The repatriation math finally works, and CIOs are doing it
For a decade, moving workloads back on-prem was a nostalgia project. In 2026 it is a spreadsheet exercise with a positive number at the bottom, and roughly 80% of enterprises are running it.
Cloud repatriation used to be an argument you lost to the CFO. Now it is an argument the CFO starts.
The number that changed the conversation
The average repatriated workload saves 32% of its annual cloud infrastructure cost once on-premises hardware amortization, colocation, and operational overhead are netted out. That is not a rounding error — on an average enterprise cloud bill of $13.7 million a year, it is real money, and it explains why roughly 80% of enterprises now expect to repatriate some workload in the year ahead. Egress fees are doing a lot of the persuading: data-intensive workloads are reporting $50,000 to $500,000 a year just to move data out of the cloud they are trying to leave.
AI made the math sharper, not softer
The workload category driving this is not the one you would have guessed five years ago. It is AI infrastructure. On-premises AI reportedly delivers more than 50% in cost savings over three years compared with cloud API alternatives, and for GPU-heavy training workloads, on-prem infrastructure can pay for itself in under four months of continuous use. That inverts the old cloud pitch — elasticity for spiky, unpredictable demand — because AI training and inference at scale are not spiky. They are close to constant, and constant is exactly the load profile where owning the hardware beats renting it.
Repatriation is not a retreat
The old read on repatriation was defensive: a company admitting the cloud migration overshot. The current read is closer to a maturity signal. Enterprises spent a decade learning which workloads are genuinely elastic and unpredictable — and belong in the cloud — versus which are steady-state and belong on owned infrastructure. That is not a rejection of cloud. It is a company finally pricing its own workloads correctly instead of defaulting everything to the vendor that made onboarding easiest.
The move worth making is not "repatriate" or "stay." It is: run the same workload-by-workload cost model the 80% ran, and let the answer be workload-specific rather than ideological. Some of it will move back. Most of it, done honestly, will surprise you either way — and a company that has never run this exercise is the one most likely to be wrong about its own cloud bill.