When the cloud stops being the cheap option

Cloud repatriation is having a moment, and most of the takes on it lean the same wrong way: they read “we moved a workload off AWS” as proof the cloud was a mistake from the start. It usually wasn’t. The cloud won its default slot for reasons that haven’t gone anywhere, elasticity you stop paying for the moment a workload goes idle, and provisioning in minutes what used to need a purchase order and a quarter. What changed is that one class of workload got big enough and steady enough to break the arithmetic.

I’ve made the rent-versus-own call at a few scales. At a healthcare startup I took an AWS bill from ~$500K to ~$200K a year while raising uptime from ~95% to over 99.9%, which meant deciding, resource by resource, where cloud pricing earned its premium and where it didn’t. At GoPro I moved transcode fleets onto Spot for $350K+ a year in savings, which is the same math run the other direction: renting cheaper by absorbing interruption. At Postscript I hold ~$4M a year flat while traffic grows. The through-line is that this decision is always a utilization question wearing a strategy costume.

Deloitte’s Tech Trends 2026 named the current trigger plainly: monthly AI compute bills reaching into the tens of millions of dollars at the high end, which pushes enterprises to evaluate hybrid or on-premises deployment once cloud cost approaches roughly 60–70% of the equivalent hardware cost. Take that threshold as Deloitte’s figure rather than a law, but the shape of it is right.

The math that actually decides it

Cloud is a rental, and rentals carry a premium. You pay above the cost of owning hardware in exchange for elasticity and for handing someone else the operational burden. That premium is a bargain for a workload that’s spiky, unpredictable, or idle half the time, because you only pay for the hours you use and you’d have to buy for the peak. It’s a bad deal for a workload that runs flat-out, around the clock, indefinitely, because you’re paying the elasticity premium on capacity you never stop using.

Utilization is the whole variable. A fleet that averages 20% and spikes on Black Friday is exactly what the cloud is for. A fleet pinned at 90% for two years straight is paying rent on something it should probably own.

AI is driving repatriation because training and steady inference are the most utilization-flat compute most companies have ever run. A GPU cluster that’s busy almost continuously for years, on hardware that costs real money whether you rent or buy, is the textbook case where ownership wins. That’s why the engagements Deloitte describes cluster around AI infrastructure: hybrid optimization, GPU cluster management, and architecture review boards that price the deployment model before committing to it.

Why most workloads should stay put

The honest caveat is that repatriation is concentrated, not general. Your web tier, your dev and staging environments, your seasonal traffic, your batch jobs that run for an hour and release, those are what elasticity was built for, and moving them onto owned hardware trades a premium you can afford for a fixed cost you can’t scale down. The companies making news by repatriating are mostly moving one specific, enormous, utilization-flat workload, and keeping the rest where it is.

The mistake worth warning against isn’t repatriating. It’s repatriating on a slide deck. I’ve watched cost decisions get made on a vendor’s headline number and a general sense that the cloud is expensive, without anyone measuring the sustained utilization the entire calculation depends on. Owning hardware also means owning capacity planning, hardware refresh, data-center contracts, and the on-call that comes with all of it, and those costs are easy to leave off the comparison that justified the move.

The work is measurement, then a decision

The valuable version of this engagement isn’t “move us off the cloud.” It’s “tell us, with real utilization data, which workloads are candidates and which aren’t, and what owning them actually costs once you count the operational tail.” Most workloads come back cloud. One or two come back as genuine repatriation candidates, and now the decision rests on numbers instead of a mood.

That’s the discipline the current moment is missing. Repatriation is a real tool, sharpened by AI compute bills that genuinely can outgrow the cloud’s economics. It’s also a precision instrument being swung like a hammer, and the difference between the two is a month of measurement nobody wants to do before they’ve already decided.