only zero-downtime for Hyper-V and Scale-out File Server clusters. Most other workloads incur some downtime (typically a couple of minutes) when they failover." If your cluster runs SQL Server or a classic file server role, there is a failover in the plan, and people need telling.
The four doors, with their disqualifying requirement
Note that the third door is written by Microsoft in the same sentence as the three passes: "or migrate to a new cluster". It is right there, with equal standing, and this far into November it is the one that has paid off most often for us, because it is the only one that does not require chaining three windows before the new year.
Where to get off
If you are going to spend passes, spend them all the way. The lifecycle dates make it plain: Windows Server 2019 ends extended support on 10 January 2029; stopping there buys two years and forces you to repeat the whole operation. Windows Server 2022 runs to 15 October 2031, but its mainstream support already ended on 14 October 2026, a month ago: it is a landing strip, not a destination —and if you do end up there, first read what hotpatching on 2022 gives you and what it does not—. Windows Server 2025 reaches 2034.
The proof that ESU is not a solution
Microsoft defines the ESU programme in its own words as "a last resort option", and there is a quick way to see what that means: ESU for Windows Server 2012 and 2012 R2 ended on 13 October 2026. Five weeks ago. The machines that were holding on with it now receive nothing, at any price, and whoever bought three years in 2023 met the same problem again, only three years later and with the hardware three years older.
ESU buys time, and sometimes buying time is exactly the right call. But it is a dated extension, not a place to settle down.
The five-column inventory
This is what we ask for before giving an opinion on any plan. Five columns, one row per machine running 2016, and an afternoon of work. With the table in front of you each machine decides itself; without it, the calendar decides on 12 January.
- ✓Is it clustered, and of what kind?
Get-Cluster | Select ClusterFunctionalLevel. A 9 is a 2016 cluster, and it is the row that governs the whole plan. - ✓What configuration version are its virtual machines on?
Get-VM | Select Name, Version. Anything below 8.0 is a scheduled outage you put in the plan now, rather than discover later. - ✓Does the virtual switch hang off an LBFO team?
Get-NetLbfoTeam. If it returns anything, you have network work on every node before you touch the operating system. - ✓Which edition is it, and does it have Software Assurance or a server subscription? PowerShell will not produce this column: whoever handles the contracts will. It decides whether ESU is even an option for that machine — and if the edition is Essentials, you already know it is not.
- ✓What role does it play, and does it still play it? The most profitable column of the five. A domain controller has its own upgrade restrictions; and more than one machine on this list is holding up something nobody uses any more.
When not to
We sell infrastructure continuity and operations, so this part runs against our own commercial interest. We say it because it is right more often than not:
- ✗Do not migrate what you can switch off. If it is a standalone server with a role nobody uses any more, or a file server whose data fits on a machine you have already upgraded, do not plan a migration: plan a decommission. The cheapest migration is the one you do not do, and in 55 days that inventory column is worth more than any other.
- ✗Do not put a cluster into mixed-OS mode in December. Starting a pass in the week of the 22nd is how a planned change becomes an incident: half the team on holiday, the backup provider on skeleton cover and a cluster running two operating systems at once. If it does not fit before the break, it goes in January and you take the gap with your eyes open.
- ✗And if it is one or two standalone machines with Software Assurance, a year of ESU may well beat a race-the-clock migration with no testing window. That is less work for us and we say it anyway: a botched December migration costs more than the extension.
The plan almost everybody has in mind
The deadline is not the problem. The problem is that the plan almost everyone has in mind —"we will upgrade it in January"— was written for a machine that is not in a cluster. If yours is, the route promised on the front page is not yours, and that is better discovered today than on 12 January. An afternoon of inventory, the decision made in the first week of December, and the work done when there are people around.
Sources (verified at origin): the up-to-four-versions limit for nonclustered systems, the "only one version at a time" for clusters, both supported-paths tables (installation media and Windows Update) and the upgrade restrictions (VHD boot, Storage Server, Core/Desktop, domain controller, disabling NIC Teaming beforehand) — "Plan Your Windows Server Upgrade Path", Microsoft Learn. The 2016 → 2019 → 2022 → 2025 example with its "or migrate to a new cluster", mixed-OS mode and its four weeks, the under-four-hours figure, the "only zero-downtime for Hyper-V and SOFS", LBFO teams no longer supported with Hyper-V on 2022 and newer, VM configuration version 8.0, the temporary node on two-node clusters, functional levels 9/10/11/12, the irreversibility and the authoritative restore from a System State backup — "Cluster operating system rolling upgrade", Microsoft Learn. Which configuration versions each host supports — "Upgrade virtual machine version", Microsoft Learn. The end-of-support dates, each from its own Microsoft Lifecycle page: 2016 (1/13/2027 6:59:59 AM PT and the three ESU years through 1/8/2030), 2019 (1/10/2029), 2022 (mainstream 10/14/2026, extended 10/15/2031) and 2025. 12 January 2027 is the date given by the Windows Server blog itself. ESU as "a last resort option" and the end of 2012/2012 R2 ESU on 10/13/2026 — "Overview of Extended Security Updates", Microsoft Learn. That 2016 ESU covers Standard and Datacenter only, and that it "generally requires Software Assurance through a Volume Licensing program (machines licensed through SPLA or a Server Subscription do not)" — the GA announcement for 2016 ESU enabled by Azure Arc. The 55-day count is ours from the dates cited, as is the reading that the squeeze is the windows rather than the hours. This post carries no figures from our own operation because it is what it says it is: a reading of the documentation crossed with a calendar. Photograph: Cyber-Complex Foundation, CC BY-SA 4.0 via Wikimedia Commons. We are not Microsoft resellers, nor resellers of any other platform, and we do not sell anybody's licences.
Do you know how many of your 2016 machines are in a cluster?
At everyWAN we build the five-column inventory and say which machine goes through which door, including the one marked "switch it off". We build and run the infrastructure underneath and the continuity plans that hold it up. If you want the table done before December, get in touch.
Talk to everyWAN