In 2022, 37signals —the company behind Basecamp and HEY— was paying $3.2 million a year for cloud. It bought about $700,000 of Dell servers, recouped them within the first year, and in its first full year off the cloud saved almost $2 million. In 2025 it announced the final phase: moving the 18 petabytes it still had in S3 to its own arrays and deleting the AWS account. Meanwhile, worldwide public cloud spending keeps growing at a 20% annual clip, according to Gartner. Both things are true at once, and they don't contradict each other: the difference is who has done the math and who renews out of inertia.
We live in both worlds: we run colocation with our own hardware in several datacenters —with Proxmox and Ceph in production on top— and we manage client infrastructure that lives in the cloud, day in and day out, when that's the right call. So this post is not an anti-cloud manifesto. It's the math we do before recommending either.
The exodus that isn't one either
Barclays' 2024 CIO survey produced a striking headline: more than 80% planned to move at least some workloads from the public cloud back to private or on-premises environments. IDC talks about more than 70% of enterprises repatriating something. It sounds like a stampede, but the fine print defuses the headline: those statistics count companies, not workloads. Move one application out of a hundred and you already count towards the percentage. And at the same time, Gartner forecasts worldwide public cloud spending to grow 21.5% in 2025, to $723 billion.
The same thing happened when we looked at the VMware migration data: the mass exodus is nowhere to be found; what you find is selective reallocation. Same with the cloud: nobody serious is "leaving the cloud". They're pulling out the specific workloads that overpay, and leaving alone the ones that are fine where they are. The useful question is no longer "cloud, yes or no?" but "which workload is overpaying where it sits today?".
The wrong workload: why your bill won't come down
Cloud pricing charges for one very specific thing: the option to grow tomorrow. Elasticity, provisioning in minutes, pay only for what you use. It's an excellent product… for workloads that actually exercise that option. The problem is that a good chunk of what companies keep in the cloud never does: the ERP, the same old database, the file server, the VMs that have been consuming the same resources every month for three years. For that stable, predictable workload, every monthly bill includes the price of a flexibility nobody is using. It's like renting an off-roader, every month, to drive on the motorway.
37signals' storage case illustrates it with round numbers:
- →18 PB on Amazon S3: about $1.5 million a year, every year, for as long as the data sits there.
- →The same capacity on their own arrays (Pure Storage, dual datacenter): ~$1.5 million once, plus under $200,000 a year to operate.
- →In other words: a single year of the S3 bill paid for the entire hardware. From year two onwards, an estimated $1.3 million in annual savings.
Bulk storage is the extreme case because it only does one thing: grow. But the pattern repeats with stable compute: 37signals recouped its $700,000 of servers within the first year. A well-sized server works for five years or more; its cloud rental is paid monthly at "you might need double tomorrow" prices. Honest caveat: 37signals has scale, an excellent operations team and very well-defined workloads. Their math is not your math. But the method is: what they did was arithmetic, not an act of faith, and that's exactly the replicable part.
The exit barrier is no longer the excuse
For years, the lock was the exit toll: egress fees made moving data out at scale cost serious money, and that number killed many analyses before they started. That changed in 2024, pushed by the European Data Act: Google Cloud dropped transfer fees for customers leaving (January), and AWS and Azure followed in March. It's not automatic —you have to request it and meet conditions— but the structural toll is gone: AWS waived about $250,000 of egress fees for 37signals' 18 PB move. If your last exit analysis predates 2024, it's outdated on one of its biggest line items.
When the cloud wins (and we say this while selling colocation)
Some workloads are a bad idea on your own iron, and saying so is part of the job:
- ✓Real elasticity: seasonal peaks, campaigns, workloads that need 10× last month's resources. There the growth option is actually exercised, and paying for it makes sense.
- ✓Projects with uncertain lifespans: if you don't know whether the product will exist in a year, don't buy five years of iron.
- ✓Managed services you don't want to operate: if a managed database saves you a role you don't have and don't want to hire, that part of the bill is money well spent.
- ✓Global presence: serving users on three continents with your own iron demands a scale almost no SME has or needs to have.
And the condition that cuts across everything: iron needs someone who cares for it. If nobody is going to size it, patch it and respond when a disk fails, don't buy disks — or hire someone to do it for you, and put that contract into the math. We do exactly that for clients; that's how we know it's a real line item, not a footnote.
The five-year math, line by line
The serious comparison isn't "instance price vs server price". It's this one, over five years and with your real workload:
- 1.Cloud side: your real monthly bill ×60, applying the data growth you already observe (storage never shrinks) and the egress you already pay.
- 2.Iron side: CAPEX for servers and storage sized to your real peak (not the fear peak), plus one refresh if the lifecycle lands within the five years.
- 3.The colocation fee: rack space, power billed on real consumption, cooling, physical security and the datacenter's remote hands.
- 4.Connectivity: transit, peering and the lines to your sites. Who your provider is matters here: a datacenter with a carrier network solves in one contract what otherwise takes three.
- 5.People: who operates, patches, monitors and answers at three in the morning. Your own payroll or a management contract — but always in the math.
- 6.The capacity buffer: in the cloud the reserve is elastic; with iron you buy it upfront. If your real peak is 10× your average, this line item can flip the whole calculation — and it's fine if it does.
With everything included, the reading is simple: if the math favours iron with a margin, don't renew out of inertia. If it comes out even or close, stay where you are — migrating costs money too, and a migration that saves 10% doesn't pay for its own risk. We won't give you a magic savings percentage here because there isn't one: the math is done with your bill, not with a sales deck's. What we will say is that for stable, storage-heavy workloads, done honestly, it favours iron more often than the "cloud-first" of ten years ago took for granted.
Colocation: the middle ground the debate skips
The debate is usually framed as "cloud vs building your own server room", and framed that way it's rigged: the room next to the kitchen, with its consumer UPS and wall-mounted air-con unit, loses to anything. The serious comparison is colocation: your hardware in a professional datacenter, with redundant power, cooling, physical security and carrier connectivity — without building or maintaining any of that. You keep what makes the math win (depreciable iron, sized to your workload) and outsource what ruins it (the building, the power, the cooling).
It's our own model: we run our own hardware across several datacenters, with Proxmox and Ceph in production on top, and a carrier network with BGP, transit and peering (the looking glass is public: lg.everywan.com). That same infrastructure is what we put underneath our clients' workloads — and when a client's math favours staying in the cloud, we tell them just the same. We don't resell anyone's platform; we take no commission on the outcome of the math.
In short
The cloud isn't a scam and iron isn't nostalgia. The cloud charges for elasticity: if you use it, it's a good deal; if you don't, you're paying every month for insurance you never claim. The mass exodus doesn't exist — what exists is people doing the math workload by workload and moving the ones that overpay, now that leaving no longer carries a toll. The worst infrastructure decision isn't cloud or colocation: it's the renewal signed without doing the math.
Sources (verified): 37signals figures ($3.2M/year of cloud in 2022, $700,000 of Dell recouped in 2023, ~$2M saved in 2024, 18 PB and $1.5M/year on S3, ~$1.5M arrays and <$200,000/year to operate, $250,000 of egress waived) — The Register (Oct 2024) and The Register (May 2025); Barclays CIO survey (2024) and IDC repatriation data, with their caveats — Channelnomics; 2025 public cloud spending forecast ($723.4B, +21.5%) — Gartner (Nov 2024); removal of exit egress fees (Google Jan 2024; AWS Mar 2024, European Data Act) — TechCrunch; Azure (Mar 13, 2024) — The Register.
Cloud renewal coming up and the bill keeps growing?
At everyWAN we do the full math —cloud, colocation or hybrid— with your real workload and every line item on the table. If it turns out you should stay where you are, we'll tell you that too.
Talk to everyWAN
Microsoft 365
ES 
CAT 
EN