Cloud Repatriation: Why Workloads Are Moving Back

The number everyone quotes is 86 percent. That is the share of CIOs in Barclays' Q4 2024 survey who said they planned to move at least some workloads from public cloud back to private or on-premises infrastructure, the highest figure the survey has recorded.

The number almost nobody quotes is 8 percent. That is the share of organisations IDC found intending to repatriate entire workloads.

The distance between those two figures is the whole story, and it is where most writing on this subject goes wrong.

What the numbers actually say

Take the headline figures at face value first, because they are real.

Barclays recorded 86 percent of CIOs planning some repatriation. IDC has reported for several years that somewhere north of 70 percent of enterprises are moving workloads back to on-premises or hybrid infrastructure, with around 80 percent expecting to repatriate compute or storage within a twelve-month window. Flexera's data has put roughly a fifth of workloads and data as already repatriated.

Now the qualifier. IDC's server and storage workloads survey found that only about 8 to 9 percent of organisations intend to repatriate entire workloads. The rest are relocating specific applications, specific data sets, specific compute.

And the sanity check. Gartner forecast global public cloud spending rising from $595.7 billion to $723.4 billion, a 21.5 percent increase. If 86 percent of enterprises were pulling out en masse, that growth would not be arithmetically possible.

So both things are true. Almost everyone is moving something. Almost nobody is moving everything.

That is not a contradiction. It is a description of what a maturing market looks like, and it is a far more useful basis for a decision than a headline about an exodus.

Why workloads come back

Four reasons, and they carry very different weight depending on what you run.

Cost on steady-state workloads. This is the largest single driver and the most straightforward. Public cloud pricing includes a premium for elasticity, and a workload running at consistently high utilisation with low variance never exercises the option it is paying for. Databases in steady production, internal applications with fixed user populations, and scheduled batch processing all fit this description. Broadcom's internal analysis has suggested a modern private cloud can deliver 40 to 50 percent lower total cost of ownership for steady-state workloads, and while that is a vendor with an obvious interest in the conclusion, the mechanism it describes is real.

Data gravity and egress. Once a dataset is large, moving it anywhere costs money in proportion to its size. Analytics platforms, media services and backup products can find transfer charges scaling with usage in a way that no configuration change addresses. We covered how those charges break down in cloud cost optimization.

AI and GPU workloads. The newest driver and the fastest growing. Sustained inference and training on rented GPU capacity is expensive in a way that intermittent workloads are not, and organisations running models continuously are doing the same arithmetic that sent database workloads back a decade earlier.

Compliance and data residency. Requirements around where data physically sits, who can access it, and what evidence you can produce for an auditor are pushing specific workloads toward environments where those questions have concrete answers.

Notice what is not on that list: dissatisfaction with the cloud as an idea. Repatriation is a placement decision, and the organisations doing it well are running the same evaluation in both directions.

What repatriation costs that nobody prices

The case against is stronger in 2026 than it was in 2024, and it deserves a fair hearing before you commit to anything.

Hardware got expensive. Estimates put hardware cost increases in 2026 at 15 to 25 percent, running ahead of the 5 to 10 percent cloud price increases that started the repatriation conversation. Memory pricing has been particularly brutal, with DRAM contract prices for 16Gb DDR5 chips moving from $6.84 in September 2025 to $27.20 by December.

If your business case was built on 2024 hardware pricing, rebuild it.

Then there is the operational surface. Owning infrastructure means owning patching, capacity planning, hardware failure, and the staffing to handle all three. Organisations that repatriate to on-premises frequently discover that the licence and instance savings were the easy part.

And the migration itself is not free. Egress charges apply to the data you are moving out, application testing takes longer than anyone estimates, and the network reconfiguration work is where timelines actually go.

The honest position is that repatriation works well for a specific workload profile and badly outside it. Anyone telling you otherwise is selling something, including anyone telling you the opposite.

Vendor lock-in, and what actually reduces it

Lock-in is cited constantly as a repatriation driver, and it is usually described too vaguely to act on.

It has three separate components.

Data lock-in is the cost of getting your data out, which rises with volume because egress is billed per gigabyte. This is the one that gets worse over time regardless of what you do.

Architectural lock-in is dependence on proprietary managed services. An application built on a provider's proprietary database, queue, and serverless runtime cannot move without being rewritten. An application built on PostgreSQL, Redis and containers can.

Commercial lock-in is committed spend. Multi-year agreements and reserved capacity are genuine savings and also genuine constraints on your ability to leave.

Reducing lock-in is mostly about the second one, and mostly about choices made at build time rather than exit time. Open APIs, portable data formats, containerised workloads and infrastructure defined as code all keep options open. Platforms built on open standards, whether OpenStack, Kubernetes or plain virtual machines, are portable in a way that proprietary managed services are not.

The practical test is simple. If you had to move this workload in ninety days, what would break? The answer is your lock-in position, and it is usually more specific and more fixable than the general anxiety suggests.

When repatriation makes sense, and when it does not

It makes sense when utilisation is high and steady, when the workload has been rightsized and committed and is still expensive, when data volumes are large and growing, when compliance requires specific answers about location and access, and when you have or can buy the operational capability to run it.

It does not make sense when demand is genuinely variable, when the estate is small enough that operational overhead outweighs the saving, when the application depends on proprietary managed services that would need rewriting, or when the migration would land in the same quarter as a hardware refresh you have not budgeted.

The middle path most organisations end up on is neither. Flexera's 2026 data put 73 percent of organisations on hybrid estates, and Gartner projects around 90 percent adopting hybrid infrastructure by 2027. Hybrid is not a compromise position here. It is what the workload-by-workload evaluation produces when you run it honestly.

How to actually decide

Five steps, and the first one is the one people skip.

Profile utilisation before costing anything. You are looking for shape rather than size. High and steady is a repatriation candidate. Spiky and unpredictable is not.

Cost the specific workload, not the estate. Estate-level comparisons hide the fact that the answer differs per workload, which is the entire finding.

Price the migration, including egress. Moving the data out costs money that has to sit in the business case.

Check the architectural dependencies. Every proprietary managed service in the stack is a rewrite, and rewrites are where repatriation projects die.

Price the operations. Whether that is headcount for on-premises, a managed provider, or colocation with remote hands, this line is real and it is the one most often left out.

We wrote a longer version of this evaluation in when private cloud beats public cloud, and when it doesn't, and a workload-by-workload breakdown in public vs private vs hybrid vs bare metal.

Frequently asked questions

What is cloud repatriation? Moving workloads, applications or data from public cloud back to private cloud, on-premises infrastructure or colocation. It is usually selective rather than complete.

Are companies really leaving the cloud? Some workloads are moving, entire estates mostly are not. Barclays found 86 percent of CIOs planning some repatriation, while IDC found only around 8 percent planning to repatriate entire workloads. Public cloud spending continues to grow.

Which workloads are the best repatriation candidates? Steady-state, high-utilisation workloads with predictable demand. Databases in production, internal applications, scheduled batch processing, and large datasets with heavy transfer.

What is an AWS exit strategy? A documented plan covering which workloads would move, what they depend on, what the data transfer would cost, and how long it would take. It is worth writing even if you never execute it, because it tells you your actual lock-in position.

Does repatriation always save money? No. Hardware costs rose 15 to 25 percent in 2026, faster than cloud prices, and operational staffing is a real cost. It saves money on the right workload profile and costs money outside it.

What is the alternative to on-premises repatriation? Managed private cloud and colocation both sit between public cloud and running your own data centre, keeping dedicated infrastructure without the facilities and staffing burden.

Where InMotion Cloud fits

We sell managed private cloud, so the disclosure matters here more than usual.

For an organisation that has identified genuine repatriation candidates and does not want to build a data centre or staff a platform team, managed private cloud is the middle option: dedicated infrastructure, fixed monthly cost, and someone else operating it. That suits the workload profile described above.

It does not suit variable workloads, and it is not an argument for moving everything. If your evaluation says three workloads should move and eleven should stay, that is a correct answer and not a failure of nerve.

Our pricing calculator will cost a specific workload against what you pay now.