86 percent of CIOs now plan to move at least some workloads off public cloud, the highest rate ever recorded. Public cloud spending is also still growing. Both of those things are true at the same time, and understanding why is the actual story.
86 percent of CIOs now plan to move at least some workloads off public cloud, the highest repatriation rate ever recorded in the Barclays CIO Survey. Public cloud spending is also still growing, from roughly $669 billion in 2024 toward a projected $840 billion by 2026. Both of those things are true at the same time, and understanding why is the actual story, not the headline percentage alone.
Cloud repatriation is real, it is accelerating, and it is not the cloud exodus some headlines make it sound like. This is what's actually moving, why, and what it means for how you should be planning your own infrastructure.
The 86 percent figure gets cited constantly, and it's accurate as far as it goes. But the same body of research contains a second number that changes the picture considerably: only 8 to 9 percent of enterprises plan a full exit from public cloud, according to IDC. The other roughly 77 percentage points are enterprises repatriating some workloads while keeping others in the cloud, and often while growing their overall cloud footprint at the same time.
Gartner's forecast makes the actual shift explicit: 90 percent of organizations are expected to adopt hybrid infrastructure by 2027. That's not a rejection of cloud computing, it's the end of "cloud-first" as a default answer and the start of workload-by-workload decisions. Repatriation is a symptom of that shift, not a trend unto itself. Selective repatriation is driven overwhelmingly by three things: AI workloads with sustained, high GPU demand, data gravity and compliance requirements, and steady-state workloads where the cloud's elasticity premium goes entirely unused.
Broadcom's internal analysis found that a modern private cloud delivers 40 to 50 percent lower total cost of ownership than public cloud for steady-state workloads, the kind that run at predictable, continuous utilization rather than spiking unpredictably. Broadcom moved critical workloads off public cloud database-as-a-service offerings onto its own infrastructure and saved more than $10 million in the process.
Training and inference run on the most expensive instance tiers any cloud catalog offers, and unlike experimental workloads, production inference runs continuously rather than in bursts. A single 8-GPU cloud instance running around the clock can cost more than $270,000 a year. Multiply that across a real production fleet and the case for owning the underlying silicon, rather than renting it indefinitely, becomes difficult to ignore.
According to the Nutanix Enterprise Cloud Index 2026, 57 percent of IT leaders now feel the need to run infrastructure within a single country, a regulatory pressure that public multi-region cloud architectures don't always accommodate cleanly.
Every additional cloud platform introduces its own pricing model, operating model, and security controls, requiring specialized engineering skill for each. Over time, the staffing and coordination cost of managing multiple cloud platforms can rival or exceed the infrastructure savings multi-cloud was supposed to deliver, pushing some organizations back toward consolidated, predictable infrastructure they already know how to run.
Andreessen Horowitz's widely cited analysis found public cloud spending averages roughly 50 percent of cost of revenue for many software companies, with at least one company reporting cloud spend at 80 percent of revenue. At that ratio, cloud cost isn't a line item to optimize, it's a structural threat to the business model. This connects directly to what your cloud provider isn't telling you about hidden billing costs.
No example illustrates the repatriation thesis as clearly as 37signals, the company behind Basecamp and HEY. Rather than a single dramatic cutover, 37signals migrated app by app, validating cost and performance on each workload before moving the next one, a methodical approach that let the company de-risk the migration and prove the savings incrementally rather than betting the business on a single move. The company's public accounting of its cloud exit became one of the most referenced case studies in the entire repatriation conversation precisely because it showed the math working in public, not just in an analyst report.
The workload-by-workload framing is not a hedge, it's the actual decision-making tool. Different workload characteristics point to different answers:
| Workload Type | Better Fit | Why |
|---|---|---|
| Steady-state, predictable compute | On-premises / private cloud | Utilization is knowable, cloud's elasticity premium goes unused |
| Variable, spiky demand | Public cloud | Elasticity is the entire value proposition here |
| Continuous AI training or inference | On-premises / owned GPU | Runs 24/7 at the most expensive instance tier available |
| Compliance-heavy, regulated data | On-premises / sovereign region | Data residency requirements often force the decision |
| Experimental, short-lived projects | Public cloud | Speed to start and stop matters more than unit cost |
The pattern across every serious analysis of this trend is consistent: cloud remains the right answer for elastic, unpredictable, or short-lived workloads. On-premises and private cloud increasingly win for workloads that run continuously at knowable utilization, where the cloud's core value proposition, elasticity, is being paid for but never actually used.
The default answer of "put it in the cloud" made sense in an era when nobody had run the numbers yet. In 2026, running the numbers workload by workload, actual utilization pattern, data residency requirement, growth trajectory, is the baseline expectation, not an advanced optimization.
Continuous GPU-backed workloads carry a fundamentally different cost profile than general-purpose compute. If production AI inference is running 24/7 at hyperscaler rates, model the ownership economics explicitly rather than assuming cloud is automatically the cheaper or simpler choice.
The 37signals model, moving app by app and validating cost and performance at each step, is the lower-risk approach compared to a single large cutover. It also produces the kind of concrete, defensible savings data that makes the next migration easier to approve.
Only 8 to 9 percent of enterprises are planning to leave public cloud entirely, and for good reason: elastic and unpredictable workloads still belong there. Building a hybrid strategy that keeps the right workloads in the cloud while bringing the right workloads home is the actual goal, not a wholesale reversal of the last decade's cloud migration.
Moving a workload's data back on-premises means paying to move it out of the cloud first. Egress fees can meaningfully change the payback timeline on a repatriation decision, and should be modeled explicitly rather than discovered after the migration is already underway.
Cloud-first was never supposed to mean cloud-everything. Repatriation in 2026 is what happens when enough companies finally do the arithmetic.
Compare AWS, Google Cloud, Azure, and alternatives like Backblaze B2 Discover how much you could save in seconds
