650 servers became 400: where the savings from leaving VMware actually come from

A tidy headline has been going round this week: Topgolf, an entertainment chain with more than a hundred venues, says it avoided or eliminated more than $11.3 million in combined infrastructure costs after moving some 12,000 virtual machines off VMware. The figure is attributed with a name and a job title, the trade press picked it up in mid-October and it will almost certainly land on a board slide this week. Possibly yours.

The number is not the problem. The problem is the word combined, which travels inside the figure and falls out of the headline. The same report carries the two lines that do not fit on a front page: they went from 650 servers to roughly 400, and the standard cluster at each venue went from six nodes with 35 TB to three nodes with 5 TB. Several different invoices live in there, and the hypervisor is only one of them.

The facts, as published

Worth putting them side by side, because taken separately each one tells a different story:

• Some 12,000 virtual machines moved from VMware to VergeOS, spread across more than a hundred venues, offices and in-house data centres.
• The published figure is "avoided or eliminated more than $11.3 million in combined infrastructure costs". It is not a stated annual saving, nor an invoice compared month by month.
• The trigger: they projected VMware costs would rise 400% after Broadcom's repackaging and licensing changes, with perpetual licences gone.
• They replaced about 650 VxRail hosts with roughly 400 Dell PowerEdge servers over six months.
• The per-venue template went from six-node clusters with 35 TB to three-node clusters with 5 TB.
• A stated reduction of 87% in comparative hardware spend and 77% in software.
• And what does not fit in a headline: twenty venues were still on VxRail and VMware Cloud Foundation when the story ran, with the full exit expected in the second week of December.
• One more fact, and it changes the whole reading: what they replaced was not just the hypervisor. According to the new platform's own vendor, what came out was "a mix of VMware software, VxRail hardware and third-party backup", and specifically "the company replaced Rubrik with VergeOS's built-in data protection". An entire backup product left the invoice.

There is one more sentence, from Topgolf's own global infrastructure manager, Scott Forehand, and it is the one that should be bolded on any slide using this story: the figure covers combined infrastructure costs, not VMware licensing alone. His words, not ours. Which means the headline is not "swapping hypervisor saves $11.3 million". It is "rebuilding the infrastructure — hardware, storage and backup included — avoided or eliminated $11.3 million, and along the way we changed hypervisor".

"Avoided" is not "no longer paid"

The exact words in the story are avoided or eliminated, together, in the same phrase and with no breakdown. And an avoided cost and an eliminated cost are not the same thing, nor do they carry the same weight in a board meeting. Avoided is a budget you never sign: the renewal they asked for and you did not pay. Eliminated is money that was leaving the till and stops leaving it. The first is legitimate — if they ask for 400% more and you do not pay it, you have avoided a real cost — but it does not show up in the P&L as income, and somebody at the table will ask.

Our criterion, offered as criterion and not as fact: when you take a migration to the board, use two separate lines. One, the euros per year that stop leaving the bank account. The other, the euros per year you are not going to sign. They add up, but they are not the same currency, and mixing them is the fastest way to lose credibility at the second meeting.

The arithmetic the headline leaves out

The hardware figures are public, so anyone can do the subtraction. We do it here and we label it as ours:

What changed	Before	After	Change
Physical servers	~650	~400	-38 %
Nodes per venue	6	3	-50 %
Capacity per venue cluster	35 TB	5 TB	-86 %
Comparative hardware spend (stated)	—	—	-87 %

The temptation is to take the last two rows and close the case: the 87% cut in hardware spend looks a great deal like the 86% of installed capacity that left each venue. That would be a stretch. They are percentages over euro bases we do not know, and the same case reports -77% in software: with what has been published there is no way to decide which line item weighed most. What you can do is count them, which is a different exercise.

Counted, what has been published contains at least four different bills inside those $11.3 million: (1) VMware licensing; (2) the hardware, some 250 fewer servers and the jump from a branded appliance — VxRail — to commodity servers; (3) the third-party backup product, which disappeared entirely; and (4) the installed storage capacity per venue, falling from 35 TB to 5 TB. Items (2) and (4) land on the same hardware invoice — in VxRail the disk sits inside the node — but they are two different decisions: how many nodes you buy and how much capacity you put in each. Three of the four are not fixed by swapping the hypervisor's licence file.

The part that does not transfer, said by the side that sells it

Our first reading of the 35 TB that end up as 5 was the obvious one for anyone who has spent years opening other people's racks: there was capacity in there that was not needed. It does not hold up in full, and it is better to say so before somebody else does. The new platform's vendor credits part of that drop to its global inline deduplication, for which it claims ratios of "4:1 or greater". Installed capacity per venue, on the other hand, falls 7:1 — 35 over 5 — and the two figures are not the same kind of quantity: 35 TB and 5 TB are installed capacity, whereas a deduplication ratio compares logical data against physical data. You cannot subtract one from the other, or divide them. And the claimed ratio is "4:1 or greater", an open floor: it could account for the entire drop and leave no slack at all. The figure you would need — how much real data sat inside those 35 TB — has never been published. With what is public, the credit cannot be apportioned.

This has a practical and not especially pleasant consequence for anyone already scaling the numbers down: part of that case's saving travels with the platform they picked, and it is not waiting on the far side of just any migration. If your destination does not deduplicate the same way, those terabytes do not evaporate. We are the ones saying it, and we make a living from migrating: before you put somebody else's percentage into your spreadsheet, you need to know which of the four bills you are touching.

Why your sum looks nothing like theirs: the fixed floor

Twelve thousand machines and a hundred venues. You probably have forty machines, two hosts and a rack. The temptation is to scale the saving down proportionally, and that is precisely the error, because the saving scales with the number of machines and the project cost does not. A set of line items costs about the same with 40 machines as with 12,000. We call it the fixed floor of a migration:

• The real inventory. Which machine exists, who uses it, what breaks if it goes off on a Tuesday at eleven. It is the line item nobody invoices and the one that decides all the others. Skip it and you are not migrating: you are copying your mess to a new address.
• The network. VLANs, firewall rules, routes, internal DNS, everything a vSwitch used to handle and now does not. You rebuild it once, whether you have two hosts or two hundred.
• Backup. Changing platform changes the agent, the proxy, the credentials and sometimes the snapshot method. And it forces you to time a full restore of your largest machine again, which is the number nobody knows by heart and the only one that tells you whether the plan holds.
• The rollback plan. The step that turns a bad night into a bounded incident, and the one most often missing. We break it down in the rollback plan almost nobody prepares.
• Boot testing for the systems nobody has restarted in three years. Nobody knows whether that machine with three years of uptime comes back on its own, and migration day is a bad day to find out.
• The window and the business sign-off. Talking to whoever invoices, picking the night, keeping somebody awake. Shutting down forty machines costs the same as four hundred: what changes is how many nights.
• Training the team that will operate it the following Monday. If whoever is on call does not know where to check the new cluster's health, you have swapped an invoice for a risk.

What does scale with machine count is the boring part: copying disks, per-socket or per-node licences, hardware. With 12,000 machines the fixed floor is statistical noise and the saving rules. With forty, the fixed floor is the project. Which is why the useful question before asking for a quote is not "how much do I save", but a different one: how many machines do I have above that floor? If the answer is "not many", the headline figure is not talking to you.

Six months was not the finish line

The six months the story cites cover the hardware replacement; that they are not the whole migration is our own deduction, and it rests on the next figure. When it ran, twenty venues were still running on VxRail and VMware Cloud Foundation, with the final exit expected in the second week of December. In other words: a company that had decided to leave kept both platforms alive for months. Two patch cycles, two ways to restore, two places to look at three in the morning, two contracts. That cost appears in no headline, and we go into it in detail in what you will maintain twice. It is not an argument against migrating: it is a budget line, with an end date written into it.

When this headline is not your problem

If your renewal is eighteen months out, your hosts run at 80% CPU and disk, and you restore a real backup every quarter, none of this moves your calendar. You can file the story away. The headline percentages are not waiting for you on the far side of any migration: they came from removing 250 servers, changing storage system and dropping an entire backup product, and none of those three happens automatically when you migrate. What is worth looking at is the other question, the renewal one, and that is a negotiation, not a technical project.

And if the temptation is to run for the exit because the number is big, it is worth rereading the inventory of contraindications we already published: there are cases where we recommend when NOT to migrate from VMware to Proxmox, and we think the same this week as last. An eleven-million headline does not change the fact that your critical application is only certified on one platform.

What we do, and what we do not claim

We run Proxmox VE with Ceph storage in production across several data centres, we have been running VMware since its early versions, we have migrated companies from VMware to Proxmox and we have also recommended staying when that made sense. We are not resellers of any platform: we resell no licences, VMware's or anyone else's. Which is why we can say the awkward part: the headline case did not pick Proxmox. It picked another platform, VergeOS. We say it because this story is already being used to sell Proxmox migrations — ours included — and the case, read in full, does not say that.

Our conflict of interest is obvious: we charge for migrations and for infrastructure consulting. Even so, the conclusion we draw from this case takes work away from the first invoice: of the four bills inside that figure, two — the hardware you do not need and the backup product you pay for — can be reviewed without migrating anything. Measure real CPU, memory and disk utilisation per host, compare it with what you have installed, and look at the gap. If the gap is large, you already have your first saving and an argument for the renewal. If there is no gap, you have just discovered your problem is capacity and not licensing, which is an entirely different conversation.

What we do not claim: we have audited none of Topgolf's figures — they are statements by their manager reported in the trade press, and the rest comes from the platform vendor's own documentation, which is an interested party; we give no savings percentage for your company, because we know neither your utilisation nor your renewal; we do not say VergeOS, Proxmox or VMware are better than one another, because that depends on what you have to run on top; and we bounded the over-provisioning reading in the body precisely because the vendor credits part of that drop to its deduplication: nobody outside Topgolf knows the exact split.

Sources (checked 17 Oct 2026): the case figures — some 12,000 virtual machines moved from VMware to VergeOS, "avoided or eliminated more than $11.3 million in combined infrastructure costs", the projected 400% rise after Broadcom's packaging and licensing changes, the replacement of about 650 VxRail hosts with roughly 400 Dell PowerEdge servers over six months, the shift from six-node clusters with 35 TB to three-node clusters with 5 TB, the stated reductions of 87% in hardware and 77% in software, more than a hundred venues, the twenty venues still on VxRail with VMware Cloud Foundation and the full exit expected in the second week of December, and the clarification by Scott Forehand (Topgolf global infrastructure manager) that the figure covers combined infrastructure costs and not VMware licensing alone — come from the original Ars Technica report (13 October 2026), which we did not read directly, and from its coverage in Business Wisconsin, cross-checked against TokenPost and the platform vendor's public documentation — its note on the project, published when the deal was announced in 2025 and therefore earlier than the closing figures, the source for the replacement of "a mix of VMware software, VxRail hardware and third-party backup", for Rubrik being replaced by built-in data protection, and for the claimed global inline deduplication ratios of "4:1 or greater"; the case study PDF is public even though its download page asks for a form. It is interested-party documentation and we treat it as such. Ours, and we label it: the distinction between avoided and eliminated cost, the calculated percentage changes (-38% servers, -50% nodes per venue, -86% capacity per venue), our express refusal to read anything into that -86% against the stated -87% in hardware, the count of four separate bills inside the combined figure, the discarding of our own over-provisioning reading once the 7:1 in installed capacity turns out not to be comparable with the claimed 4:1 in deduplication, the fixed-floor list, the scale arithmetic and the section on when this is not your problem.