Cloud Repatriation Is Real, but It Is a Correction, Not an Exodus

After more than a decade in which moving to the public cloud functioned as a default strategic decision requiring little justification, a countercurrent has become too large to dismiss as anecdote: companies moving specific workloads back onto owned or leased hardware, publishing detailed cost accounting to explain why. The pattern has acquired a name — cloud repatriation — and a predictable shape. The workloads coming back are steady, high-volume, and predictable: large databases, storage-heavy archives, video processing pipelines, and services whose load curves look the same every week. The economics driving repatriation are unambiguous for that specific workload profile. Public cloud pricing embeds a premium for elasticity — the ability to scale up in minutes and pay nothing for idle capacity — and a workload that never uses that elasticity pays the premium anyway, every month, indefinitely. Companies that have published their numbers report the same finding repeatedly: for stable workloads running at consistent utilization, owned hardware amortized over several years costs a fraction of the equivalent cloud bill, with data egress fees — the charge for moving data out of a cloud platform — often the line item that finally forces the analysis. What the repatriation stories rarely support is the broader conclusion sometimes drawn from them, that cloud computing was a mistake being unwound. The same companies moving stable workloads back overwhelmingly keep their variable ones in the cloud: anything spiky, experimental, geographically distributed, or new enough that its resource needs are unknown. Startups continue to launch cloud-first for the same reason they always did — capital not spent on server racks is capital spent on the product. The correction is to the totalizing version of the strategy, the one that migrated everything on principle without asking which workloads actually fit the pricing model. The durable lesson emerging from the trend is that infrastructure placement is a per-workload economic decision that deserves revisiting as a workload matures, not a one-time identity choice between cloud company and data-center company. A workload that was correctly cloud-hosted while its growth was unpredictable can become correctly self-hosted once its curve flattens, and the organizations handling this well are the ones that instrumented their costs finely enough to notice the crossover point — an accounting capability that, more than any hardware expertise, turns out to be the prerequisite for making the decision at all.