Cloud Adoption Report 2026: The Repatriation Year
86% of CIOs now plan to move workloads back from public cloud — the highest rate recorded. What is driving it, what the AI workload numbers show, and why this is a correction rather than a reversal.
For fifteen years the cloud conversation had one direction. In 2026 the data shows something else: 86% of CIOs plan to move at least some workloads from public cloud back to private or on-premises infrastructure — the highest rate ever recorded.
This is not the end of cloud. It is the end of cloud-by-default, which is a different and more useful development.
What the data shows#
| Measure | 2026 |
|---|---|
| CIOs planning some repatriation | 86% — highest recorded |
| Enterprises that have already repatriated some workloads | 67% |
| Plan to within 12–24 months | 87% |
| Workloads originally moved that have been pulled back | ~20% |
| Enterprises repatriating, in process, or evaluating AI workloads | 93% |
| Have already moved some AI workloads | 79% |
| Public cloud as primary environment for production AI inference | fell 56% → 41% YoY |
| Now run or plan production AI inference on private cloud | 56% |
| Estimated on-premise AI saving over three years | 50%+ |
The AI inference number is the story#
Public cloud as the primary home for production AI inference fell from 56% to 41% in a single year, while private cloud rose to 56%.
That is a large shift in a short time, and the reason is arithmetic rather than ideology. Inference is a steady, predictable, high-volume workload — exactly the profile where per-token or per-hour cloud pricing is most expensive relative to owned capacity. Training is bursty and suits rented capacity. Serving a model to users all day does not.
Organisations discovered this the way organisations usually discover cost problems: on the invoice.
Why this is a correction, not a reversal#
The workloads coming back share a profile:
- Steady-state and predictable. No elasticity to capture, so no elasticity premium worth paying.
- Storage- or compute-heavy with stable demand.
- High egress. Data transfer out is charged per gigabyte by every major provider, and it accumulates quietly.
- Production inference, per above.
The workloads staying are also consistent: variable and spiky demand, new projects, global reach, disaster recovery, and managed services that would be expensive to reproduce.
That is not a retreat. It is workload placement finally being decided per workload — which is what should have been happening all along.
The uncomfortable finding#
Roughly one fifth of everything moved to public cloud has already been pulled back.
That represents two migrations paid for, two sets of re-architecture, and an outcome that could have been reached by analysis. The cost of "cloud-first" as a blanket policy is visible in that number.
The lesson is not "we were wrong about cloud". It is that blanket policies about infrastructure are expensive, in either direction. A "repatriate everything" policy in 2027 will produce the same number pointing the other way.
What makes repatriation possible#
Worth being explicit, because organisations that cannot do this are not choosing:
- Data portability, tested — not assumed
- Container images rather than provider-specific packaging
- Infrastructure as code, even if provider-specific; at least the intent is documented
- Standard database engines where the difference from a proprietary one is small
Deep use of proprietary managed services buys real productivity at the cost of optionality. That is a legitimate trade — but it should be made deliberately, and the test is simple: if this service vanished, how long to replace it? Weeks is fine. A year means it is a strategic dependency.
Our read#
The mature position is that cloud is a tool, not an identity. Steady workloads on owned hardware, variable and new workloads rented, is a coherent architecture rather than a compromise.
The organisations doing well in this data are not the ones that picked a side. They are the ones that can move a workload either way without a project — and that capability was built by the portability decisions above, usually years earlier.
The full reasoning is in our Cloud guide, which took this position before we compiled this data.
Method and limitations#
Synthesis of published 2026 CIO surveys and analyst reporting. Not primary research.
Repatriation surveys have a known bias: they are frequently published by vendors selling on-premises or private-cloud infrastructure, and "plan to move some workloads" is a low bar that most large estates would meet at any point in history. The AI inference figures are more interesting because they measure a change in where production workloads actually run, rather than intent. Where a figure came only from a vendor source, treat it as directional.
Published 2026-08-04.
Sources#
- Cloud Repatriation in 2026: Costs, Compliance, and Control — Volico
- Cloud Repatriation: The Strategic Shift in IT — HyScaler
- What's Driving Cloud Repatriation in 2026? — mrc
- Cloud Repatriation: Trends and Statistics — NovoServe
- The Great Cloud Repatriation of 2026 — Practical Logix
- On-Premise AI Statistics 2026
What else is coming for Cloud Adoption
Report Ready
The findings, with sources.
Data Not yet
The underlying figures.
Method Not yet
Where each number came from.
Updates Not yet
What changed since publication.