Back to blog

That "86% of CIOs are leaving the cloud" number is two years old and doesn't mean what you think

8 min read Infrastructure Cloud Hosting
On this page

If you have read anything about cloud costs this year, you have met the statistic. Eighty-six percent of CIOs plan to move workloads out of the public cloud, it says, usually in the first paragraph of an article published by a company that would very much like to rent you a dedicated server. It turns up in pitch decks, in conference talks and in at least one draft of this very post, which is how we ended up going to look for where it came from.

It comes from a Barclays CIO survey, and the figure that everyone quotes as the pulse of 2026 is, as far as anyone citing it can show, from late 2024. That alone ought to slow people down a little, but the age is the smaller problem. The question the survey asked was whether CIOs planned to move some workloads back from the public cloud to private cloud or on-premises, and a CIO who intends to repatriate one batch job answers yes in exactly the same way as one who is closing the AWS account and buying a building. The survey does not distinguish between them, and the articles quoting it are careful not to either.

The number that sits next to it, and that almost never makes the headline, comes from IDC's research of the same period, which found that only about 8 to 9 percent of companies were planning anything resembling a full exit. Put the two side by side and the story changes completely, because what they describe together is not an exodus from the cloud but a great many organisations doing the unremarkable thing of moving the two or three workloads that never belonged there in the first place.

Why the misleading version travels so well

Nearly everyone writing about repatriation is selling one side of it. Hosting companies and hardware vendors quote the 86 percent because it makes leaving sound inevitable, the hyperscalers and their partners answer with studies showing that nobody really leaves, and the rare first-hand account, of which 37signals is the famous one, gets recycled as proof of whatever the author already believed, usually without mentioning that a company with flat, predictable, thoroughly understood workloads is about the best case for owning hardware that exists.

We should declare our own position, since we are a hosting company quoting statistics at you. We run our own servers in Frankfurt and Amsterdam, with our own network, and we also build and operate a large share of our client work on AWS, Azure and Google Cloud, because that is frequently the right place for it. We get paid either way, which does not make us neutral, but it does mean we have no reason to push you in one direction, and it means we work with both kinds of infrastructure every day rather than arguing about one of them from the outside.

What actually decides it

The question of where a workload should live comes down to three properties of the workload itself, and the size of the company, the fashion of the year and the opinion of whoever spoke at the last conference have very little to do with it.

The first is how steady the load is. Cloud pricing is a premium for elasticity, and elasticity is worth a great deal when you need it. We run cloud scaling for research projects in healthcare and biotech that reach several thousand hosts at a time, and there is no sane way to do that on hardware you own, because you would be paying for a small data centre to sit idle whenever a project isn't running. Anything with sharp peaks has the same problem in miniature, since hardware sized for the busiest day of the year is mostly unused on all the others. An application whose graph is a flat line, on the other hand, is paying that premium every hour for a flexibility it never uses.

The second is how much data leaves. This is the line most people have never looked at, and it is the one that most often settles the argument. Outbound traffic from AWS costs around $0.09 per gigabyte after the first 100 GB a month, so a service pushing ten terabytes out is paying roughly $900 a month for bandwidth before it has bought any compute at all. If that traffic leaves a private subnet through a NAT gateway there is a further processing charge of about $0.045 per gigabyte on top, and traffic between availability zones is billed at about a cent per gigabyte in each direction, which a chatty application talking to a database in another zone can turn into real money without anyone noticing. On dedicated hosting, traffic at that volume is normally included in the price of the server.

The third is how much of the platform you are really using. If your application is built on queues, serverless functions and a managed database that would take you months to replace, then the cloud bill is partly the price of not having built and operated those things yourself, and that is often a perfectly good trade. If, on the other hand, what you actually run is a handful of virtual machines, a database and a load balancer, then you are renting ordinary servers at a premium and using none of the things the premium is for.

The costs that the savings claims leave out

The figures in repatriation articles, typically somewhere between 30 and 60 percent saved, are nearly always a comparison of the cloud invoice against the price of the hardware, and that comparison flatters the hardware because it leaves out the people.

Somebody has to patch the operating system, replace the failed disk, notice at three in the morning that replication has stopped, and design the backups so that losing one site does not lose the data. On AWS a good deal of that is included or at least hidden, whereas on your own servers it is somebody's job, and if it is nobody's job then it does not get done, which is how the cheap option quietly becomes the expensive one. Our own setup keeps Frankfurt's backups in Amsterdam and Amsterdam's in Frankfurt and tests restores every week, and none of that work appears in a price list for a dedicated server.

There is also the cost of moving, which is paid once but is rarely small, and the cost of getting it wrong, which is paid by whoever is on call. A migration that saves a few hundred a month and takes three engineers a quarter has not saved anything for years.

None of this means the savings are imaginary, because for steady, data-heavy workloads they are large and they are real. It means that the honest calculation has four lines in it rather than two, and that anyone who shows you only the first two is selling you something.

How to work it out for your own bill

You do not need a consultant for the first pass, only your last three invoices and an hour.

Start by splitting the bill into compute, storage, data transfer and managed services, because the total tells you nothing and the split tells you nearly everything. If data transfer and NAT charges come to a fifth of the bill or more, you have found the part that dedicated infrastructure would simply remove, whereas if managed services dominate, you are paying for a platform and leaving would mean rebuilding it.

Then look at how utilisation behaves over a month. Machines that sit at a steady load all day and all night are the candidates, and anything that genuinely scales up and down with demand should stay exactly where it is.

Then price the alternative in full, by which we mean the servers, a second location, the backups, the monitoring and the hours of somebody competent, whether that person works for you or for a company like ours. If the gap is still wide after all of that, it is worth doing, and if it has shrunk to nothing, you have just saved yourself a migration.

And if you are on a one or three year commitment, find out when it ends before anything else, because moving a workload you have already paid for does not save you anything at all.

What usually comes out of this exercise is not a decision to leave or to stay but a split: the steady base load and the heavy storage on hardware, the spiky and the specialised in the cloud, and a much smaller bill for both, which is precisely the unglamorous arrangement those survey numbers were describing all along, once you read past the headline.

If you would like a second pair of eyes

Send us the bill. We will tell you which parts of it would be cheaper elsewhere, which parts should stay exactly where they are, and whether the difference is worth the trouble of moving, and if the answer is that you are fine as you are, we will say so and you will have lost half an hour. The first call is free, it is with an engineer who runs both kinds of infrastructure, and there is no slide deck involved. Talk to us.

We wrote the longer, line-by-line version of the cost comparison in We run both AWS and our own servers, which goes through the hidden charges one at a time.

Written by

Blendbyte Team

Written by the engineers who run Blendbyte's hosting.

About us

Need help with your setup?

We build and run infrastructure for customers every day. If you need help with a server, cloud or software setup, talk to the engineers who do that work.

Talk to an engineer

New enquiries get a reply from an engineer within one business day.