Cloud repatriation enterprise strategy has moved from a fringe cost experiment to a board-level priority, with the Barclays Q4 2024 CIO Survey recording the highest repatriation intent ever measured: 86% of CIOs plan to pull at least some workloads back from public cloud to private infrastructure or on-premises environments. For enterprise IT leaders, this is no longer a trend to monitor - it is an architectural decision that is already arriving on the agenda.
Sources: IDC 2024, Barclays CIO Survey Q4 2024, Flexera State of the Cloud 2025
| Value (%) | Share (%) |
|---|---|
| Enterprises expecting repatriation (IDC 2024) | 80% |
| CIOs planning to repatriate (Barclays Q4 2024) | 86% |
| Workloads already repatriated (Flexera 2025) | 21% |
Record intent, but execution is still catching up
The gap between planning and action is the defining feature of the current repatriation moment. While 86% of CIOs signal intent to repatriate and approximately 80% of enterprises expect some level of compute and storage repatriation within the next 12 months (IDC, 2024), Flexera's 2025 State of the Cloud report puts the share of workloads and data already moved back at just 21%. That 60-plus point gap represents a large execution wave that has not yet landed - and the window to plan for it properly is closing.
The fact that intent is so far ahead of action is not necessarily a sign of hesitation. Many organizations are still working through workload classification, vendor contract timing, and the capital planning needed to build or expand private infrastructure. But the organizations that delay the planning work risk being forced into rushed decisions that fragment their environments without adequate governance tooling in place.
The gap between 86% planning intent and 21% workloads already moved represents a large execution wave - and the window to plan for it properly is closing.
Why this repatriation wave is different from earlier ones
Previous repatriation cycles, most notably around 2018 to 2020, were driven almost entirely by unit cost concerns - surprise bills, egress fees, and the realization that lift-and-shift migrations had not delivered the savings once promised. The 2025 wave has those same cost pressures but adds two significant new factors that were absent before.
The first is AI workload economics. Running persistent inference workloads on hyperscaler compute at scale is genuinely expensive, and latency between model serving infrastructure and proprietary data stores adds operational complexity. Enterprises are increasingly finding that owning GPU capacity - or colocating it closer to their data - produces better unit economics for workloads that run continuously rather than in bursts. The second factor is data sovereignty. Stricter regulatory requirements in finance, healthcare, and public sector - particularly in the EU and across Asia-Pacific - are pushing regulated data out of multi-tenant public cloud environments entirely, regardless of cost.
The 37signals case study, analyzed by Kyndryl in June 2025, remains the most cited concrete example: the company projected roughly $7 million in savings over five years by repatriating off a $3.2 million annual public cloud bill. Not every organization will see those ratios, but the directional logic - that predictable, high-volume workloads are often cheaper to own than to rent - is now widely accepted among enterprise architects.
What this means for cloud spending overall
CIOs should be precise in how they frame repatriation internally, because the narrative is easily misread as a retreat from cloud altogether. It is not. Gartner forecasts global end-user cloud spending at roughly $723 billion in 2025 - a number that does not suggest any broad abandonment. What is happening is a workload placement recalibration: organizations are becoming more deliberate about which workloads belong in public cloud and which do not.
Bursty, experimental, and development workloads remain strong fits for public cloud, where elastic capacity and managed services reduce operational overhead. Predictable, data-heavy, latency-sensitive, and compliance-constrained workloads are the repatriation candidates. The hybrid model - not a binary choice between cloud and on-premises - is the dominant landing position emerging from current planning cycles.
What architects need to do before moving anything
The repatriation organizations that succeed will be the ones that run workload classification before they move anything. That means categorizing production workloads by three dimensions: cost predictability (is this workload bursty or steady-state?), regulatory sensitivity (does the data carry jurisdictional or sector-specific requirements?), and latency or data gravity (does the workload need to sit close to a large on-premises dataset or AI model?). Workloads that score high on at least two of those dimensions are repatriation candidates.
A parallel track on vendor lock-in mitigation is equally important. Organizations moving workloads back to private infrastructure should standardize on open formats where possible - OCI-compatible container images, Kubernetes orchestration, and open table formats such as Apache Parquet for analytical data. This preserves optionality and avoids trading hyperscaler lock-in for vendor lock-in on private infrastructure hardware or software stacks.
FinOps tooling that spans both public and private environments is the governance requirement that most organizations underestimate. Without unified cost visibility across hybrid environments, the savings case for repatriation becomes difficult to validate, and shadow spending in either direction becomes harder to catch. CIOs presenting repatriation business cases to the board should build the FinOps investment into the program budget from the start, not as an afterthought.
