Executive Summary
Repatriation stopped being an argument about pride and became a spreadsheet question. The market lost faith in the idea that cloud is always cheaper, and providers answered by loosening exit fees. What remains is a test of load shape. Flat workloads amortize. Spiky ones do not.
Start with the monthly baseline, not the egress bill. Amazon Web Services charges $0.09 per gigabyte for the first 10 terabytes, and it now waives transfer out when a customer leaves entirely. That toll is rarely why the case pays. 37signals spent about $700,000 on servers and reported annual savings near $2 million, then moved 10 petabytes out of object storage and projected more than $10 million over five years. Its load held still. Budget the operations salary too, and move storage before compute. Most large companies land on hybrid, whether or not that was ever the plan.
Cloud repatriation used to be a confession. Now it is a line item. The question is no longer whether leaving is respectable. It is whether the math works for your workload. For steady-state systems it often does. For anything that moves, it usually does not.
The bill that starts the conversation is egress. Amazon Web Services charges $0.09 per gigabyte for the first 10 terabytes of internet egress each month, then steps down by volume. There is a 100 gigabyte free allowance. Google Cloud and Microsoft Azure run similar ladders.
Egress is the visible number. Cross-zone and cross-region transfer is the one that surprises people. Traffic between two availability zones costs money in both directions. A busy service mesh spread across three zones can quietly burn thousands a month before a byte leaves the region. None of that traffic crosses the public internet, and none of it is free.
Egress Fees Start the Argument but They Do Not Win It
Be careful here. Egress fees are real, and they are the easiest thing to point at. They are rarely the reason repatriation pays. Providers have also softened the exit. AWS now offers free data transfer out when an eligible customer moves all their data off its platform, and European rules pushed changes to egress pricing. That removes the one-time toll. It does not change the monthly economics.
The number that actually decides the case is the monthly baseline. If your load is flat, you are renting capacity you would otherwise buy once. If your load is spiky, cloud burst is worth a premium, and owning racks of idle servers is the mistake.
The Payoff Comes From a Flat Baseline and Full Amortization
37signals made this argument in public and with numbers. The company moved its main apps off AWS, spent about $700,000 on servers, and reported annual savings near $2 million. It later moved 10 petabytes out of object storage to on-prem arrays and projected total savings above $10 million over five years.
Their load was steady. That is the whole trick. Predictable demand lets you size hardware to actual use and run it hard for five to seven years. Cloud pricing builds in a buffer for capacity the provider must hold ready. You pay for that insurance whether you claim it or not. When you do not need the insurance, you are overpaying for it every month.
Amortization is what makes the math work. A server bought once and used at high utilization for years beats a rented instance that bills by the hour forever. The catch is the word used. Under-utilized hardware is just a cloud bill with extra steps and a depreciation schedule. The clearest candidates to move are internal platforms, batch systems, and storage, where demand is known and latency is not a contract.
Storage is usually the first thing that should move, and compute is the last. Storage is flat by nature. You write data once and read it for years. Compute has to flex with demand.
The Cost Nobody Budgets Is the Team
Here is where most business cases fall apart. The spreadsheet compares hardware to instance pricing and stops. It rarely prices the people. Someone has to own firmware, backups, capacity planning, and the 3 a.m. page. That is a real salary, and it does not shrink with the hardware. Run it in a colo and you also buy the network gear, the power contract, and the spare parts. That is the cost that turns a clean spreadsheet into a two-year slog.
Repatriation is often a mistake for early-stage companies, for products with unpredictable growth, and for teams with no operations bench. It is also wrong when you depend on managed services. Replacing a managed database, queue, and identity provider with self-run versions is a multi-year project, not a migration.
Be honest about the buffer too. On-prem capacity has to cover your peak. If peak is double the baseline, you buy for peak and idle the rest. That is exactly the waste you left the cloud to escape. Hybrid is the honest answer for most large companies. Keep the system of record on owned hardware. Keep the elastic tier and the global edge in the cloud.
The rule is simple. Repatriate flat, heavy workloads you already understand. Keep the spiky, the new, and the global in the cloud. Run both numbers twice, and include the salary you will not admit you need.
Related reading. Your Cloud Bill Follows You Out the Door. Gartner’s Container Grid Names Seven Leaders. Only Two Are Not Hyperscalers.. Sovereign Cloud Puts a Border Around Your Data. The 2026 State of Enterprise Infrastructure.
Get the next one before it is old news
Independent analysis of cloud-native infrastructure, Kubernetes and data centre economics. No vendor spin.