Back to Blog

Cloud Repatriation in 2026: When Leaving AWS Saves Money, and When It Is Just a Good Blog Post

• 3 min read • Ahmed Daoudi

"Cloud first" was the default answer for a decade. In 2026 it is a question again. Platform teams are moving selected workloads back to owned or leased hardware, and the arguments on both sides have become louder than the numbers.

The case everyone quotes

37signals, the company behind Basecamp and HEY, made repatriation mainstream by publishing its numbers. It was spending about $3.2 million a year on AWS, moved to owned and colocated hardware, finished in 2023, and has reported large ongoing savings since. Ahrefs has made similar claims about avoiding cloud costs with bare metal.

These are real results, but they come from companies with steady, predictable load and strong operations teams. That context matters more than the headline percentage.

When leaving tends to pay off

When it rarely pays off

Do this before you buy a single server

Most teams I see overspending on AWS are not victims of the cloud. They are running oversized instances, paying on-demand prices for steady workloads and routing traffic through NAT gateways they do not need. Start here:

  1. Right-size instances and move steady workloads to Savings Plans or Reserved Instances. My guide to EC2 instance types covers the families that matter.
  2. Try ARM-based Graviton instances, which are usually cheaper for the same work.
  3. Use VPC endpoints for S3 and other AWS services instead of sending that traffic through NAT.
  4. Put a CDN in front of anything that serves files.
  5. Tag everything and review the bill line by line once a month.

If the bill is still painful after that, run the full comparison: hardware, colocation, bandwidth, staff time, hardware refresh every few years, and the one-time cost of moving data out. On the last point, ask AWS first, since it has offered to waive data transfer fees for customers who are migrating away.

My take

Repatriation is not a rebellion against the cloud. It is a placement decision: each workload goes where it is cheapest and safest to run. For most startups that is still the cloud, configured well. For a mature company with a flat traffic graph and a capable ops team, a rack of servers can be the boring, correct answer.

Sources