After 15 years on the cloud, 37signals concluded the promised savings never came and the team never shrank — so they bought their own servers and left.

37signals — the company behind Basecamp and HEY — spent about fifteen years on Amazon and Google cloud services. By 2023 the AWS bill was around $3.2 million a year, and their complaint had two halves: outrageous bills and ever-increasing complexity.

The cloud’s core promise is that you trade capital costs and operational toil for a bill that’s smaller in total and simpler to run. 37signals’ verdict on that promise, after a decade and a half, is one sentence: the savings never materialized and the team never shrank.

Their situation is specific, and it’s what makes the exit rational for them. 37signals runs a stable roster of applications — Basecamp, HEY, and five heritage apps — with predictable, steady load. That’s not a workload that needs to conjure a thousand servers for an hour and release them; it’s a workload that needs a known amount of capacity, running all the time.

For that shape, the cloud’s central value proposition — elasticity you rent by the second — is something you pay for continuously and rarely use.

The decision was to leave and run on their own hardware, keeping the same team. They project savings of roughly $10 million over five years — cutting infrastructure costs by between half and two-thirds — against a server investment on the order of $600,000. By the time of writing they’d already realized about $1 million a year in savings.

The storage leg is its own line item: exiting S3 was worth about five grand a day, roughly $1.8 million a year. And crucially, they moved the applications without adding any new staff — the same people who ran it in the cloud run it on the metal.

Leaving wasn’t free of engineering. 37signals built and released Kamal, a deployment tool (now through 1.0 and 2.0), and ran a “de-k8s” initiative to move off Kubernetes as part of simplifying the stack they’d own. The exit and the tooling are the same project: you can only run your own hardware with a small team if deploying to it is boring.

The load-bearing caveat isn’t in the cost table — it’s in who this applies to. The whole case rests on a steady, predictable workload and a team that was already doing the operational work. A company with spiky, unpredictable traffic, or one genuinely relying on the cloud to avoid an infrastructure team, would be copying the conclusion without the premises.

The cloud sells two discounts — a lower bill and a lighter team — and if you’re only ever getting one of them, you may be paying full price for a promise that doesn’t fit your workload.