Cloud repatriation is 21% of workloads, not an exodus. 37signals and GEICO show what actually moves back.

Cloud repatriation entered 2026 as a phrase that means everything and nothing. Trade coverage and hardware vendors cite numbers as high as 80 or 90 percent of companies supposedly leaving the cloud. The research behind those headlines rarely holds up to a second look. Flexera's 2026 State of the Cloud Report puts workload and data repatriation among enterprises at 21 percent, up two points on the year before. That is a real shift, not an exodus, and the gap between the headline figure and the measured one is where the useful part of this story sits.

The cloud repatriation number nobody agrees on

Flexera surveys IT leaders directly and defines repatriation narrowly: workloads and data that moved from public cloud back to on-premises or private infrastructure, not merely a pause in cloud growth. By that definition, enterprises with more than 1,000 employees report 21 percent of their cloud footprint has moved back. Managing cloud spend is now the top priority for 84 percent of respondents, ahead of security for the first time in the report's history.

Vendor-funded surveys with hardware or colocation to sell tend to report far higher figures, sometimes above 80 percent, often by counting stated intent rather than a completed move. Flexera's own 2026 report and reporting from CIO Dive describe a narrower, slower shift, and the two sets of numbers are answering different questions.

What moved at 37signals: the small end of the range

37signals is the case most engineers already know. After publicly committing to leave AWS in 2022, the Basecamp and HEY maker cut its annual cloud bill from roughly $3.2 million to about $1.3 million during 2025, a saving of close to $2 million in a single year, according to reporting in The Register.

Compute moved first. Storage followed in summer 2025, when 37signals shifted 18 petabytes off Amazon S3 onto its own Pure Storage arrays, a $1.5 million hardware purchase that costs under $200,000 a year to run. AWS waived roughly $250,000 in egress fees to let the migration happen. David Heinemeier Hansson's own projection puts total five-year savings from the combined exit above $10 million, without adding headcount to run the new hardware.

What moved at GEICO: the other end of the range

GEICO sits at a completely different scale. A decade of cloud migration took the insurer's annual bill past $300 million, and its parent, Berkshire Hathaway, wanted a different trajectory. GEICO's answer was not a full exit. It adopted Open Compute Project hardware, sourced through the Taiwanese manufacturer Wiwynn, to run its own compute and storage on commodity, open-standard racks, as described in reporting from The Stack.

The reported result is compute costs down more than 50 percent per core and storage down more than 60 percent per gigabyte against GEICO's prior cloud pricing. None of it happened quickly. GEICO's public roadmap targets repatriating at least half its workloads by 2029, a multi-year programme with its own tooling and staffing investment, not a weekend migration.

The pattern underneath both moves

Line up 37signals and GEICO and a pattern appears that the aggregate repatriation number hides. Both companies moved workloads with flat, predictable demand: a SaaS product with a known user base, an insurer's core policy and claims systems. Neither repatriated the workloads that spike unpredictably, because that is exactly the case public cloud still wins.

The same logic is starting to show up in AI infrastructure decisions. Sustained, high-volume inference and training runs behave more like 37signals' steady compute load than a startup's bursty traffic, which is why 2026 FinOps conversations increasingly put a repatriation column next to the reserved-instance one. Flexera's data shows both halves of this story happening in the same year: FinOps practice adoption reached 59 percent of organisations in 2026, up eight points, while wasted AI cloud spend still rose to 29 percent, the first increase in five years. Better cost discipline and rising AI waste are not a contradiction. They describe the same organisations getting good at managing the workloads they understand and still overspending on the ones they do not.

Where the cloud still wins

None of this is an argument for defaulting to on-premises. GEICO's own account of the project describes operational gaps that surfaced only once the hardware was running: tooling that AWS provided invisibly, expertise that took years to build internally, and a capital outlay that only makes sense at GEICO's volume. 37signals runs lean by design and had unusually predictable, self-hosted-friendly products to begin with; a company with seasonal traffic, multi-region latency requirements, or a small platform team is solving a different problem.

Egress fees remain the sharpest trap in the other direction. AWS waived them for 37signals as a goodwill gesture during a well-publicised exit. Most companies moving a smaller share of their data won't get that waiver, and need to model the standard egress bill before signing anything.

A framework for cloud repatriation before you model it

Three questions do most of the work before a repatriation business case is worth building:

• Is the workload's demand flat and forecastable, or does it spike in ways that would need idle on-premises capacity to cover?
• Does the team already have, or is it willing to build, the operational expertise the cloud provider currently supplies for free?
• Does the five-year total cost of ownership, including hardware refresh and egress, beat the cloud bill by enough to justify the capital risk?

37signals and GEICO both answered yes to all three, at completely different scales. Most companies running the numbers honestly will find they only clear the bar for a subset of their infrastructure, which is close to exactly what Flexera's 21 percent describes.

The workloads worth repatriating in 2026 are the ones you could describe in a single sentence to someone outside engineering: steady, well understood, and expensive enough at cloud list price to justify owning the hardware. Everything else is probably staying exactly where it is.