Reviewed by: Mansoor Ali, Technical Editor, PenPonder | Last Updated: August 2026
You have seen the stat. 83% of companies are pulling workloads off the public cloud. Sometimes the number is 86%.
It shows up in vendor blogs. It shows up in sales decks. It shows up on LinkedIn every few weeks.
The number is real. What people do with it is not.
Where the 83% actually came from
It comes from a Barclays CIO Survey. Michael Dell posted the chart in April 2024 and it spread from there.
The finding was that 83% of enterprise CIOs planned to repatriate at least some workloads. That was up sharply from 43% in late 2020. A later wave of the same survey put the figure at 86%.
Two details get dropped every time it is quoted.
First, it measures plans. Not moves. Nobody checked back to see who followed through.
Second, it says at least one workload. Moving a single database off AWS puts a company in that 83%. So does a full exit. The survey treats them the same.
What actually moved
IDC published a report in June 2024 called “Assessing the Scale of Workload Repatriation.” That is the source of the other big number, the one saying 80% of organizations expected to repatriate something within a year.
Read further into it and the picture changes. CIO reported that fewer than one in ten respondents had actually moved an entire workload back.
The same report noted something else worth knowing. That 80% figure was in line with what earlier surveys had found. It was not a spike. Companies have been saying this for years.
| What gets quoted | What the source says |
|---|---|
| 83% of companies are leaving the cloud | 83% planned to move at least one workload |
| 80% are repatriating this year | 80% expected some repatriation, in line with prior years |
| A mass exodus is underway | Fewer than 10% had moved a whole workload |
Both surveys are from 2024
This is the part that surprised me most.
The Barclays data is from 2024. The IDC report is dated June 2024. Yet articles published this year still present both as the current state of things.
Two year old survey data is not a 2026 trend. It is a 2024 mood, reprinted.
The cloud providers just reported their numbers
If companies were really leaving in bulk, it would show in hyperscaler revenue. It does not.
The July 2026 earnings round told a different story. CNBC reported AWS revenue of $42.23 billion for the June quarter, growing about 37%. That was its fastest growth since 2021.
Microsoft put growth for Azure and other cloud services at 43%. Google Cloud grew roughly 82% to nearly $25 billion.
You can argue about how much of that is AI demand. You cannot call it an exodus.
Who gains from the story
Look at who publishes repatriation content. Hosting companies. Colocation providers. Bare metal sellers. Private cloud vendors.
All of them make money when you leave AWS.
Broadcom put the 83% figure on stage at VMware Explore in 2024. It was shown to customers and partners as proof of the opportunity to sell them virtualization software. That is not a neutral use of a statistic.
None of this means the number is fake. It means you should ask who is holding it up before you act on it.
Where leaving the cloud genuinely works
The best documented case is 37signals, the company behind Basecamp and HEY.
Their AWS bill had reached about $3.2 million a year. Their workloads were steady and predictable. They were paying for flexibility they no longer used.
So they bought hardware instead. Around $700,000 of Dell servers. CTO David Heinemeier Hansson wrote in 2024 that the hardware cost was fully recouped during 2023. The annual bill had dropped to $1.3 million. Projected savings now run past $10 million over five years.
Those numbers are real. Two caveats matter.
Data Center Dynamics pointed out what the savings figure leaves out. Hardware refresh cycles. The extra operations staff needed to run your own gear. The power and cooling bill for a data center. Those are not small.
And when 37signals moved its stored data out, The Register reported that AWS waived $250,000 in egress fees for them. Most companies will not get that call answered.
How to decide for your own systems
Forget the survey. The question is about your workloads, one at a time.
| Workload type | Usually better to |
|---|---|
| Runs 24/7 at steady load | Own it. You are renting the same capacity every hour. |
| Spiky or seasonal traffic | Rent it. This is what cloud pricing is built for. |
| Early stage or still changing | Rent it. Do not buy hardware for a guess. |
| Constant GPU training or inference | Do the math. Owning often pays back inside a year. |
| Large stored datasets | Check the exit cost before you commit either way. |
Then run three numbers.
- What you spend now, per workload, not as one bill.
- What the equivalent hardware, rack space, power and cooling would cost over five years.
- What it costs to get your data out, including egress fees.
Then add the staff. Somebody has to patch, monitor and replace that hardware. If you no longer have those people, hiring them is part of the price.
If the numbers favour owning, own it. Plenty of workloads do. Just get there through your own spreadsheet, not through a chart on LinkedIn.
What this actually is
Companies are not abandoning the cloud. They are getting choosier about what goes in it.
That is a duller story than a mass exodus. It is also the true one, and it is far more useful if you are the person signing the invoice.
For the wider picture on where computing is moving, see our guide to edge computing. If you are weighing up how to run workloads across your own hardware, our containerization and Kubernetes guide covers the tooling side. For the rules that apply wherever your data sits, see cloud security compliance. And for everything else we cover on infrastructure and careers, start with the complete technology guide.
Sources: Barclays CIO Survey (2024), as circulated publicly by Michael Dell; IDC, “Assessing the Scale of Workload Repatriation” (June 2024), reported by CIO; Amazon, Microsoft and Alphabet quarterly results for the June 2026 quarter, reported by CNBC; David Heinemeier Hansson, 37signals cloud exit update (October 2024); The Register on the 37signals storage migration and egress fee waiver; Data Center Dynamics on the costs excluded from those savings figures. PenPonder does not provide investment or technology purchasing advice. Readers should run their own numbers and consult primary sources.
