Every few months another company publishes the blog post: we left the cloud, we bought our own hardware, we're saving seven figures. The posts are usually honest, the math is usually real, and the day one goes viral, somebody forwards it to your CTO with a one-line email. "Should we be doing this?"
It deserves a better answer than either reflex. Not the cloud-vendor reflex ("never, cloud is always cheaper at scale," which stopped surviving contact with invoices years ago) and not the contrarian reflex ("the cloud was a mistake," which conveniently forgets why everyone went there).
Why the math changed
Three things moved. First, a generation of workloads matured. The startup that needed elastic everything in 2016 is now a business with steady, predictable load, and the cloud's premium buys elasticity that steady workloads don't use. Paying for burst capacity you never burst into is renting a stadium for a dinner party.
Second, egress. Moving data out of a cloud has always been priced like the cloud would rather you didn't. Bandwidth-heavy businesses feel this as a tax on their own product.
Third, AI. GPU capacity is expensive to rent and sometimes simply unavailable when you want it, which has even cloud-native companies pricing out owned hardware for training workloads. That was unthinkable five years ago. It's a spreadsheet exercise now.
The workloads that genuinely qualify
There's a profile. Steady, predictable utilization, the flatter the better. Heavy bandwidth between the workload and its users. Hardware you'd keep busy around the clock, which is where owned GPUs earn their price. Data anchored somewhere specific by regulation. If a workload checks most of that list, running your own iron for it is a legitimate business decision, not nostalgia.
Notice what's not on the list: your spiky web tier, your quarterly batch jobs, the analytics cluster that runs hard three days a month, anything you're still iterating on weekly. Elasticity is exactly what those are paying for, and they should keep paying for it.
The invoice nobody prints
The published savings numbers are real. What they sometimes underweight is that the cloud bill was quietly covering things that come back when you leave: capacity planning, hardware refresh cycles, the ops depth to carry a 3 a.m. hardware failure, physical security, a disaster recovery answer that isn't "we have a second closet." None of that is impossible. All of it is payroll and attention. If the spreadsheet says you'll save 40% and the plan doesn't include hiring, the spreadsheet is missing a tab.
Usually, the answer is two workloads
The pattern that actually works, in our experience, is surgical. Identify the one or two systems that fit the profile, move those, and keep everything elastic exactly where it is. Hybrid isn't a compromise position; for mid-sized companies it's usually the correct one.
And before any of it: run the discipline check. We've taken as much as 70% off client cloud bills without moving a single workload, because the problem was never location. It was idle capacity, orphaned resources, and commitments that didn't match usage. If your bill is scary and your utilization graph looks like a heart monitor, you have a FinOps problem. Repatriation won't fix it, and it'll follow you to your own datacenter, where it's harder to see.
Five questions before you decide
Is the workload's utilization steady across a year? Would the hardware stay busy most of every day? Have you priced the people, not just the servers? Is egress a meaningful slice of the current bill? And the honest one: have you already done the cost discipline work in the cloud, or are you hoping a move makes the problem someone else's?
Three or more wrong answers and you're not a repatriation candidate. You're a cost-optimization candidate wearing a repatriation costume.
The honest take
Repatriation is a tool, not a movement. The companies doing it well treat it the way they'd treat any placement decision: workload by workload, with the people costs on the sheet and the boring cloud cleanup done first. The companies doing it badly are making an emotional decision with a spreadsheet as a witness.
Your infrastructure doesn't need a philosophy. It needs each workload running where the math says it should.