Skip to content

Business BreakdownsCase Study · 3 min read

37signals left the cloud by naming the conditions it no longer met

Most arguments about infrastructure are arguments about principle. 37signals argued about its own conditions, with its own number, and that is the part any founder can reuse.

Sources checked 23 October 2026 · 2 sources, listed at the end

What happened

Reported facts, each from the sources listed below.

On 19 October 2022, David Heinemeier Hansson of 37signals announced that the company was leaving the cloud to run on its own hardware. He gave a figure for part of what it was paying at the time:

“We’re paying over half a million dollars per year for database (RDS) and search (ES)”
David Heinemeier Hansson, 2022

His argument was not that the cloud is wrong for everyone. Renting, he wrote, makes sense when an application is very simple and low-traffic, or when its load is highly irregular, with wild swings or sharp peaks in usage. 37signals, in his account, was in neither situation.

The 2022 post gave no projected savings figure.

On 23 June 2023 he reported the move finished: the last application had been brought onto the company’s own hardware. In the same post he put the previous year’s cloud budget at $3.2m and called his savings estimate “back of the napkin math”: at least $1.5 million a year. These are his estimates, not audited results.

What we take from it

Our analysis: FoundersGPT’s reading of the facts above, not something the sources say.

The transferable move is not “leave the cloud”. It is the shape of the argument: name the conditions under which a default earns its premium, then check honestly whether you still meet them.

Defaults are adopted at one stage and paid for at the next. A setup chosen when traffic was unpredictable, the team was two people and speed mattered more than cost can be right at the time and wrong two years later, and nobody notices because nobody wrote down why it was chosen.

Publishing the number mattered too. An argument about principle cannot be checked; “we pay this much for these two services” can. Inside a company the same applies: a cost nobody has written down is a cost nobody re-examines.

Do this this week

  1. List your five most expensive defaults

    Hosting, agencies, software billed per seat, office space, a role hired for an earlier stage. Use the bank statement, not memory.

  2. Write the condition for each

    One sentence: “This is worth paying for while …”. For hosting it might be unpredictable load; for an agency, a skill you cannot hire yet.

  3. Check the condition against today

    Not against the day you signed. If the condition no longer holds, the default is now a habit.

  4. Put the annual cost beside it

    A yearly figure next to a condition that no longer holds is usually enough to make the decision for you.

  5. Diarise the next check

    Conditions change again. Re-read the list every quarter, before renewals rather than after.

What this does not show

This is one company’s decision, at its own size and, in its founder’s account, with steady and predictable load. It is not evidence that leaving the cloud saves money in general, and the savings figures are the founder’s own estimates.

Sources

  1. Why we’re leaving the cloud — David Heinemeier Hansson, world.hey.com, 19 October 2022.
  2. We have left the cloud — David Heinemeier Hansson, world.hey.com, 23 June 2023.

How these were checked: how we source. Found an error? Corrections policy.

One founder lesson a week, by email

Each issue gives one lesson, the mechanism behind it and one thing to do this week. Read Issue #1 before you subscribe.

Unsubscribe in one click. How we handle your address.

← All articles