Cloud platform services: when colocation costs less than the public cloud
For workloads that run the same every day, colocation or a hosted private cloud usually costs less than AWS or Azure; public cloud earns its premium on workloads that spike, shrink or are still being built. The best-known analysis, from a16z, found repatriation could cost one-third to one-half as much as the cloud for equivalent workloads. Few companies move everything back: the usual answer is a mix of cloud platform services, with each workload placed on its cost profile.
What are cloud platform services, and where does colocation fit?
"Cloud platform services" covers more than the three hyperscalers. For a mid-market company, the realistic menu looks like this:
- Public cloud (AWS, Azure, Google Cloud): rent compute and storage by the hour or second, pay for what you use, scale up and down on demand.
- Hosted private cloud: dedicated infrastructure run by a provider, billed at a steady monthly rate.
- Colocation: your servers in someone else's data center, which supplies the building, power, cooling, network connectivity and physical security.
- Backup and disaster recovery: copies of your systems at a second site, ready to take over if the first one fails.
Most companies end up with more than one. In Flexera's 2026 State of the Cloud report (753 respondents), 73% said they use hybrid cloud. The question is less "cloud or not" than which workloads belong where.
When does public cloud cost more than colocation?
On-demand pricing is built for workloads that change. You pay for the ability to add a hundred servers on a Tuesday afternoon. If your ERP, file servers and reporting database run at roughly the same load all year, you keep paying for flexibility you never use.
The industry data shows how common overspending is:
- Flexera's 2026 report estimates wasted IaaS and PaaS spend rose to 29%, the first increase in five years, and 85% of respondents called managing cloud costs their top challenge.
- IDC reported in October 2024 that close to half of cloud buyers overspent in 2023, and 59% expected overruns in 2024.
- In their 2021 paper "The Cost of Cloud, a Trillion Dollar Paradox," a16z's Sarah Wang and Martin Casado estimated about $100 billion of market value lost across the top 50 public software companies because of cloud's effect on margins.
That last figure is about large software companies, and the authors are clear it only makes sense for some workloads. It is still the clearest statement of the principle: at steady scale, renting costs more than owning or leasing.
Is cloud repatriation real, or a few loud examples?
Both. The loud examples are real, and the broader trend is smaller than the headlines suggest.
The examples
- 37signals had an AWS bill of $3.2 million a year. It bought about $700,000 of Dell hardware, paid back within 2023, and David Heinemeier Hansson of 37signals put savings at about $2 million a year, projected at "well over ten million dollars over five years."
- 37signals then left Amazon S3 when its contract ended on 30 June 2025. It moved 18 PB onto Pure Storage across two data centers for about $1.5 million in hardware; the S3 bill had been about $1.5 million a year. AWS waived about $250,000 in egress fees.
- Dropbox saved nearly $75 million over two years by moving off public cloud, according to the a16z paper.
The trend
IDC's Server and Storage Workloads Survey (2,250 respondents, March 2024) found about 80% expected some repatriation. But IDC also found only 8% to 9% plan full workload repatriation. What usually moves back is production data, backup and DR, and compute. Flexera's 2026 data shows repatriation of workloads and of data each up 2 points year on year: a steady drift, not a stampede.
One barrier has come down. In March 2024, AWS removed data-transfer-out fees for customers leaving AWS.
Which workloads belong in colocation, and which in public cloud?
A rough guide to sorting workloads across cloud platform services, before anyone looks at your actual bills:
| Workload pattern | Usually cheaper in | Why |
|---|---|---|
| Steady, always-on (ERP, databases, file and print) | Colocation or hosted private cloud | No premium for elasticity you don't use |
| Large, growing data sets | Colocation or private cloud | Storage priced per GB per month adds up at scale |
| Backup and disaster recovery | A second site you control or contract | Predictable cost; one of the categories IDC sees moving back |
| Seasonal or spiky (retail peaks, batch jobs) | Public cloud | You pay for peaks only when they happen |
| New products still finding their load | Public cloud | Hard to size hardware for something that doesn't exist yet |
| Built on hyperscaler-only managed services | Public cloud | Moving them means rebuilding them |
How much can colocation save compared with AWS or Azure?
It depends on your workloads, so be suspicious of any single number, including ours. The a16z estimate of one-third to one-half of cloud cost comes from large software companies. Liberty Center One, the Royal Oak, Michigan data center we work with, says typical migrations to its hosting cost 30% to 50% less than the hyperscalers. Treat that as the provider's own figure, and ask for it to be proven on your workloads.
Two published figures from Liberty Center One, written up on our results page: one client eliminated $181,400 in VMware expenses by moving to its OpenStack cloud, and one organization quoted more than $1 million a year for disaster recovery by Microsoft and Amazon got the equivalent from Liberty Center One for about 75% less. Treat both as examples, not averages; your own bill is the only number that decides it.
Savings can also disappear. Colocation means you own or lease hardware, plan capacity and refresh it. If your team has no time to do that, a hosted private cloud, where the provider runs the hardware, is often the better middle ground.
How should you decide between colocation and public cloud?
Start with evidence you already have:
- Pull 12 months of cloud bills and tag each line to a workload. Anything that barely changes month to month is a candidate.
- Map dependencies. In Flexera's 2026 report, understanding application dependencies was the top migration challenge (54%). A database that moves without the three apps that call it will cost you in latency and egress.
- Compare like for like. Comparing on-premises and cloud costs was a challenge for 43% of Flexera's respondents. Include hardware refresh, staff time, support, connectivity and DR on the colocation side; include reserved-instance discounts on the cloud side.
- Move one steady workload first, measure it for a quarter, then decide on the next.
- Keep what the cloud does well in the cloud. A hybrid mix of cloud platform services is not a failure to decide.
If you want those numbers run for you, Liberty Center One costs your workloads line by line before anything moves. We make the introduction; Liberty Center One pays us if an engagement goes ahead, and you pay us nothing. Our infrastructure page covers hosting, colocation and DR in more detail.
Sources
- Liberty Center One, "VMware Alternative" and "Disaster Recovery" service pages (libertycenterone.com, accessed September 2026)
- a16z, "The Cost of Cloud, a Trillion Dollar Paradox" (May 2021)
- David Heinemeier Hansson, "Our cloud-exit savings will now top ten million over five years" (October 2024)
- The Register, "37signals on-prem migration to save millions, abandon AWS" (May 2025)
- Flexera, "Flexera 2026 State of the Cloud Report: The convergence of cloud and value" (March 2026)
- Flexera, "What 5 years of Flexera’s State of the Cloud data reveals" (September 2026)
- IDC, "Storm Clouds Ahead: Missed Expectations in Cloud Computing" (October 2024)
- CIO.com, "The great repatriation? IT leaders reset cloud strategies to optimize value" (2024)
