Cloud Repatriation Turns Out to Be Selective, Not Wholesale

The cloud repatriation debate has resolved into something less dramatic than either side predicted: sorting. Companies are moving steady, predictable workloads, databases, batch processing, sustained inference, onto owned or leased infrastructure, while keeping the cloud for the variable demand it was always best at.

The arithmetic is unsentimental. A workload running flat-out around the clock can cost several times more in the cloud than on owned hardware amortized over its life, while a workload that spikes tenfold on demand would require ruinous overprovisioning anywhere else. Engineering teams now model the crossover point explicitly.

AI sharpened the calculation. Sustained inference is exactly the profile that favors ownership, and the availability of colocation capacity with modern accelerators gives mid-sized companies an option that once required hyperscale expertise.

Cloud providers have responded rationally, discounting committed steady-state usage and building services that span owned and rented infrastructure. The likely end state is the boring one: hybrid by spreadsheet, with ideology retired from the discussion.