The cloud got more expensive but so did the alternative
Your cloud bill is high, and someone has come to you with a link to the 37signals cloud exit and a suggestion that you look into doing the same. At least in part.
It is a fair suggestion, and for a lot of use cases it had a clear answer during a period of time. This year the answer shifted. Not because of fundamental changes in the inner workings of cloud computing, but because of what is happening inside a handful of chip factories.
Here is the short version:
- There are three ways out of a large cloud bill, not two as most sources claim. Spend less where you are, buy your own hardware, or move to a different provider.
- Buying your own hardware is the route many write about, and it is the one that got harder this year. AI has made server memory scarce and expensive.
- That squeeze reaches anyone who provides cloud services to you, us European providers included.
- The third route, moving to a smaller or more local provider, gets the least attention and is growing quickly in Europe.
Is the case for a cloud exit a good one?
There is not a better answer than that it depends.
The strongest, most transparent, and most talked about voice that leads the argument to leave practically belongs to David Heinemeier Hansson these days. He is the co-owner of 37signals, the company behind Basecamp and HEY. In 2022 he announced they were leaving Amazon1, Then he did something unusual. He published the numbers as they came in, year after year. That included the parts that took longer than expected to pan out.
The savings
By their own account in early 2023, 37signals expected to save around seven million dollars over five years. Many conversations on leaving end with the same realisation. It is a zero sum game, because you need staff to run your own hardware. They moved seven applications onto their own hardware in 2023 without hiring anyone new2.
Conservative estimates
Four years in, their estimates turned out to be conservative. The initial five year, seven million dollar estimate was revised upwards and landed on ten million dollars. All in all, they describe it as cutting their infrastructure costs by between half and two-thirds.
The total infrastructure bill went from $3.2 million a year to, as Hansson puts it, well under a million. The operations team is still the same size.
The caveats
Those numbers are not in dispute, and Hansson has been more open about both the arithmetic and caveats than most companies are about anything. So let’s take a look into the caveats that applied to 37signals, and what made this work out in their favour. Who knows, perhaps they apply to your organisation too.
In short, three things made their move cheaper than it would be for most companies. The people they already employed, the space they already had, and the timing of their hardware cycle.
People
They had engineers who already knew how to do the work. Hansson has pointed out that most of those skills transfer directly between running someone else’s hardware and running your own3. This is why they added nobody to the team. If your team would need to learn this from scratch, or is already at capacity, that saving is not available to you.
Data centre
He is also straightforward about the fact that 37signals already had data centre racks and power headroom to run the machines. A company starting from nothing would need to pay for that space and equipment as well. Even though Hansson reckons that it is cheap next to cloud, it’s not nothing.
Hardware
The savings are estimated and compared over five years, and Hansson expects the hardware to last five to seven. Replacement therefore sits mostly outside the window. That makes the five-year figure the cheapest stretch of owning hardware: you buy once at the start, and the next big cheque falls due after the comparison ends. When it does, they will be buying into whatever the market looks like then.
A point for clarity
Just to be perfectly clear. This isn’t an attempt to dissect their remarkable story and to turn the caveats into something that should deter you and your case. It’s just some things to consider in your own thought process and potential exit exercise.
But then the price of a server changed
The cause sits a long way upstream, in chip manufacturing. AI needs a specific kind of memory, and making it uses the same production capacity as the ordinary memory that goes into ordinary servers. Three companies manufacture almost all of the world’s server memory. And when those companies started allocating more production capacity towards the more profitable, AI-specific memory, there became less of everything else to go around4.
Naturally, the price of ordinary memory went up for everyone.
Prices projected to rise by six times
Octave Klaba, the founder of OVHcloud, wrote a clear, public account of what this does to an IaaS company. The account is a rarity in and of itself. It is also more interesting than most, because OVHcloud assembles its own servers and has a global supply chain.
He put the rise in memory and storage costs at between 15 and 300 percent against 2025 levels. How much depends on the configuration. He has since reported that memory costs OVHcloud six times what it did a year earlier, with storage following the same trajectory. He does not expect prices to return to their “normal” levels before 20285.
Ordering ahead without knowing the price
The part worth reading twice is not about price at all. To be sure of getting the memory and disks they need, he wrote, providers now order up to twelve months ahead. They do it without knowing what they will pay. The final price is confirmed one to two months after delivery. He described a market where visibility rarely lasts longer than a week or two.
An account like that is more use to a customer than a quiet adjustment to a price list.
Old advice in a new reality
That is what buying hardware looks like at the moment. And again, OVHcloud is doing it while assembling its own servers, with a global supply chain behind it. We don’t assemble our own, and you would be doing it without either.
The point here is that the advice that came out of the 37signals cloud exit was formed when servers were cheap and available. The story has survived and echoed for a number of years. Right now servers are neither cheap nor available, and the shortage is expected to last into 20276.
This reaches us too
Our cloud services don’t run on hopes and dreams but the very same kind of hardware, so we do not float above any of this.
Hetzner raised its cloud server prices by 30 to 37 percent in Germany and Finland on 1 April. It told customers it could no longer absorb what it described as drastic cost increases across the IT sector7. OVHcloud raised its own prices from 1 April, holding the average increase to between 9 and 11 percent on services deployed from 2026 onwards5. Both are European. Both buy from the same suppliers we all buy from.
There is a more uncomfortable aspect of this, and it belongs in an honest article. Size helps in a shortage, and the hyperscalers have it. They buy years ahead, in volumes no European provider can match, which locks in prices long before a squeeze arrives. And this seems to show in their price lists too: Amazon’s published price increases this year have landed on reserved GPU capacity rather than on ordinary compute8. Smaller providers buy much nearer to the market, and pay what the market is asking.
In this particular respect the advantage is theirs, not ours.
Cleura buys hardware in the same market. We have not made any decision about our own prices as of right now. But we are not going to pretend that the pressure everyone else is describing somehow stops at our door.
The statistic everyone quotes does not say what people think
If you have read anything about companies leaving the cloud, you have probably seen the number. Somewhere between 80 and 86 percent of technology leaders plan to move at least some workloads back from public cloud. It appears in nearly every article on the subject, including articles by companies who would like to sell you a physical server.
The German technology publication Digital Chiefs made the point that undoes it9. The survey behind that number, run by Barclays in late 2024, asked whether companies plan to move any workloads back from the cloud. A company moving one database because a regulator asked answers yes. So does a company moving a tenth of what it runs. So does a company leaving entirely.
All three land in the same line of the results. The number tells you that things are moving. It tells you very little about why, where they are going, or how much of anything is going there. On the same analysis, the share of companies pulling everything out of public cloud is in the single digits.
This matters because the number gets used to describe a general retreat to the server room. Some companies are doing exactly that, and 37signals is one of them. Most are moving one thing, for one reason, and leaving the rest where it is.
Where companies are actually going
Not, for the most part, into their own server rooms.
Gartner expects spending on sovereign cloud IaaS to grow by more than a third this year. Europe is among the fastest-growing regions anywhere, and on course to overtake North America in 2027. Their analysts also expect a share of workloads currently running on global platforms to move to local providers instead10.
That figure needs one qualification. Not all of that money goes to European companies. The largest American platforms sell what they claim to be EU sovereign cloud offerings of their own. They therefore take a share of what is merely moving back to Europe geographically. What the number shows is where the demand is pointing, not who ends up banking it in the end.
So, this is the third option, which the debate tends to skip, because it frames the choice as a fight between two camps. Cloud, or own it yourself. Let’s not forget that there is a good deal of room between those.
A cloud exit does not have to mean buying hardware. Moving to a smaller or more local provider does not require you to buy anything, hire anyone, or find somewhere to put a rack. It does not leave you queuing for memory that is not being made fast enough. It is still a migration, with all the work that implies, but it is the kind of work your team has done before.
Which of the three paths should you take?
You don’t need a spreadsheet to get most of the way to an answer. You just need honest answers to a few questions.
One thing before the questions. This is not one decision for everything you run. Most organisations end up with a mix, and the sensible unit of decision is a single workload, not a whole company. So answer these about one specific system, and then run them again for the next one.
Five questions worth answering
Is your demand steady, or does it spike? If your traffic is roughly the same on a Tuesday in February as it is on your busiest day of the year, you are paying for flexibility you never use. If it spikes hard and without warning, that flexibility is the product, and it is worth the money.
Do you already have people who can run hardware? Not could learn. Already can, and would still have time to do their actual jobs.
Do you need machines this year? If you are buying into the current market, price the hardware before you price anything else, and get the quote in writing with a date on it. Quotes are not lasting long.
Is the location of your data a legal question for you? For some organisations this decides the matter before cost enters the conversation. If that is you, the choice is narrower than it looks, and it is worth reading up on where your data actually sits and who can reach it.
What is actually driving the bill? Before you conclude that cloud is the problem, find out what you are being charged for. In our experience a surprising share of a shocking invoice is not compute at all. It is the charges around the edges, and the cost of getting your own data back out.
If a cloud exit is the answer
If the honest answers point at owning your own hardware, then own it. That is a legitimate outcome and we would rather you reached it with clear eyes than talked yourself into something that does not fit.
Ask us for a second opinion
If you want a second opinion on where a particular workload belongs, ask us. Tell us what you are running and what it is costing you. We will tell you plainly whether we think we can help. Sometimes the answer is no, and it saves everyone time if we say so early.
What we would say about ourselves
Three things, since we have spent this article asking you to be sceptical.
- We are not the cheapest place to rent a virtual machine. If your only criterion is the lowest number per month, you can find a lower number.
- Our platform is built on OpenStack, and it does not do everything that the largest hyperscaler platforms do. We keep a list of what it does not do, in public, in our documentation11, because you should find that out before you migrate rather than after.
- And the hardware market is doing to us what it is doing to everyone else.
What we would say for ourselves is that we run in Sweden and Germany, under European law. We run on open source that you can both scrutinise and leave if you want to. For a certain kind of organisation that combination answers questions a lower invoice does not.
Yes, the cloud got more expensive. So did the alternative. The useful question is no longer whether to cloud, but who to cloud with, and what you are actually paying for.
Sources
- David Heinemeier Hansson, “Why we’re leaving the cloud”, 37signals, October 2022. world.hey.com ↩
- 37signals, “We left the cloud”. basecamp.com ↩
- David Heinemeier Hansson, “Sovereign clouds”, 37signals, May 2023. world.hey.com ↩
- Bloomberg, “Why the AI-driven memory chip shortage is making technology more expensive”, July 2026. bloomberg.com ↩
- Octave Klaba, “Pricing evolution of Public Cloud, Bare Metal and VPS at OVHcloud”, OVHcloud, March 2026. us.ovhcloud.com ↩
- TrendForce, “Diverging memory market outlook in 2027 as DRAM supply remains tight while NAND flash supply conditions ease”, 30 July 2026. trendforce.com ↩
- Tom’s Hardware, “German data center giant hikes prices up to 37% starting April 1”, February 2026, reporting Hetzner’s own price change notice. tomshardware.com ↩
- InfoQ, “AWS hikes EC2 Capacity Block rates by 15% in uniform ML pricing adjustment”, January 2026. A second increase of roughly 20 percent followed on 1 July 2026. infoq.com ↩
- Benedikt Langer, “Cloud Repatriation 2026 Is a Statistical Illusion – Those Who Fall for It Miss the Real Architectural Shift”, Digital Chiefs, April 2026. digital-chiefs.de ↩
- Gartner, “Gartner says worldwide sovereign cloud IaaS spending will total $80 billion in 2026”, February 2026. gartner.com ↩
- Cleura documentation, known limitations. docs.cleura.cloud ↩
Özgür Bal
Marketing Manager · Cleura AB
Özgür is the Marketing Manager at Cleura. He works across brand, communications, content and go-to-market strategy, with a particular focus on cloud infrastructure and digital sovereignty.
